Tracking and recognizing actions of multiple hockey players using the boosted particle filter

This article presents a system that can automatically track multiple hockey players and simultaneously recognize their actions given a single broadcast video sequence, where detection is complicated by a panning, tilting, and zooming camera. There are three contributions. Firstly, we use the Histograms of Oriented Gradients (HOG) to represent the players, and introduce a probabilistic framework to model the appearance of the players by a mixture of local subspaces. We also employ an efficient off-line learning algorithm to learn the templates from training data, and an efficient online filtering algorithm to update the templates used by the tracker. Secondly, we augment the boosted particle filter (BPF) with a new observation model and a template updater that improves the robustness of the tracking system. Finally, we recognize the players' actions by combining the HOG descriptors with a pure multi-class sparse classifier with a robust motion similarity measure. Experiments on long sequences show promising quantitative and qualitative results.