Enhancing the Robustness of Vehicle Re-Identification in Intelligent Transportation Systems

Vehicle Re-identification (ReID) is a critical component of Intelligent Transportation Systems (ITS), enabling effective traffic management and law enforcement. However, achieving robust and efficient ReID remains challenging due to the significant variability in vehicle appearances and environmental conditions. This thesis addresses these challenges and makes several key contributions to enhance robustness, accuracy, and efficiency in vehicle ReID.

We developed a comprehensive data set generation pipeline that uses vehicle detection algorithms with confidence scores to select optimal Regions of Interest (ROI) for image cropping. This ensures high-quality datasets tailored for ReID tasks. Comprehensive experiments demonstrated that our designed ROIs can significantly improve vehicle ReID accuracy and tracking performance. Furthermore, we created a unique dataset that combines images from highway cameras and drone footage captured across nine pairs of weaving areas, enriching the diversity of the data and enabling robust model training.

To enhance robustness in real-world scenarios, we conducted an aspect ratio analysis and demonstrated that Vision Transformer (ViT)-based ReID models can achieve improved robustness when trained with varying aspect ratios. This approach effectively addresses challenges arising from diverse vehicle shapes and imaging conditions. We incorporated a mixup method applied at the patch level during the ViT’s conversion of images into patches—using spatial attention weights as guidance—and adopted non-uniform strides to more accurately match the inherent aspect ratios of objects. Finally, we proposed a dynamic feature fusion ReID network to further improve model robustness. Our method achieved a 10\% improvement in mean average precision (mAP) on the VehicleID dataset, surpassing existing state-of-the-art methods.

Furthermore, to address the domain generalization challenge in ReID - where models trained on urban road data struggle to generalize to highway data - we proposed a novel framework to bridge this gap, improving cross-domain performance in vehicle ReID. This work builds upon an existing framework that combines image-level and feature-level disentanglement for improved generalization.We leveraged Conditional Value at Risk (CVaR) to boost robustness and employed a loss landscape smoothing technique—commonly known as sharpness-aware minimization—to foster improved generalization, our approach achieved a 3\% improvement in mAP compared to baseline, setting a new benchmark in cross-domain ReID.

Together, these innovations address key challenges in vehicle ReID, providing a pathway to improved accuracy, robustness, and efficiency, and positioning this work as a valuable contribution to real-world ITS applications.