A Real-Time Vehicle Detection and Re-ID System from the Drone Perspective

To address the application demand for unmanned aerial vehicles (UAVs) to identify important ground vehicles, a software system for detecting and recognizing significant vehicles from UAV images has been developed. The system employs the YOLOv11 model, utilizing domain adaptive training with vehicle data captured by UAVs to adapt to the UAV perspective for effective vehicle detection. Additionally, a lightweight vehicle re-identification model with channel grouping is proposed, maintaining high re-identification accuracy while significantly reducing the number of parameters and computational load. The system, running on a workstation with an RTX 4070 graphics card, processes single-frame images in less than 10 ms. It also supports functions like setting targets autonomously and adjusting recognition matching thresholds, making it suitable for rapidly and accurately searching and identifying vehicles such as hazardous chemical transporters and hit-and-run vehicles, demonstrating high practical value.