|
Tracking and modelling of team game interactions |
| PhD project - Chris Needham 1999-2003 |
Sport is a rich and diverse domain. It captivates the minds of the general public and sports are followed by millions of people around the world. To the computer vision researcher, it provides an abundant source of captivating footage from which to work. This thesis has two goals: to track the movements of sports (specifically football) players on an indoor pitch and to model the behaviour of the sports players within team games. Alongside these two goals, there are two primary motivations for this work. Firstly, the sports science industry is very interested in being able to know how much ground athletes have covered, and how quickly they have moved, during the course of a game. This information would allow more specific training to be designed to suit individual players. Secondly, the analysis of positional data from such a system to identify team game interactions is a fascinating research area. Sports games that involve two teams of players provide a rich environment for modelling cooperative, collaborative and adversarial actions of individuals and for modelling the behaviour of the teams as a whole. FA Premiership clubs are becoming more aware of the possibilities of team and individual performance analysis of football matches.
 |
 |
| Football players well spread out over the pitch | Football players: busy, occluded scenes |
The soccer domain exhibits many challenging aspects: the size of the pitch means that the resolution of an image of the game varies greatly between the nearest and the furthest parts of the pitch; sports games are busy areas; sports players' shapes vary significantly, often in short periods; and the players move at variable speeds, often suddenly changing direction, which makes their movements hard to predict. In addition, environmental variations can be considerable, even indoors. Players cast shadows on the floor, and the lighting varies and also produces some areas of `glare' at the camera end of the pitch.
A framework for multi-object tracking, using a condensation based approach is developed. Each player being tracked is independently fitted to a model, and the sampling probability for the group of samples is calculated as a function of the fitness score of each player. Image perspective effects are overcome through image transformation and tracking in the ground plane rather than in the image plane.
 |
 |
| Tracked 5-a-side football footage | |
[Example footage MPEG (4.3 MB)] [A tracked sequence MPEG (1.9 MB)] [Tracked sequence Quicktime (3.4 MB)]
Needham, C J; Boyle, R D. Tracking multiple sports players through occlusion, congestion and scale in: Cootes, T & Taylor C (editors) British Machine Vision Conference 12th 2001, vol. 1, pp. 93-102 BMVA. 2001. gzip'd postscript | pdf
 |
| Subsampled soccer players |
A novel multi-resolution template kernel approach to locating soccer players is developed. A set of multi-resolution template kernels may be learned through unsupervised clustering of example convolution masks extracted from subsampled bounding boxes around example players. Application of these kernels to subsampled image regions provides a robust shape descriptor/object locator for a range of image resolutions and object poses.
 |
 |
 |
 |
 |
| Five exemplar multi-resolution template kernels | ||||
There are many ways in which the performance of a computer vision system can be evaluated. For evaluating the performance of a `sports player tracker', the aim is to evaluate how well a tracker is able to determine the position of a target object. Few metrics exist for positional tracker evaluation. A set of metrics for analysing the similarity of trajectories have been pulled together and designed which allow spatial, temporal and spatio-temporal differences between trajectories to be analysed.
Needham, C J; Boyle, R D. Performance evaluation metrics and statistics for positional tracker evaluation in: Crowley, J L, Paiter, J H, Vincze, M & Paletta, L (editors) Computer Vision Systems Third International Conference, ICVS 2003, pp. 278-289 Springer-Verlag. 2003. abstract on Springer Website | PDF article from Springer
In addition to this positional tracker evaluation, it is also possible to evaluate the performance of each component a computer vision system. These evaluation can often provide an important insight into a methods strengths and weaknesses. The multi-resolution template kernels were evaluated in a number of ways. The sensitivity/robustness of a bounding box around a player has been investigated, w.r.t. a shift in the x-direction, a shift in the y-direction, a scaling in the x-direction, a scaling in the y-direction, and a fixed aspect-ratio scaling.
 |
 |
 |
 |
| Evaluation of the multi-resolution template kernels' robustness and sensitivity to changes in horizontal and vertical location | |
Capturing the behaviour of a set of sports players would allow many exciting activities to be performed; identifying tactics, predicting future movements, recognising set-plays, identifying teams, and evaluating teamwork. An advanced model of the interactions of players may allow for speculative tactical attacking moves to be simulated by a team and for probable responses for the defending team to be generated. A behaviour model to aid in predicting movements of sports players in a tracking system is the ultimate goal. From a cognitive viewpoint, a model which could learn the rules of the game would be very powerful indeed.
 |
 |
| Probability density estimation: A 100 component Gaussian mixture model of player position | Vector quantisation: 200 prototypes (position and velocity) |
This research produced a PhD thesis under the guidance of Prof Boyle. It is available to download (PDF).
Chris Needham