Dynamic global tracker for online multi camera multi vehicle tracking

Abstract

Abstract Multi-camera multi-target Tracking (MCMT) is often regarded as a downstream task of Multi-Object Tracking (MOT). Traditional methods typically follow an offline pipeline involving detection, re-identification, single-camera tracking, and post-hoc clustering, which leads to poor real-time performance, high computational cost, and weak adaptability in dynamic environments. Moreover, trackers tailored for specific locations overly rely on manually crafted information like road topology and camera calibration, reducing their effectiveness in varied scenarios. We propose Dynamic Global Tracking (DGT), an innovative online framework for Multi-Camera Multi-Target (MCMT) vehicle tracking. Unlike traditional methods that rely on full trajectory extraction and then clustering, the DGT integrates cross-camera associations directly into the tracking process. This transformation reduces the computational burden and enhances real-time performance. Especially, our framework includes a Hybrid Fusion Module (HFM) to address resolution disparities and a Stable Trajectory Manager (STM) to improve stability and robustness. Extensive experiments demonstrate that DGT significantly improves tracking accuracy and adaptability in various environments, achieving an IDF1 score of 61.19 on the HST dataset (speed version) and 70.49 (performance version) with FPS of 90.