Abstract
Efforts at player motion tracking have traditionally involved a range of data collection techniques from live observation to post-event video analysis where player movement patterns are manually recorded and categorized to determine performance effectiveness. Due to the considerable time required to manually collect and analyse such data, research has tended to focus only on small numbers of players within predefined playing areas. Whilst notational analysis is a convenient, practical and typically inexpensive technique, the validity and reliability of the process can vary depending on a number of factors, including how many observers are used, their experience, and the quality of their viewing perspective. Undoubtedly the application of automated tracking technology to team sports has been hampered because of inadequate video and computational facilities available at sports venues. However, the complex nature of movement inherent to many physical activities also represents a significant hurdle to overcome. Athletes tend to exhibit quick and agile movements, with many unpredictable changes in direction and also frequent collisions with other players. Each of these characteristics of player behaviour violate the assumptions of smooth movement on which computer tracking algorithms are typically based. Systems such as TRAKUS(TM), SoccerMan(TM), TRA PERFORMANCE(TM), Pfinder(TM) and Prozone(TM) all provide extrinsic feedback information to coaches and athletes. However, commercial tracking systems still require a fair amount of operator intervention to process the data after capture and are often limited by the restricted capture environments that can be used and the necessity for individuals to wear tracking devices. Whilst some online tracking systems alleviate the requirements of manual tracking, to our knowledge a completely automated system suitable for sports performance is not yet commercially available. Automatic motion tracking has been used successfully in other domains outside of elite sport performance, notably for surveillance in the military and security industry where automatic recognition of moving objects is achievable because identification of the objects is not necessary. The current challenge is to obtain appropriate video sequences that can robustly identify and label people over time, in a cluttered environment containing multiple interacting people. This problem is often compounded by the quality of video capture, the relative size and occlusion frequency of people, and also changes in illumination. Potential applications of an automated motion detection system are offered, such as: planning tactics and strategies; measuring team organisation; providing meaningful kinematic feedback; and objective measures of intervention effectiveness in team sports, which could benefit coaches, players, and sports scientists.
The capacity of portable data recording and analysis systems to process large amounts of information for motion tracking has improved markedly in the last 10 years. Indeed, equipment such as global positioning systems (GPS), high-speed video and accelerometers are increasingly used in the measurement and evaluation of human physical activity. Consequently, the quantitative analysis of team and player activity is now an important aspect of the coaching process in sport.[1] Obtaining accurate positional information about sports players is of interest to coaches and high-performance support teams because of the potential to relate performance to tactics, and to assist in the design of better training programmes. Furthermore, sport scientists can utilize such information to understand the coordination dynamics of player activity and the most influential constraints acting upon them. For example, such data could be processed to calculate player velocities and accelerations and establish performance profiles in successful compared with unsuccessful games. Therefore, it would be of value to several groups to access an objective player tracking system that can depict performers' movement patterns with accuracy, reliability and efficiency. However, despite considerable advances in technology, issues in automatic player tracking still present many challenges.
This review evaluates existing published research on human motion tracking in sports. Given the recent and rapid developments in this field of study, an overview of available research would make an important and timely contribution to the literature. The following Internet search databases were used between January and March 2007 to identify relevant articles: PubMed, SportDiscus, Ovid and Web of Science. The search reference terms 'movement patterns', 'movement analysis', 'spatio-temporal trajectories', 'automatic tracking' and 'sport' were used. The inclusion criteria specified articles published between 1970 and 2007, in peer-reviewed scientific journals (n = 52), as papers published in scientific conference transactions or proceedings (n = 27), technical reports (n= 1) and unpublished dissertations (n = 2). Additionally, three peer-reviewed scientific journals were included on recommendation by an anonymous reviewer. Papers that tracked movement using accelerometry or magnetic field (e.g. Polhemus) technology were excluded. Further, a Google Internet search was conducted using the reference terms 'commercial tracking systems', 'motion analysis systems' and 'performance analysis systems' to obtain information about existing commercial systems used in the analysis of sporting activities.
This article is structured into three sections in a manner that is loosely chronological. The first section describes manual notational analysis methods that have largely involved the subjective monitoring of player activity. The second section, reviews automated tracking systems in sport that have emerged over the last 10 years and the exciting new possibilities for multiple, simultaneous person tracking. Finally, the strengths and weaknesses of some commercially available motion tracking systems that have been used for research and performance analysis purposes are discussed. The aim is to establish what visionbased tracking systems exist and to what extent they are able to record multiple player movement in different sports environments (i.e. outdoor and indoor sports).
1. Manual Vision-Based Tracking Systems: Notational Analysis in Sport
Notational analysis is often recommended to coaches as an inexpensive way of providing insight into the physiological and technical demands of sports activities, by recording and quantifying the player movement patterns that characterize skilled performance.'21 This involves the subjective quantification of individual player(s) movements by an investigator, and the frequency and timing of particular movements are then related to their relative success.'31 Notational analysis has been used to investigate the activity patterns and techniques in various sports, such as rugby,[2,4] squash,[3,5,6] badminton,[7] football,[8-12] netball,[13-15] basketball,[16,18] Australian Football League (AFL),[19] volleyball[20] and roller hockey.[21] This section provides an overview of methods and scope of notational analysis. Racket sports and football were chosen as our examples because they represent the most commonly analysed sports using notational analysis, they show the greatest technological developments and they demonstrate significant differences in player numbers, patterns of play and size of capture environment. For further examples of notational analysis in sport, the reader is recommended to consult Hughes and Franks.[22]
1.1 Racket Sports
Historically, researchers have used different procedures to carry out notational analysis, from simple observations with written notes to more complex systems using video and computers.[7] Sanderson and Way[6] first developed a hand notation method for sequential stroke analysis in squash using illustrative symbols to analyse 17 different strokes and court plans to gather accurate positional information. This method was later modified by Sanderson[5] and applied to match play to define winning and losing patterns of play in squash. Hughes et al.[23] introduced computerized movement tracking to squash, entering player position data with a digitizing pad and using custom-designed software to calculate distances travelled and velocity and acceleration time series. A more sophisticated notational system was developed by Hughes and Clarke[24] to study the effect of court surface on elite tennis strategy. Players' court positions, the time taken per shot and the type of event itself were recorded using a graphical user-interface, which was analysed post event by video recorder. This analysis provided both temporal and positional information as well as frequency distribution of shots and rally-ending conditions.[24]
Hong and colleagues[3] also used a notational analysis method in studying game strategies of 12 of the world's top ranked male squash players. Video cameras were placed at the back of the court, at ground level, to provide a clear view of the movement of the player and the ball. Several variables were recorded and subsequently analysed including the strokes played by each individual, the position from where the shot was played and where the ball landed.[3] Strokes were identified according to 13 pre-determined types and each return shot was classified into one of four categories (effective, ineffective, winning and losing) to determine the playing effectiveness of the competitors.
