A Novel Local Features Based Athlete Detection Method in Sports Video

This paper presents a novel athlete detection method in sports video based on local features. Firstly, the framework of the proposed athlete detection system is illustrated. In the proposed system, noise and shadow are eliminated in advance, and the region growing and connectivity analysis process are implemented as well. After carrying out the global motion estimation process, action clips are detected and the shape of athletes is obtained from an athlete action database. Afterwards, the athletes in the sport video can be detected from background. Secondly, the athlete detection algorithm is described in detail. In this algorithm, the SIFT descriptors are used to describe the local features of video sequences, and all the point pairs from two adjacent video sequence are extracted, and the residual error for each point pair is calculated. Furthermore, the video sequences are represented by the histograms of feature vectors. Then, to obtain the similarity between two histograms, several traditional similarity measures are integrated together using linear fusion. Finally, the final athlete detection results are obtained utilizing the final classification based on adding the scores of all the similarity measures for the histograms. Finally, experiments are designed to make performance, and very positive experimental results are obtained.