Player tracking and heatmap generation from broadcasted football games

Abstract

Data analytics have become an integral part of sports. Gone are the days when player scouting agents would travel all over the world on the search for new talent. The introduction of tracking sensors and vision systems, made data analysis possible without even having to watch the game. Data analysts sift through thousands of time-stamped data to extract metrics that detail the performance of players and teams to gain insights into improving the results, preventing injuries, and increasing the revenue. However, obtaining the necessary data is quite a complex task. This work aims at developing an offline computer vision algorithm that takes broadcasted footage of an attacking or defending scenario as its input, and aims to extract the players’ positional data throughout the sequence of frames in which they are visible. The data is then visualised in the forms of a heat map and a tracking sequence. The implemented system is divided into four major modules: the camera homography estimation module, the player detector module, the player tracking module, and finally the statistics module. This dissertation investigates the problem of keeping track of the image player positions amidst all the challenges that make the problem much more complex. Players move in and out of the camera field-of-view when the camera is not static. Every player needs to be detected and tracked in a sequence of frames, by a unique label. The tracker is expected to keep those labels in case the detection of the players is lost, and then found again after a number of frames, or during set-pieces where players congregate and occlude each other. The creation of a heatmap requires the computation of a camera homography matrix, which may be estimated by establishing a correspondence between the 3D field model and the 2D image. This process needs to be performed in every frame in which there are enough field lines visible to compare to the field model. Each module was then tested separately under varying conditions. The results show that the camera calibration module performs reliably when the weather conditions are favorable, and there are enough field lines visible in the image frame. The player detection module and the player tracking module proved to be very robust, and it was shown that it’s worth using Convolutional Neural Networks for such an application.