Multi-Target Multi-Camera Tracking of Vehicles by Graph Auto-Encoder and Self-Supervised Camera Link Model

Hung-Min Hsu, Yizhou Wang, Jiarui Cai, Jenq-Neng Hwang; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2022, pp. 489-499

Multi-Target Multi-Camera Tracking (MTMCT) of vehicles is a challenging task in smart city related applications. The main challenge of MTMCT is how to accurately match the single-camera trajectories generated from different cameras and establish a complete global cross-camera trajectory for each target, i.e., the multi-camera trajectory matching problem. In this paper, we propose a novel framework to solve this problem using the self-supervised trajectory-based camera link model (CLM) with both appearance and topological features systematically extracted from a graph auto-encoder (GAE) network. Unlike most related works that represent the spatio-temporal relationships of multiple cameras with the laborious human-annotated CLM, we introduce a self-supervised CLM (SCLM) generation method that extracts the crucial multi-camera relationships among the vehicle trajectories passing through different cameras robustly and automatically. Moreover, we apply a GAE to encode topological information and appearance features to generate the topological embeddings. According to our experimental results, the proposed method achieves a new state-of-the-art on both CityFlow 2019 and CityFlow 2020 benchmarks with IDF1 of 77.21% and 55.56%, respectively.