1.2 Association Football (Soccer)
Several approaches have been adopted to analyse the movement activities of football players. Notational analysis has been employed to explore the success of different playing formations, and in particular goal scoring patterns, and attacking moves.[25-29] Erdmann[30,31] developed a system for player tracking by televising the whole football pitch and gathering information on every moving object. A video camera, fitted with a wide-angle lens (130�), was placed in a stand above the pitch to allow a view of the full field of play. White tape marked 1-m intervals longitudinally across the field prior to the game for camera calibration and to control for the radial distortion that occurs with a wide-angle lens. The position of the white tape was then transferred from a TV monitor onto a transparent foil to form a semi-permanent reference grid. Another foil was used to manually record the positional information about each player of the field. The resulting displacement, velocity and acceleration were calculated by measuring the relative displacement of the two foils.[31]
Ali and Farrally[32] were able to identify and distinguish typical patterns of play for three professional teams, two club sides and an international team. Data were collected by an experimenter using 20 shorthand symbols that were representative of common game features. During matches, the movement of the ball was manually recorded in real-time on diagrammatic sheets that were later converted onto a 2-dimensional (2-D) coordinate grid (figure 1).
Similarly, Yamanaka and colleagues[25] and Garganta et al.,[28] performed investigations into patterns of play associated with teams from different countries. Yamanaka et al.,[25] used a computerized notational analysis method for eight games in the Asian Qualifying tournament for the 1994 World Cup. It was observed that Japan's playing strategies were more predictable than those of many other countries.
Garganta et al.[28] analysed goal scoring patterns from video recordings of five successful European football teams from different countries (Barcelona, Porto, Bayern Munich, A.C. Milan, and Paris Saint Germain), using a hand notation system. Data were collected from 44 matches (including 104 goals), with set plays filtered out and key features recorded for each scoring movement. It was concluded that the teams exhibited similar player movement patterns before scoring goals, independent of the number of matches observed, match difficulty, or the teams' style of play.[28] Although the direct style of play seemed to be the most successful, it was also noted that the efficiency of the direct style was dependent on the team's ability to change and adapt to the rhythm of the game and vary their attacking methods to outplay the opposition.
A video-based notation analysis system was also used to conduct a time-motion analysis of football players in the 1994 World Cup.[33] The input data from each of the matches included the time taken for each action with the ball, spatial locations of players as ? and y coordinates, selected qualitative manoeuvre variables, stops in play, injury time, and time in possession of the ball for each player.[33] Brazil demonstrated the most successful attacking plays, the most scoring opportunities, and the most shots for scoring goals.[29] A breakdown of their attacking manoeuvres showed more frequent passing and a more free style of play, snorter runs with the ball, more shots after receiving the ball and more direct free kicks than their opponents.[33]
1.3 Limitations and Reliability of Manual Notational Analysis Techniques
A potential limitation of notational methods described so far is the reliability of the data entry procedure, or the researcher's ability to reproduce the observed value when the measure is repeated.[2] Inter-observer consistency is considered crucial in establishing the reliability of motion analysis systems where the total time, frequency and mean duration of movements can exhibit large variations. Blomqvist et al.[7] investigated the reliability of a notational analysis system in badminton. Two matches were analysed post-event by three trained observers who coded the matches twice using SAGE Game Manager(TM) for Badminton software (Newcastle, England). The results highlighted the good reliability and validity of the analysis system for evaluating playing time, player position and the type, quantity and length of shots, but showed less reliability in relation to shot execution.[7]
The reliability of notational analysis in squash was evaluated by comparing the results in a repeated analysis.[3] The categorization of the 13 squash shot types and the classification of winning and losing shots found no errors. An error of 1.85% was found when classifying shots as effective or ineffective, a value that was considered acceptable.[3]
Similarly, Duthie et al.[2] analysed the reliability of video-based time-motion analysis with ten rugby union players who were individually tracked during the 2001-2 Super 12 competitions in New Zealand. The footage was then analysed by a single researcher on two occasions, 1 month apart. The test-retest reliability was quantified as the typical error of measurement (TEM) and rated as good, moderate or poor. The total time spent in the movement categories (walking, jogging, striding, sprinting, static exertion and being stationary) had moderate to poor reliability (5.8-11.1% TEM). The frequency of individual movements had good to poor reliability (4.3-13.6% TEM), while the mean duration of individual movements had moderate reliability (7.1-9.3% TEM). It was concluded that the time motion analysis system was a moderately reliable evaluation tool for examining movement patterns in competitive rugby.[2]
The laborious process and extensive time requirements involved with manually tracking multiple players in team games has meant that researchers have typically avoided trying to analyse all the available players, and instead have focused only on the movements of players involved in active play. In one exception, Franks and Goodman (cited in Hughes et al.[23]) attempted to analyse player movements in an entire team of football players. Unfortunately, the analysis was presented incomplete because of difficulties with identifying players from the distance required to film the entire pitch, and the time needed to complete the digitization process.[23]
Investigations by Docherty et al.[4] and Kingman and Dyson[21] alternately monitored players from two teams in rugby and roller hockey, respectively. Player movements and activities were manually coded into standard categories and general movement patterns were analysed individually. However, while both of these investigations involved the filming of multiple players, no information was provided regarding interactions between players and team movement patterns.
In summary, early efforts at tracking player motion by methods of notational analysis have typically involved a range of data collection techniques from live observation to post-event video analysis. Notational analysis has emerged as a field of study involving observer(s) subjectively categorizing player movement patterns and relating this information to performance effectiveness. Due to the considerable time required to manually collect and analyse such data, research has tended to focus only on small numbers of players within pre-defined playing areas or on the player in possession of the ball and those in close proximity. Whilst notational analysis is a convenient, practical and typically inexpensive procedure, the validity and reliability of the process can vary depending on a number of factors including how many observers are used, their experience, and the quality of their viewing perspective.[22]
2. Automated Vision -Based Tracking Systems
Unlike the manual visual tracking systems described in section 1, automatic motion tracking does not require human operators to locate manually and continually record the position of the tracked object. Theoretically, the process required for automated tracking involves a number of steps. The initial stage of automatic tracking is the detection of moving targets in the captured images. These images must then be processed and segmented before shape models are fitted and the foreground shapes are extracted from the overall image. Further filtering and tracking methods are then used to predict the future locations and shapes of the tracked object.[34] Whilst some commercial online motion analysis systems (e.g. MAC, Santa Rosa, CA, USA) alleviate the requirements of manual tracking, to our knowledge a completely automated system for sports performance is not yet commercially available. Automatic motion tracking has been used successfully in other domains outside of elite sport performance, notably for surveillance in the military and security industry where automatic initialization of moving objects is achievable because identification of the objects is not necessary. Human motion analysis, particularly automated motion analysis, is a broad field of study.[35] For this reason we will briefly discuss automatic tracking from a surveillance perspective, for background and context, but the focus of this section is on systems that are currently used for sports analysis or those that have potential to be used to measure athletic performance. For a more detailed overview of developments in computer image processing see Poppe[36] or Moeslund and colleagues.[37,38]
2.1 Surveillance Tracking Systems
Tracking human motion in an indoor environment is of interest in applications of surveillance particularly in security-sensitive areas such as, banks, borders, airports and department stores.[38-40] As well as the obvious security applications, smart surveillance has also been proposed to measure traffic flow, monitor pedestrian congestion and behaviour and compile consumer demographics in shopping malls.[39] The monitoring of these activities requires the viewing system to be able to follow the image of the tracked object(s) within a large volume of space over a long period of time.[38]
Human motion analysis in sport presents considerable difficulties to potential surveillance systems due to large variations in human motion and appearance, camera viewpoints and environment settings.[36] There have been many attempts at human motion detection, with varying degrees of success. Human body models that were initially described in 2-D have now evolved into highly articulated 3-D models. Deterministic linear tracking has been replaced by samplingbased tracking frameworks that sample according to some cost function, and the role of machine learning is playing an increasingly important role in human motion analysis.[36] Recent contributions have addressed previously limiting assumptions, and some systems can now cope with natural outdoor scenes and operate on long sequences of video containing multiple interacting and occluding people.[36,39,41] It has been suggested that these developments are due to advancement in segmentation algorithms.[35,36,38] Other explanations suggest improved modelbased pose estimation or in the representation and interpretation of actions and behaviour.[36-38]
There are a number of motion-sensing devices with surveillance applications that do not rely upon video technology (e.g. infrared, magnetic systems, ultrasonic, IEEE 802.11 radio frequency-based systems and radio frequency identification [RFID]). However, the purpose of surveillance systems does not typically demand a high level of spatio-temporal tracking precision, as long as the movements of individuals can be tracked, their identity, for surveillance purposes, is not usually necessary.[42] It is necessary in sport, however, to track individuals and know when errors in player identification occur, which may pose serious analytical problems. Additionally, the characteristics and complexity of player activities in many sports, means that the frequency of player occlusion behind other players and officials pose serious problems for accurate uninterrupted tracking. Together, these factors currently prevent the direct application of surveillance systems to sporting environments.
