Scene-Specific Multiple Cues Integration for Multiperson Tracking

Robust multiperson tracking requires the correct associations of online detection responses with existing trajectories. In this paper, we propose to integrate multiple cues to resolve the ambiguities in data association for multiperson tracking. Unlike most existing algorithms which integrate multiple cues in the same manner for different scenes, we learn scene-specific parameters to integrate multiple cues for different scenes, as the discriminative power of each cue may vary in different scenes. The scene-specific integration parameters are learned offline by supervised learning method. Min-cost multicommodity flow is employed to solve the data association task. The edge cost of the multicommodity network, which is crucial for the data association, is determined by integrating the multiple cues extracted from the detection response based on the learned scene-specific integration parameters. The experimental results on public multiperson tracking data set demonstrate the effectiveness of the proposed scene-specific integration method.