Extracting field hockey player coordinates using a single wide-angle camera

In elite level sport, coaches are always trying to develop tactics to better their opposition. In a team sport such as field hockey, a coach must consider both the strengths and weaknesses of both their own team and that of the opposition to develop an effective tactic. Previous work has shown that spatiotemporal coordinates of the players are a good indicator of team performance, yet the manual extraction of player coordinates is a laborious process that is impractical for a performance analyst. Subsequently, the key motivation of this work was to use a single camera to capture two-dimensional position information for all players on a field hockey pitch. The study developed an algorithm to automatically extract the coordinates of the players on a field hockey pitch using a single wide-angle camera. This is a non-trivial problem that requires: 1. Segmentation and classification of a set of players that are relatively small compared to the image size, and 2. Transformation from image coordinates to world coordinates, considering the effects of the lens distortion due to the wide-angle lens. Subsequently the algorithm addressed these two points in two sub-algorithms: Player Feature Extraction and Reconstruct World Points. Player Feature Extraction used background subtraction to segment player blob candidates in the frame. 61% of blobs in the dataset were correctly segmented, while a further 15% were over-segmented. Subsequently a Convolutional Neural Network was trained to classify the contents of blobs. The classification accuracy on the test set was 85.9%. This was used to eliminate non-player blobs and reform over-segmented blobs. The Reconstruct World Points sub-algorithm transformed the image coordinates into world coordinates. To do so the intrinsic and extrinsic parameters were estimated using planar camera calibration. Traditionally the extrinsic parameters are optimised by minimising the projection error of a set of control points; it was shown that this calibration method is sub-optimal due to the extreme camera pose. Instead the extrinsic parameters were estimated by minimising the world reconstruction error. For a 1:100 scale model the median reconstruction error was 0.0043 m and the distribution of errors had an interquartile range of 0.0025 m. The Acceptable Error Rate, the percentage of points that were reconstructed with less than 0.005 m of error, was found to be 63.5%. The overall accuracy of the algorithm was assessed using the precision and the recall. It found that players could be extracted within 1 m of their ground truth coordinates with a precision of 75% and a recall of 66%. This is a respective improvement of 20% and 16% improvement on the state-of-the-art. However it also found that the likelihood of extraction decreases the further a player is from the camera, reducing to close to zero in parts of the pitch furthest from the camera. These results suggest that the developed algorithm is unsuitable to identify player coordinates in the extreme regions of a full field hockey pitch; however this limitation may be overcome by using multiple collocated cameras focussed on different regions of the pitch. Equally, the algorithm is sport agnostic, so could be used in a sport that uses a smaller pitch.