2.2 Outdoor Sports Tracking Systems
This section focuses on tracking systems that were designed to capture human motion performed in an outdoor environment. Outdoor tracking systems differ to those developed for indoor sports, as they have larger capture areas, more interacting players and variable lighting conditions. Traditionally, this has meant that analysis systems have needed multiple cameras positioned around stadia and have focused on the game's 'active play'.
One of the first investigations into video-based automatic tracking of athletes involved the development and application of what has been called the 'closed-world' to an outdoor sports environment.[43] A closed-world was defined as a spatio-temporal region in which all identifiable objects present, such as players and boundary lines, were coded manually and each display pixel was allocated to one of those objects. This information enabled context-specific features to be dynamically selected as the basis for tracking. Intille and Bobick'431 define a context-specific feature as one that has been chosen, based upon its context, to maximize the chance of successful tracking between frames. In their analysis, two consecutive frames from television footage were subtracted to detect the players. The area around these players was then expanded to form a closedworld relative to the selected and relevant players. This closed-world information was then used to match the visual characteristics of each player in future frames. Indile and Bobick[43] demonstrated the use of contextual knowledge and its application to closed-world football player tracking and assert its potential to work more effectively than common template matching in tracking isolated players. However, the authors acknowledge that the method is somewhat limited by its inability to accurately track more than one player.[43]
Scene event analysis using image processing is becoming increasingly common. For example, tracking of players/athletes makes strategy analysis and scene recovery possible and can be achieved using many different methods.[12,44-48] Several studies have used computer vision processing, which groups together image pixels to form blob-like entities based on proximity and visual appearance. The blob representation was developed as a method of extracting compact and structurally meaningful descriptions of multispectral imagery.[46] Spatial coordinates are added to the spectral component of the image forming feature vectors for each pixel. These coordinates are clustered so image properties (colour and spatial similarity) combine and form coherent connected regions or 'blobs'. Tracking of the blobs' location on a playing surface is then carried out using motion features instead of structural features.[12]
A fully automatic and computationally efficient framework for analysis and summarization of football videos was proposed by Ekin et al.'491 using cinematic and object features. The framework included some novel low-level football video processing algorithms including dominant colour region detection, robust shot boundary detection, and shot classification. High-level algorithms were also incorporated for goal detection, referee detection and penalty box detection. The system was capable of three types of summary outputs: slow motion segments, goals scored, and slow motion segments classified by object features.[49] The results showed that this system was successful at detecting real-time events (using cinematic features) and filtering slow-motion replays (using object-based features) to allow semantic labelling. This automatic video analysis and summarization method, however, is currently only capable of detecting events (i.e. goal scoring) and is not able to provide more detailed information regarding player(s) location throughout the match.[49]
Qi et al.[50] presented a multiple-player tracking framework using low resolution video of various sports, at eight frames per second. Tests showed that the algorithm could successfully deal with 2 minutes of non-occluded game situations and occluded situations between members of opposite teams. Motion model and shape information with a data association method could track objects when no occlusions occurred. However, when occlusion between members of the same team occurred, the similar appearance of players made differentiation very difficult.[50]
Iwase and Saito[51,52] proposed a method for parallel tracking of all football players during games using multiple fixed cameras. From 25 people (22 players and 3 referees) tested, 14 of these could be tracked through a scene of 500 frames without problems. Each player image was provided with an identification number in the original video image frame. On occasions where the scene became too crowded and there were occlusions, identification numbers were replaced by numbers belonging to other players, allowing problems to be easily identified and easily corrected by manual intervention.
A recent investigation by Barros et al.[53] used an automatic tracking system (DVideo, Campinas, Brazil) to measure distances covered by football players. In each game, four digital cameras were fixed at the highest points of the stadiums, each covering approximately a quarter of the field, but in overlapping regions.[53] The trajectories of 112 different players were tracked in four games, although only the results of the players who participated in whole games were analysed. Situations where players were not tracked automatically only occurred when players were occluded and were corrected manually.[53]
Most investigations that have attempted to auto-track objects have used fixed camera placements and focused on a single moving object.[54,55] Alternative automatic tracking methods for analysis of football, using a moving camera were suggested by Utsumi et al.[54] and Araki et al.,[55] The investigation by Utsumi et al.[54] detected the object for tracking and each pixel was evaluated based on its colour rarity and its local edge property, which meant that player regions could then be detected and tracked. Players without occlusion were easily identified and the average sizes of the upper and lower halves of the 'player region' of the image were calculated separately to provide more accurate positional information. Players who were occluded by another player or object were separated using colour-based template matching. While the overall player detection rate was not sufficient to label specific players and precisely describe the game, the additional information on object shape and colour was shown to improve tracking ability. Utsumi et al.[54] suggested that further improvement was required to automatically detect occluded players and that a more robust method of detecting changes in player's positions could accurately measure player movement patterns.[54]
Similarly, Araki et al.[55] proposed a method for real-time tracking from a moving camera image sequence using robust statistics and active contour models. Detection of moving objects was achieved by cancelling the apparent motion of a static background and applying split and merge models that set initial contours and surround the moving object on the first frame. Although this method was shown to detect and track multiple moving objects in real-time, it could not accurately track dynamic objects that stop temporarily and frequently change direction as it is limited to continuously moving edges.[35,55]
2.2. 1 Limitations of Outdoor Automatic Motion Tracking Systems
Automatic tracking systems developed for outdoor applications appear to be restricted by several common limitations. The first of these is the inability to accurately track more than one player.[43,49,51,56] Problems with player tracking are often encountered during periods of player congestion, where players are frequently occluded by others, particularly players of the same team. Situations where multiple players cluster in restricted playing areas (e.g. penalty area during a corner in football) often encounter such problems and consequently are typically low in tracking accuracy.;45,53] Conversely, players who are largely isolated during the game, (e.g. football goalkeepers) are tracked with the most accuracy. It has been suggested that the problems associated with occlusion could be overcome by increasing the number of cameras, or changing the placement of the existing cameras.[53] In instances where players are lost or cannot be tracked the operator is usually required to correct errors in the player's path manually.[48]
Studies that use computer vision processing to create blob-like entities for player tracking also report occlusion as the most significant limitation.[12,44-48] This is primarily because of the camera placement, which is usually on the side of the pitch at ground level and causes inherent difficulties when multiple players surrounded the ball.
