Fully-Convolutional Siamese Networks for Football Player Tracking
Yuejie Ma, Shuang Feng, Yongbin Wang · 2018
Automatically tracking footballer over a large-scale video dataset has a wide application in the intelligent explanation and analysis of football video. However, it is non-trivial due to two reasons: occlusion in targets are getting close and even if they collide with each others, and large variations in their silhouettes. Towards this end, we propose a scheme comprising of two components. In particular, we build a Fully-Convolutional Siamese Network (FCSN for short) trained end-to-end on the ILSVRC15 dataset to extract a rich set of visual features and obtain a generic target representation. We then, in the second component, fine-tune FCSN on a large number of videos and images which contain many similar objects to increase the ability of the network to deal with occlusion. FCSN is capable of extracting rich visual features and distinguishing similar objects, which effectively solves the large variations and occlusion problem. Extensive experiments on a large-scale football dataset have well-validated our model.