IAMPDNet: Instance-aware and Multi-part Decoupled Network for Joint Detection and Embedding

Abstract

Video surveillance applications play an important role in smart cities. Recently, intelligent video surveillance methods have been widely investigated to address large-scale video data, among which multi-object tracking (MOT) is the most popular method, which aims to track every object appearing in the video for monitoring. MOT is essential for deep intelligent perception. Considering inference speed, joint detection and embedding (JDE) has become a new paradigm for MOT. JDE is to obtain detection results and object features through one forward propagation. However, most JDE models lack instance awareness ability and multi-part feature extraction ability, which may lead to the lack of discrimination of extracted instance features. To address these problems, in this paper we propose an instance-aware and multi-part decoupled network (IAMPDNet), which can perceive all instances in the environment and extract multi-part features from the instances. Specifically, our IAMPDNet consists of three key modules: a complementary attention module used to perceive all instances in the environment, a feature extraction module used to decouple multi-part features from the instances, and an adaptive aggregation module used to fuse multi-level features of instances. Extensive experiments on multiple MOT benchmarks demonstrate that our IAMPDNet achieves higher tracking accuracy and lower identity switches against recent MOT methods.