Finally, there are limitations associated with using a moving camera. Utsumi et al.[54] and Araki et al.[55] presented automatic tracking methods for analysis of football using a moving camera, and reported that the overall player detection rate was not sufficient to label specific players. They also suggest that further improvement is required to automatically detect occluded players and that a more robust method of detecting changes in player's positions is needed to accurately measure player movement patterns.[54,55] Furthermore, the use of a moving camera increases the complexity of player tracking as both the objects and the sensor are moving independently with respect to the reference coordinate frame, requiring continuous calibration.[57-60] It appears that the moving camera approach is limited to predominately planar activities like walking, running, skating or skiing where athletes can remain within the calibrated space and the relationship between cameras can remain constant.[55]
2.3 Indoor Sports Tracking Systems
This section focuses solely on tracking systems that were designed to capture human motion performed inside a building. Although automatic tracking and image processing ability is developing rapidly, its application to sport has undoubtedly been hampered because of inadequate video and computational facilities available at sports venues. The complex nature of movement inherent to many physical activities also represents a significant hurdle to overcome. In many sports, such as basketball, handball and netball, participants exhibit quick and agile movements, with many unpredictable changes in direction and also frequent collisions with other players. Each of these characteristics of player behaviour violate the assumptions of smooth movement on which computer tracking algorithms are typically based.[61] As previously discussed in section 1, notational approaches have partially addressed the motion acquisition problem through manual tracking of single players for each recorded field with observational sheets or computerized digitizing pads.[30,31,57,61]
Segen and Pingali[62] developed a fixed camera system for transforming recorded video footage of people moving through an environment into spatio-temporal coordinates. Feature points were identified in each video frame and these points were matched across frames to develop feature 'paths'. The trajectories consisted of spatial (vertical and horizontal) and temporal coordinates and data were collected continuously as the individual moved through the environment in real-time.[62] This technique was developed further by Pers and Kovacic[59,61] who devised a computer vision system capable of tracking multiple players with a known accuracy, in real time and with minimal human operator intervention.[59,61] The system presents spatio-temporal trajectories for five players in a handball match and allows for further context-specific analysis (see figure 2).
The automated tracking method of Pers and Kovacic'59�61�631 required the players to remain in the field of view for the duration of the capture period. Theoretically this requirement can be achieved though camera panning, a moving camera system or appropriate camera placement. Pers and Kovacic[59,61] decided to place two static cameras with 103� wide-angle lenses at a height of 10 m above a handball playing court. Footage of the players was captured at 50 fields per second with an image resolution of 384 ? 288 pixels (see setup in figure 3).
Motion detection, template and colour-based tracking algorithms were tested by Pers and colleagues[59,61,63,64] to determine the most effective method for following player movements. Each video frame in a sequence was subtracted from the reference frame (image of an empty court) leaving the difference image. The image was then thresholded and filtered, resulting in the production of 'blobs' (that correspond to players), which were then analysed. The frequent collisions between players in handball required constant human intervention as sudden changes in player position occur when the tracker confuses two players during collisions.[57,59,61,63]
The colour identification algorithm was shown to be successful even when the colour of the player is changed with signal distortion during tape recording.[59,61] The algorithm searched for the pixel most similar to the recorded colour of the player, in a limited area surrounding their previous known location. However, a highly diverse background with coloured areas (court surface advertisements) could potentially cause differentiation problems for the system if the background colour is similar to the players' kit.[59,61,63] The colour identification method is not always successful because the number of pixels required to construct the player images are often too great, and cannot be provided by the captured footage.[59,61,63]
Vuckovic et al.[65] used a motion detection tracking algorithm to compare the work rates of international and national squash players during matches. Tracking of the players was initiated manually, by clicking near each player's centre of gravity, and then automated at 15-20 frames per second. Video images of player positions were smoothed using the Gaussian smoothing kernel to reduce measuring errors in path length and players' velocity.[65] Information relative to the player's movement patterns was then presented graphically as spatio-temporal trajectories, which showed that typical x-shaped motion patterns (see figure 4) of varied magnitudes were executed by both groups of athletes during rallies. The similarity of the player trajectories suggest that common play tactics are employed by players of different abilities. However, significantly shorter diagonals in motion trajectory can be observed for the national player, indicating national players play with less accuracy, which directly influences the distance covered and work-rate of opposing players.[65]
Kristan et al.[66] used a colour-based tracking algorithm to study the interaction of multiple targets during team sports. Intille and Bobick's[43] closed-world concept was applied to model the scene context, because of the sportspecific task constraints. This concept can be applied to a sporting context on the assumption that the camera above the playing field is static and its optical axis is perpendicular to the floor.[66] Furthermore, the texture of the players was known and varied throughout the game due to non-uniform lighting conditions, the background and variations in the players pose.
Needham and Boyle[67] presented an alternative method to those proposed by Pers and colleagues[57,59,61,63] by using a condensationbased approach. They tracked the movements of football players from a single camera on an indoor court. Each player automatically tracked was independently fitted to a model and the sampling probability was calculated as a function of the fitness score of each player. To evaluate the automated tracking system, a sequence was manually tracked four times and each resulting trajectory was compared with the other trajectories. The automated system showed a mean difference of 2.5 m to the manually digitized trajectories. Needham and Boyle[67] were able to make improvements to the tracking system using Kaiman filtering, preventing it from switching between players, and consequently reducing the mean error to 1.16 m (where an error of up to 0.5 m was considered acceptable for handtracked data).
2.3.1 Limitations and Reliability of Indoor Automatic Motion Tracking Systems
Pers and colleagues[63,68] validated the use of their automated indoor tracking system by reporting the errors and mistakes associated with player tracking. First, the most significant sources of error capable of influencing accuracy were identified as: movement of player extremities, video tape noise; quantization error; imperfect camera calibration; operator mistakes. Error related to the movement of player extremities was attributed to viewing a 3-D situation through a 2-D medium. Ideally, player positions would change only when the player walked or ran from one position to another. However, because 2-D movement is assumed, changes in position due to vertical movement or movement of the extremities are not accounted for and can result in error.[63,68] Quantization error relates to the effect that radial distortion has on the input images, which is problematic particularly at boundary regions (see figure 5).
