Multi-camera multi-player tracking with deep player identification in sports video. (June 2020)

• Record Type: Journal Article
• Authors: Zhang, Ruiheng; Wu, Lingxiang; Yang, Yukun; Wu, Wanneng; Chen, Yueqiang; Xu, Min
• Is Part Of: Pattern recognition. Volume 102(2020:Jun.)
• Publication Date: 2020-06
• Subjects: Identity switch -- Multi-target multi-camera tracking -- Object detection -- Player identification -- CNN; 00-01 -- 99-00; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107260
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a robust tracking framework for basketball players in the multi-camera sports videos. We introduce deep-learning-based player identification into the player tracker to leverage the players ID. The framework consists of three modules: DeepPlayer, IPOM and KSP-ID. A Cascade Mask RCNN and a PoseID model are designed for the DeepPlayer. Our framework yields superior performance in two benchmarks. Abstract: Identity switches caused by inter-object interactions remain a critical problem for multi-player tracking in real-world sports video analysis. Existing approaches utilizing the appearance model is difficult to associate detections and preserve identities due to the similar appearance of players in the same team. Instead of the appearance model, we propose a distinguishable deep representation for player identity in this paper. A robust multi-player tracker incorporating with deep player identification is further developed to produce identity-coherent trajectories. The framework consists of three parts: (1) the core component, a Deep Player Identification (DeepPlayer) model that provides an adequate discriminative feature through the coarse-to-fine jersey number recognition and the pose-guided partial feature embedding; (2) an Individual Probability Occupancy Map (IPOM) model for players 3D localization with ID; and (3) a K-Shortest Path with ID (KSP-ID) model that links nodes in the flow graph by a proposed player ID correlation coefficient. With the distinguishable identity, the performance of tracking is improved. Experiment results illustrate that our framework handles the identity switches effectively, and outperforms state-of-the-art trackers on the sports video benchmarks.