Nighttime Vehicle Detection Algorithm Based on Improved YOLOv7

Aiming at the problems of low visibility, fuzzy target features and high leakage rate of small targets in nighttime vehicle detection, this paper proposes a nighttime vehicle detection algorithm E-YOLOv7 based on the improved YOLOv7. Firstly, a hybrid feature enhancement module (SFE) is designed to enhance the expression of key features through the grouped-channel attention mechanism and feature reorganization to alleviate the loss of semantic information caused by fuzzy features of nighttime vehicles. Second, a lightweight efficient feature fusion network (GS-EFF) is constructed to optimize multi-scale feature alignment by combining the jump connection and GSConv modules to reduce the information loss in the feature fusion process and lower the number of parameters. The Soft-EloU-NMS post-processing algorithm is further proposed to effectively reduce the leakage detection rate of dense small targets by fusing the multi-dimensional evaluation of overlap, center distance and aspect ratio. Experiments on the BDD100K nighttime vehicle dataset show that E-YOLOv7 achieves an AP value of 94.9%, which is a 6.6% improvement over the original YOLOv7, with a significant increase in recall. The model also demonstrates strong real-time performance. Ablation experiments verify the synergistic optimization effect and efficiency of each module. Furthermore, a comparison with other state-of-the-art algorithms like SSD and DETR confirms the superiority of our approach. Visualization results validate the robustness of the algorithm in long-distance, multi-target, and complex lighting scenarios. This study provides a high-precision real-time detection solution for nighttime environment sensing for autonomous driving.