Multi-Camera Multi-Vehicle Tracking with Domain Generalization and Contextual Constraints

In this paper, we propose a system for Multi-Camera Multi-Target (MCMT) Vehicle Tracking in Track 1 of AI City Challenge 2022. There are many technical difficulties to the MCMT problem such as a common lack of labeled data in real scenarios, a distortion of vehicle detailed appearances in recording, and ambiguity between highly similar vehicles. Taking those into account, we develop a 3-component MCMT system that exploits vehicle behavior, leverages synthetic data and multiple augmentation techniques, and enforces contextual constraints. Specifically, our system involves a motion-driven vehicle tracker for obtaining robust trajectories, applying MixStyle domain generalization on the TransReID model to exploit as much labeled data as possible, and experimenting with contextual constraints such as our proposed neighbor matching to address ambiguity in terms of vehicle appearances. Overall, our system achieved an IDF1 score of 0.7255.