To verify the ability of the tracking system, Pers et al.[63,68] conducted four experiments to investigate the system's accuracy. Separate experiments were designed to look at (i) the influence of 'jitter' (short trajectory length on stationary players) in the output trajectories; (ii) the influence of filtering on the accuracy of player positioning; (iii) the influence of filtering on calculated player velocity; and (iv) a comparison between the multi-object tracking system and a manual single-object, 3-D tracking system (Ariel Performance Analysis System [APAS]).[63,66,68] Five players participated in each experiment, initially they all stood still in various positions on the court before performing 180 seconds of physical activities.[63] Trajectory smoothing was found to significantly reduce the errors in player velocity and distance covered, but had minimal influence on player position error.[63,66] The second experiment found that heavy smoothing hid rapid changes in player trajectories.[63] In the third experiment, velocity errors were calculated from the length of the player's circular trajectory and the time taken to complete the path. Average error ranged from 0.07 to 0.35m/sec depending on the size of the smoothing kernel used.[63] In the final experiment, the participants ran around three markers in one corner of the court in a 3.4-second (25-Hz) video sequence. The results were consistent with previous experiments with a 0.36-m error in position. However a 0.5 m/sec error was reported in velocity, highlighting the important difference in capture detail between the systems.[63]
Undoubtedly both the multiple and singleobject tracking systems describe the actual motion of the player. However, while the APAS system provided accurate information of the player's acceleration from their centre of mass and considered the movement occurring at the extremities, the multi-object system provided more global information, accurately showing acceleration and changes in direction on a larger scale. This information was considered potentially more relevant to game play and necessary to identify interactions in team game situations.[63,68]
3. Commercially Available Vision-Based Analysis Systems
3.1 Semi-Automatic/Online Systems
Several sports performance analysis systems now exist on the commercial market providing different features of player tracking capacity.[34] In this section, examples of these systems will be introduced and their common limitations will be summarized. Basic video analysis systems such as Dartfish(TM) (Fribourg, Switzerland), Digital Soccer(TM) (Italy), Game Breaker(TM) (Sportstec, Australia) and Utilius VS (CCC-Software, Germany) are designed to provide extrinsic feedback information to coaches and athletes to enhance performance and learning. Systems such as Game Breaker and Utilius VS do not perform actual analysis, but edit video recordings to create movie presentations and time-coded video clips to allow a frame by frame analysis of individual movements and skills. They do not currently provide information about movement patterns or player interactions. Systems that require players to wear special tracking devices (e.g. TRAKUS(TM) TKS Inc., MA, USA) are often unsuitable for competition because of regulations and the risk to player safety. The TRAKUS(TM) software is an example of a customized real-time system based on microwave receivers that analyse signals emitted from transmitters worn by the players to various interactive media.[68] To date, the TRAKUS(TM) system has been applied in sports such as ice hockey, golf and motor sports. The system provides data on location, speed, acceleration endurance and intensity. It is also capable of reconstructing sequences of events or plays and creates views of the action from multiple angles.[69]
SoccerMan(TM) (Berne, Switzerland) is a football game-reconstruction system that is designed to generate a descriptive animated 3-D scene from a given video sequence. This description includes information on the ground texture as well as a 2-D description of the players. This 3-D scene can then be examined from any virtual viewpoint with a 3-D viewer, assisting in the verification of referee decisions, television game analysis, training support, teamwork evaluation and the classification of football scenes.[67,70] Initially, the cameras are calibrated to extract background and player information from the video sequence and the system assumes that the camera remains in the same position throughout the match.
LucentVision provides a product for analysing and visually presenting tennis games.[69] The features of the product include presence maps that highlight where players spend most of their time during a match, virtual replays, where the ball can be viewed from any position and a variety of numerical statistics (velocity, distance). This system is a real-time networked visual information system that archives sports action using visual processing. Computer vision techniques are used to track each player's movement including the distance travelled and speed.[69]
TRAKPERFORMANCE(TM) (Sportstec, Australia) is a video-based system developed by Sportstec that provides player tracking information from games in real-time. The system relies upon operators to manually track player positions frame by frame before providing information such as distances travelled, speed breakdowns of individual players, work to rest ratios and positionspecific information. It does not require players to wear additional transmitting devices, which means it is also capable of tracking opposing teams or individuals. This programme is marketed to be accurate to within 5% of true values and the breakdown of speeds is accurate to within �10% (Sportstec, Australia).
Online motion analysis systems developed by companies such as Motion Analysis Corporation (MAC, Santa Rosa, CA, USA) capture movement in digital format in real time. Such systems are often used in animation production and for industrial measurement and control. MAC have developed a powerful 28-camera Eagle-4 Digital System that has the ability to capture ten performers at 60 frames per second in a 40 � 40 � 15 feet (12�12�5 m) capture area. Performers are prepared in advance of tracking with infrared light reflecting markers fixed to various body segments. The Eagle-4 cameras capture the motion of the markers and Motion Analysis EVaRT(TM) software identifies the actors' movements in realtime. Some examples of other companies that produce similar motion analysis systems include Vicon(TM) (Vicon Motion Systems, Oxford, UK), BTS Elite(TM) (Milan, Italy) and Qualysis Motion Capture Systems(TM) (Gothenburg, Sweden).
Pfinder(TM) (Cambridge, MA, USA) is a realtime digital video system for tracking and interpreting human movement.[46] Pfinder(TM) uses a multi-class statistical model of colour and shape to segment the video image and obtain a 2-D representation of head and hands in a wide range of viewing conditions.[46] These 2-D representations are 'blobs' or clusters of pixels with similar colour and spatial image properties. The 'blobs' are then automatically tracked by software that uses a background subtraction technique to determine positional changes.[12,44,46]
Amisco(TM) (Sport Universal, Nice, France) provides a passive (marker-less) tracking system that measures all moving objects on the football pitch. Sampling at 25 Hz, real-time information is provided from the multiple capture systems installed around the stadium. However, operators are required to note all events, such as fouls, off sides and cautions that occur throughout the game.[69]
Similarly, ProZone(TM) (West Yorkshire, England) is another video tracking system designed for the analysis of football that requires an elaborate multi-camera system that is custom-fitted at sports stadia. The footage from these cameras is used by the Prozone(TM) company to track player positions manually over quadrants that make up the total playing area. The positional information of each player is then used to provide valuable statistics to the coaching staff.[34,69] Di Salvo et al.[71] attempted to validate the Prozone(TM) measurement system while analysing displacement and velocity on a football field. Participants performed a series of runs while being tracked by the Prozone(TM) system in different zones of the Old Trafford and Reebok Stadiums. The Prozone(TM) data were compared with timing gate data collected for the same runs.[71] The results indicated that the average speed recorded by the Prozone(TM) system during the paced runs of 60 m and 50 m showed an excellent correlation (r= 0.999) with the average speed measured by timing gates. Furthermore, the maximal 15-m sprint showed a correlation (r= 0.970). Therefore, it was concluded that Prozone(TM) represented a valid motion analysis system for analysing movement patterns of football players on a football field.[71]
3.2 Limitations of Commercial Tracking Systems
Clearly the commercial tracking systems described in this section have provided enhanced motion tracking capability in recent times. Such systems typically operate though highspeed digital video or infra-red light-detecting cameras. However, several of the systems are limited by some common factors. For example, the main limitation of the SoccerMan(TM), TRAKORMANCE(TM) and ProZone(TM) systems is that significant manual intervention is required to accurately track the players. Thus, similar to notational analysis, it can be assumed that the accuracy of the positional information generated by these two systems will be dependent upon the training and experience of the observers. The MAC system is relatively expensive, has a limited capture volume, and the infra-red light-detecting cameras cannot be used for outdoor sports. The Pfinder(TM) tracking system is limited by its reliance on dynamic scenes and the assumption that only one person occupies a space at any one time, preventing multipleobject tracking.[46]
A number of commercial motion tracking systems are now available with reasonably high levels of accuracy and reliability. The tracking software that accompanies some of these systems has developed to the extent now that much of the information can be encoded in real-time (online) meaning that reduced processing times are required. However, commercial tracking systems still require significant operator intervention to process the data after capture. Whilst several systems are expensive, purchase costs have decreased substantially in the last 10 years or so. The main limitations of the systems described in this section are the restricted capture environments that can be used and the necessity for individuals to wear tracking devices. Such limitations can prevent their usage in a number of competitive sports situations.
