Multi-object Tracking by Joint Detection and Identification Learning

Abstract

Multi-object tracking (MOT) is closely related to video-based object detection and target re-identification. In recent years, with the representation power brought by deep learning, the majority of state-of-the-art methods on object detection and re-identification are based on deep neural networks. However, it is still an open problem to improve the performance of MOT in real challenging scenes. Specifically, recent MOT algorithms have not been optimized together with object detection, which hinders the performance of tracking. Inspired by recent progress on object detection and recognition, we propose a MOT method via joint learning on detection and identification by using existing MOT datasets without external training data. We further introduce a feature enhancement module based on the ConvGRU structure, which helps to deal with deterioration of image quality in video object detection and re-identification, such as motion blur and camera losing focus. Experimental results show that the proposed method achieves competitive performance compared with state-of-the-art methods in video-based object detection, cross-dataset person re-identification, and multi-object tracking.