Non-local attention association scheme for online multi-object tracking

Abstract

Online multi-object tracking (MOT) is a fundamental problem in video analysis and multimedia applications. The major challenge in the popular tracking-by-detection framework is knowing how to associate candidate detections results with existing tracklets. In this, we propose a non-local attention association approach and apply it to a unified online MOT framework that integrates the merits of single object tracking and data association methods. Specifically, we use non-local attention association networks (NAAN) to incorporate both spatial and temporal characteristics to associate new detections. The non-local attention mechanism generates global attention maps across space and time, enabling the network to focus on the whole tracklet information, as opposed to the local attention mechanism to overcome the problems of noisy detections, occlusion, and frequent interactions between targets. Experimental results on MOT benchmark datasets show that the proposed algorithm performs favorably against various online trackers on the basis of identity-preserving metrics.