A lightweight tunnel vehicle re-ldentification model based on YOLOv11n and FaceNet.

Abstract

With the advancement of computer vision, vehicle re-identification (Re-ID) in tunnel environments faces critical challenges like low-resolution imagery, lighting variations, and occlusions, which greatly limit the effectiveness of existing algorithms. This study presents a novel framework for intelligent tunnel vehicle monitoring, integrating lightweight detection and enhanced feature learning. Specifically, YOLOv11n is embedded as the front-end for lightweight detection; for vehicle Re-ID, the FaceNet model is optimized by replacing its Inception-ResNet backbone with MobileNetV3 and adding a Coordinate Attention module, along with a proposed joint loss function combining IoU-based hard triplet mining and Center Loss. A tunnel-specific dataset with 12,000 vehicle images is constructed, incorporating data augmentation to handle real-world surveillance complexities. Experimental results show: YOLOv11n achieves 98.63% mAP at 242 fps; the improved Re-ID model reaches 94.18% accuracy at 25.43 fps (0.81 GFLOPs, 3.51M params), outperforming baselines; ablation studies validate components, and AUC improves by 2.44%. This work provides a robust solution for real-time tunnel vehicle monitoring, with potential extensions to multi-modal fusion and cross-tunnel transfer learning.