ETD

• Computer Vision
• FastReID
• Kinematic-Guided ReID
• Multi-Class Tracking
• Multi-Object Tracking
• Re-Identification
• Wide Residual Network

Multiple Object Tracking is a fundamental problem in autonomous-vehicle perception, yet many state-of-the-art trackers struggle in multi-class automotive environments. This weakness is not fully exposed by conventional single-class benchmark protocols. Most existing tracking pipelines were designed for single-class settings and therefore apply a single association strategy to all object classes, despite large differences in the motion and visual characteristics of agents such as vehicles and pedestrians. In dense multi-class scenes, this limitation is compounded by the cubic computational cost of global association, which becomes a major obstacle to real-time deployment.

This thesis proposes the Hybrid Multi-Class (HMC) Tracker, a parallel class-oriented multi-paradigm framework designed to address these limitations. Given a scene with $N$ objects, the HMC Tracker performs a class-oriented parallel association to reduces the computational complexity of association from $O(N^3)$ to $O(n_{\max}^3)$ where $n_{max}$ is the highest number of objects of the same class, among all detectable classes. Building on this decomposition, the framework combines motion-based matching, residual spatial association, and a final kinematic-guided ReID stage, together with controlled feature admission and class-specific noise calibration, to balance computationaal efficiency with robust association under occlusion.

Experimental results on the MOT17, KITTI, and Waymo V2 benchmarks show that the HMC Tracker improves both accuracy and efficiency. On Waymo V2, it reduces identity switches by 42.6\% relative to the OC-SORT baseline while maintaining a high frame rate of 25.5 FPS. Across the three benchmarks, it also achieves an average frame rate roughly 4.5$\times$ higher than comparable high-accuracy ReID-based frameworks. These results demonstrate that reducing association complexity through class-oriented parallelization, together with class-specific motion modeling, selective ReID usage, and class-dependent uncertainty calibration, enables robust and scalable multi-class tracking in autonomous-driving scenarios.