4. Conclusions
The increasing capacity of digital technology to collect, manage and organize video images has made it possible to enhance existing sportspecific analytical procedures. A variety of systems and methods have been employed to analyse the motion of athletes during sports where the movements vary in duration, field position and surface, speed, direction technique and tactics.[3-5,13,25.30,32,45.53,67,70,72-74] The traditional methods of manual data entry, using either hand or computer, can be lengthy and monotonous. Codes that are assigned to different actions or player's position if not designed to provide some meaning to the operator, often results in the process being inefficient and inaccurate.[22] Video cameras linked with computers and automated analysis software provide a more sophisticated approach to motion analysis, but such systems have also experienced a number of practical limitations. For example, either the field of view of the camera has been too small to include the entire playing area or the system requires athletes to wear joint markers and move within a pre-set, restrictive calibration volume.[23,73]
Notational analysis linked to a computerized coding system enables quick and immediate analysis of the video images but fast feedback is restricted to simple kinematic and temporal data. Such techniques are also vulnerable to inaccuracies within the qualitative analysis of the data collected and interpreted.[2,23,75] However, despite these limitations and advances in the field of image processing and automatic tracking, notational analysis systems remain an important sports analysis tool. While semi-automatic or automatic computerized systems provide objective measurements about player movements, notational analysis is still required to compliment these new systems and provide event-specific information about player activities such as, jumps, throws, shots and fouls.
Tracking objects that take part in sporting events is inherently difficult because of athletes' interaction with other athletes and hence the unpredictability of movement.[45,62] The development of an automated system that collects uninterrupted footage of games and provides accurate positional information would be desirable and advantageous to sporting teams and coaches. Although the results of the studies reviewed have shown various levels of successful tracking, they also have significant limitations and most are restricted by their inability to track multiple, dynamic, frequently changing movements in a large environment.
The application of automated systems into sports is further restricted by the amount of equipment required to collect and process the video images. For example, multiple cameras are regularly used to film the movements of multiple players in football. Iwase and Saito[51] used 15 cameras to record the movements of 22 players and three referees. Barros and colleagues[53] used four cameras and needed five computers to segment and process the movements of football players over 10 hours. M�ller and Anido[45] used five cameras and needed 66 computers to process a short sequence of a professional football match. In addition to these large hardware demands, the sampling frequency of these systems is generally low. Iwase and Saito[51] sampled at 15 frames per second, and other systems sampled at a lower frequency (i.e. 7.5 Hz[3]or 8 H[50]). Furthermore, the duration of the playing sequences tracked is typically quite brief.[12] Iwase and Saito[51] recorded playing sequences of 500 frames (33.3 seconds), Utsumi et al.[54] tracked 30 seconds, Needham and Boyle[67] tracked 835 frames, whereas M�ller and Anido[45] analysed 4 minutes, 42 seconds and Qi et al.[50] tracked 2 minutes of real football game video.
The indoor systems tested by Pers and colleagues[57,59,61,63,66,68] had similar processing problems. The system described by Pers et al.[64] tracked 120 seconds of player activity at 10 frames per second and captured the image using one video camera. Another investigation by Pers et al.'6'1 used two cameras, captured sequences of 30 and 50 seconds and sampled at 25 Hz, processing this information on one computer running at 500 MHz. A third investigation also used two cameras, and tracked 14 players at a speed of 4.5 frames per second, requiring approximately one operator intervention per player during the 30-second sequence.'631
These requirements, of multiple cameras and multiple computers, mean that these systems are not easily installed next to playing courts or fields. The time required to process entire games also means that investigations are currently limited to only short playing sequences. Finally, sampling at low frequencies may influence the tracked data by reducing the number of frames processed and potentially minimizing the influence or effect of some movements recorded.
5. Future Research
The application of computer technology to the sporting domain presents an arduous research challenge because of the variable nature of human movement, the complexity of team ball sports and significant equipment requirements. The current challenge is to obtain appropriate video sequences that can robustly identify and label people over time, in a cluttered environment containing multiple interacting players in the absence of markers. Although systems that can achieve this goal are currently being developed for surveillance and security purposes, the problems associated with their application into the sporting environment is compounded by the quality of video capture, the relative size and occlusion frequency of people within that scene, and also changes in illumination.[35-37,61] Significant progress towards fully automated human motion tracking and reconstruction has occurred in recent years. These advancements have been largely in the initialization of model shape appearance and pose, tracking of multiple people in unstructured outdoor scenes, the reconstruction of human movement from multiple views, monocular motion reconstruction, pose estimation in natural scenes and the understanding of behaviour and actions.[35-37,61] Future developments may focus on body part detectors, body shape and movement of clothing, which would improve tracking reliability and pose estimation in clutter scenes.[38] Other areas of focus are on the development of reliable detection of people and behaviour from low quality imagery, and on the advancement of behaviour representation for dynamic scenes.[38,39] Further research is also required to reduce the computational costs while maintaining processing capacity and therefore increase processing speeds of automatic tracking.[35]
An automated system capable of capturing, analysing and quantifying player information in real-time would greatly enhance research in game analysis. For example, in indoor sports such as handball or basketball, detailed information regarding player's velocities, distances travelled, speed of attacking play, and frequencies related to the direction of passes, and the number of passes can be obtained. Furthermore, the ability of the system to track the movement of the ball and its interaction with players means that information regarding successful and unsuccessful passes between team mates can be quantified. The expression of players' activities as measurable physical quantities is important for developing position-specific training programmes for players, creating more informed team strategies and developing profiles on opposing teams.
As well as being of benefit to coaches and athletes, sport and exercise science researchers would benefit from information derived from an automated player tracking system. For example, sport psychologists could examine team interactions to relate theories of group dynamics to actual team performances. Sports biomechanists might use the information to create simulations of team behaviour and to model hypothetical game situations. Motor control researchers could examine the stability and variability of intraperson coordination as a function of different decisions made by key players. Clearly, a wealth of fascinating information related to multiple player interaction will soon become available that has the potential to enhance understanding of sports performance to a new level.
