Multiple-Person Tracking with Criss-Cross Attention and Deep Matching

Multi-pedestrian tracking is one of the important application fields of computer vision, among which the real-time tracking of pedestrians and the ability to solve occlusion problems are two big challenges to improve the recognition accuracy. In this paper, we propose a new standardized tracking-by-detection framework that can effectively alleviate the above two problems. We first introduce the Criss-Cross Attention (CCA) mechanism in the detector. The CCA module can reduce the computational complexity of the detector and improve the detection accuracy. Then a pedestrian deep feature extractor fused with cosine metric learning for two-layer convolution is added to further improve the tracking accuracy and real-time performance. Finally, Kalman Filter prediction combined with pedestrian deep features effectively reduces identity switching under the occlusion. We conduct experiments on the MOT16 datasets and evaluate on HOTA. Experimental results demonstrate the competitiveness of our method against some selected state-of-the-art methods.