An MLP-based player detection and tracking in broadcast soccer video

This paper addresses the automatic player detection and tracking problem applied to broadcast images of soccer games. We propose an approach for player detection in long view of broadcast soccer video. We detect long view frames using dominant color ratio. We use k-means clustering algorithm for calculating dominant color thresholds. Our method for player detection has two phases: in the first phase we perform a preprocessing for deleting some non-player regions, this is done with knowledge of soccer video domain and morphology operation. Remaining regions will be classified into player or non-player groups with using a Multilayer Perceptron Neural Network as classifier in second phase. Our used feature vector for every region consist of two parts: color value of region in Cb and Cr layers from YCbCr color space and area of the region. Finally we track each player. Experiments show that our method can detect player effectively and with high accuracy.