Edge-Optimized Modular Architecture for Real-Time Vehicle Re-Identification

This work addresses the challenge of real-time vehicle re-identification (ReID) on resource-constrained edge devices. Our research aims to design and deploy a modular, edge-optimized system that integrates object detection and ReID for intelligent transportation applications. The proposed architecture combines a lightweight detection module with a ReID embedding generator, enabling accurate multi-camera vehicle tracking under strict computational limits. We evaluate the system on NVIDIA Jetson platforms, applying quantization, pruning, and TensorRT optimization to achieve low-latency inference while maintaining accuracy. Additionally, we explore scheduling strategies for day and night operations to optimize energy consumption and performance. The impact of 360-degree fisheye camera distortion on detection and ReID accuracy is analyzed. Experimental results demonstrate that the system achieves real-time performance on edge hardware without sacrificing reliability, making it suitable for large-scale deployment in smart city environments. This study highlights the feasibility of deploying deep learning computer vision pipelines on edge AI platforms for robust, privacy-preserving vehicle analytics.