Acknowledgements
The authors would like to thank Professor Roger Bartlett, Kathryn Phillips and two anonymous reviewers for their help and advice whilst preparing the manuscript. No funding, no potential conflicts of interest.
[Reference]
References
1. Eom H, Schutz R. Transition play in team performance of volleyball: a log-linear analysis. Res Q Exerc Sport 1992; 63 (3): 261-9
2. Duthie G, Pyne D, Hooper S. The reliability of video based time motion analysis. J Hum Move Stud 2003; 44: 259-72
3. Hong Y, Robinson P, Chan W, et al. Notational analysis on game strategy used by the worlds top male squash players in international competition. Aust J Sei Med Sport 1996; 28 (1): 18-23
4. Docherty D, Wenger H, Neary P. Time-motion analysis related to the physiological demands of rugby. J Hum Move Stud 1988; 14: 269-77
5. Sanderson F. A notational system for analysing squash. Phys Educ Rev 1983; 6: 19-23
6. Sanderson F, Way K. The development of an objective method of game analysis in squash rackets. Br J Sports Med 1977; 11: 188
7. Blomqvist P, Luhtanen P, Laakso L. Validation of a notational analysis system in badminton. J Hum Move Stud 1998; 35: 137-50
8. Mayhew S, Wenger H. Time motion analysis of professional soccer. J Hum Move Stud 1985; 11: 49-52
9. Grehaigne J. Game systems in soccer from the point of view of coverage of space. Sci Football 1988: 1: 316-21
10. Grehaigne J, Bouthier D, David B. Dynamic-system analysis of opponent relationships in collective actions in soccer. J Sports Sci 1997; 15: 137-49
11. Hughes M, Clarke A. Computerised notational analysis of rugby union to examine the effects of law changes upon patterns of play by international teams, [abstract]. J Sports Sci 1994; 12(2): 180
12. Figueroa P, Leite N, Barros R. Tracking soccer players aiming their kinematical motion analysis. Comput Vision Image Understand 2006; 101: 122-35
13. Steele J, Chad K. Relationship between movement patterns performed in match play and in training by skilled netball players. J Hum Move Stud 1991; 20: 249-78
14. Palmer C, Hughes M, Borrie A. A comparitive study of the centre pass patterns of play of successful and non successful international netball teams, [abstract]. J Sports Sci 1994; 12 (2): 181
15. Loughran B, O'Donoghue P. Time motion analysis of work rate in club netball. J Hum Move Stud 1999; 36: 37-50
16. Miller S, Bartlett R. Notational analysis of the physical demands ofbasketball. [abstract]. J Sports Sci 1994; 12(2): 181
17. Mendes L, Janeira M, editors. Basketball performance: multivariate study in Portuguese professional male basketball teams. Cardiff: UWIC, 2001
18. Tavares F, Gomes N. The offensive process in basketball: a study in high performance junior teams. Int J Performance Analysis 2003; 3 (1): 34-9
19. Dawson B, Hopkinson R, Appleby B, et al. Player movement patterns and game activities in the Australian Football League (AFL). J Sci Med Sport 2004; 7 (3): 278-91
20. Hughes M, Daniel R. Playing patterns in elite and non-elite volleyball. Int J Performance Analysis Sport 2003; 30 (1): 50-6
21. Kingman J, Dyson R. Player position, match half and score effects on the time and motion characteristics of roller hockey match play. J Hum Move Stud 1997; 33: 15-29
22. Hughes M, Franks I. Notational analysis of sport: systems for better coaching and performance in sport. 2nd ed. London: Routledge, 2004
23. Hughes M, Franks I, Nagelkerke P. A video system for the quantative motion analysis of athletes in competitive sport. J Hum Move Stud 1989; 17: 217-27
24. Hughes M, Clarke S. Surface effect on elite tennis strategy. Sci Racket Sports 1995; 1: 272-7
25. Yamanaka K, Liang D, Hughes M. An analysis of the playing patterns of the Japan National team in the 1994 World Cup qualifying match for Asia. In: Reilly T, Bangsbo J, Hughes M, editors. Science and football III: proceedings of the Third World Congress of Science and Football; 1995 Apr 9-13; Cardiff. London: Taylor & Francis, 1997: 221-9
26. Sforza C, Dunani S, Mauro F, et al. Repeatability of the football penalty; a statistical evaluation by the morphological variation analysis. In: Reilly T, Bangsbo J, Hughes M, editors. Science and football III: proceedings of the Third World Congress of Science and Football; 1995 Apr 9-13; Cardiff. London: Taylor & Francis, 1997: 240-6
27. Sforza C, Michielon G, Grassi G, et al. Bivariate analysis of the repeatability of football offensive schemes. In: Reilly T, Bangsbo J, Hughes M, editors. Science and football III: proceedings of the Third World Congress of Science and Football; 1995 Apr 9-13; Cardiff. London: Taylor & Francis, 1997: 233-40
28. Garganta J, Maia J, Basto F. Analysis of goal-scoring patterns of European top level soccer teams. In: Reilly T, Bangsbo J, Hughes M, editors. Science and football III: Proceedings of the Third World Congress of Science and Football; 1995 Apr 9-13; Cardiff. London: Taylor & Francis, 1997: 246-51
29. Grehaigne J, Bouthier D, David B. A method to analyse attacking moves in soccer. In: Reilly T, Bangsbo J, Hughes M, editors. Science and football III: proceedings of the Third World Congress of Science and Football; 1995 Apr 9-13; Cardiff. London: Taylor & Francis, 1997: 258-67
30. Erdmann W. Quantification of games: preliminary kinematic investigations in soccer. In: Reilly T, Clarys J, Stibbe A, editors. Science and football II. London: Taylor & Francis, 1991
31. Erdmann W. Gathering of kinematic data of sport event by televising the whole pitch and track. In: Rodano R, Ferrigno G, Santambrogio G, editors. Symposium of the International Society of Biomechanics in Sports. Milan: International Society of Biomechanics, 1992 Jun: 159-62
32. Ali A, Farrally M. An analysis of patterns of play in soccer. Sci Football 1990; 3: 37-44
33. Luhtanen P, Korhonen V, Ilkka A. A new notational analysis system with special reference to the comparison of Brazil and its opponents in the World Cup 2004. In: Reilly T, Bangsbo J, Hughes M, editors. Science and football III: proceedings of the Third World Congress of Science and Football; 1995 Apr 9-13; Cardiff. London: Taylor & Francis, 1997: 229-33
34. Needham C. Tracking and modelling of team game interactions [dissertation]. Leeds: University of Leeds, 2003
35. Aggarwal K, Cai Q. Human motion analysis: a review. Comput Vision Image Understand 1999; 73 (3): 428-40
36. Poppe R. Vision-based human motion analysis: an overview. Comput Vision Image Understand 2007; 108: 4-18
37. Moeslund T, Granum A. A survey of computer vision-based human motion capture. Comput Vision Image Understand 2001; 81 (3): 231-68
38. Moeslund T, Hilton A, Kruger V. A survey of advances in vision-based human motion capture and analysis. Comput Vision Image Understand 2006; 104: 90-126
39. Wang L, Hu W, Tan T. Recent developments in human motion analysis. Pattern Recognition 2003; 136: 585-601
40. Fujiyoshi H, Lipton A, Kanade T. Real time human motion analysis by image skeletonisation. IEICE Trans Inf Syst 2004;E87-DNo 1: 113-20
41. Hu W, Tan T, Wang L, et al. A survey on visual surveillance of object motion and behaviours. IEICE Trans Inf Syst Man Cybernetics 2004; 34 (3): 334-52
42. Ni L, Liu Y, Lau Y, et al. Indoor location sensing using active RFID. Wireless Networks 2004; 10: 701-10
43. Intille S, Bobick A. Visual tracking using closed worlds. MIT Media Laboratory Perceptual Computing Section, 1994. Technical report no. 294. Los Alamito (CA)
44. Bobick A, Davis J. The recognition of human movement using temporal templates. IEEE Transact Pattern Analysis Machine Intelligence 2001; 23 (3): 257-67
45. Muller B, Anido R. Distributed real-time soccer tracking. ACM 2nd International Workshop on Video Surveillance and Sensor networks conference proceedings. New York: ACM Press, 2004 Oct 15; 97-103
46. Wren C, Azarbayejani A, Darrell T, et al. Real-time tracking of the human body. IEEE Transact Pattern Analysis Machine Intelligence 1997; 19 (7): 780-5
47. Figueroa P, Leite N, Barros R. Background recovering in outdoor image sequences: an example of soccer players segmentation. Image Vision Comput 2006; 24: 363-74
48. Taki T, Hasegawa J, Fukumura T. Development of motion analysis system for quantitative evaluation of teamwork in soccer games. IEEE; International Conference Image Processing 1996; 3: 815-8
49. Ekin A, Tekalp M, Mehrotra R. Automatic soccer video analysis and summarization. IEEE Transact Image Process 2003; 12 (7): 796-807
50. Qi F, Luo Y, Hu D. Visual tracking of players through occlusions in low resolution. IASTED International Conference; 2004 Aug 23-25; Honolulu (HI), 375-80
51. Iwase S, Saito H. Parallel tracking of all soccer players by integrating detected positions in multiple view images 17th International Conference on Pattern Recognition; 2004 Aug 23-26; Cambridge, UK, 2004: 751-54
52. Iwase S, Saito H. Tracking soccer players based on homography among multiple views. Proceedings of SPIE; 2003 Visual Communications & Image Processing (VGP) VCIP 2003: 283-92
53. Barros R, Misuta M, Menezes R, et al. Analysis of the distances covered by first division Brazilian soccer players obtained with an automatic tracking method. J Sports Sci 2007; 6: 233-42
54. Utsumi O, Miura K, Ichiro I, et al. An object detection method for describing soccer games from video; 2002 Multimedia Expo 2002. ICME. Proceedings of the IEEE International Conference 2002; 1 : 45-8
55. Araki S, Matsuoka T, Yokoya N, et al. Realtime tracking of multiple moving object contours in a moving camera image sequence. IEICE Trans Inf Syst 2000; E83-D (7): 1583-91
56. Ohashi J, Miyagi O, Nagahama H, et al. Application of an analysis system evaluating intermittent activity during a soccer match. In: Spinks RaM, editor. Proceedings of Science and Football IV; 2002 Aug 26-29; Lausanne, Switzerland: Taylor & Francis, 2002: 329-33
57. Pers J, Kovacic S. Tracing people in sport: making use of partially controlled environment. Ljubljana: Faculty of Electrical Engineering, University of Ljubljana, 2001
58. Cai Q, Aggarwal K. Tracking human motion using multiple cameras. IEEE Proceedings ICPR 1996; Vienna, 1996: 68-72
59. Pers J, Kovacic S. Computer vision system for tracking players in sports games. First International Workshop on Image and Signal Processing and Analysis; 2000 Jun 14-15; Pula: IWISPA, 2000: 81-6
60. Lafontaine D, Lamontagne M. 3-D Kinematics using moving cameras. J Appi Biomech 2003; 19: 372-7
61. Pers J, Kovacic S. A system for tracking players in sports games by computer vision. Electrotechnical Rev 2000; 67 (5): 281-8
62. Segen J, Pingali S. A camera based system for tracking people in real time. IEEE Proceedings of 1 3th International Conference on Pattern Recognition ICPR; 1996: 63-7
63. Pers J, Bon M, Kovacic S, et al. Observation and analysis of large-scale human motion. Hum Move Sci 2002; 21: 295-311
64. Pers J, Vuckovic G, Kovacic S, et al. A low cost real time tracker of live sport events. Ljubljana: Faculty of Electrical Engineering, University of Ljubljana, 2001
65. Vuckovic G, Dezman B, Pers J, et al. Motion analysis of the international and national rank squash players. 4th International Symposium on Image and Signal Processing and Analysis (ISPA); 2005 Sep 15-17; IEEE, 2005, 334-8
66. Kristan M, Pers J, Perse M, et al. Multiple interacting targets tracking with application to team sports. 4th International Symposium on image signal processing and analysis; 2005 Sep; ISPA, 2005; 322-7
67. Needham C, Boyle R. Tracking multiple sports players through occlusion, congestion and scale. Proceedings British Machine Vision Conference; 2001 Sep 10-13; Manchester, 2001; 1:93-102
68. Pers J, Bon M, Kovacic S. Errors and mistakes in automated player tracking. Sixth Computer Vision Winter Workshop; 2001; Bled: 2001 Feb 7-9; IEEE, 2001: 25-36
69. Setterwall D. Computerised video analysis of football: technical and commercial possibilities for football coaching [dissertation]. Stockholm: CID, NADA, 2003
70. Bebie T, Bieri H. Reconstructing soccer games from video sequences. International Conference on Image Processing; 1998 Oct 4-7; Chicago (IL): IEEE, 1998: 898-902
71. Di Salvo V, Collins A, McNeill B, et al. Validation of Prozone: a new video-based performance analysis system. Int J Perform Analysis Sport 2006 Jun; 6 (1): 108-19
72. Cruz J, Tavares F, editors. Notational analysis of the offensive patterns in cadet basketball teams. Cardiff: UWIC, 2001
73. Hughes M, Bartlett R. Performance analysis. J Sports Sci 2002; 20: 735-7
74. Grehaigne J, Godbout P, Bouthier D. The foundations of tactics and strategy in team sports. J Teach Phys Educ 1999; 18: 159-74
75. Liebermann D, Katz L, Hughes M, et al. Advances in the application of information technology to sport performance. J Sports Sci 2002; 20 (10): 755-70
[Author Affiliation]
Sian Barris and Chris Button
Human Performance Centre, School of Physical Education, University of Otago, Dunedin, New Zealand
[Author Affiliation]
Correspondence: Stan Barris, School of Physical Education, University of Otago, Dunedin, New Zealand.
