Enhanced vehicle detection via YOLOv7-tiny: Advancements in feature extraction and bounding box regression

Abstract

With the proliferation of intelligent transportation systems, vehicle detection plays a pivotal role in autonomous driving and traffic management. However, real-time vehicle detection faces challenges, particularly in dense traffic scenarios with overlapping and occluded vehicles, small targets at long distances, and variable shapes of large vehicles. To address these issues, this paper presents improvements to the YOLOv7-tiny algorithm. Firstly, we introduce the Sim_DFC feature extraction module, which combines multi-scale feature extraction with an attention mechanism to enhance recognition accuracy for large and special vehicles. Secondly, we optimize object box regression by integrating Shape IoU with Inner IoU, improving the detection of overlapping vehicles and small targets. Finally, we adopt an improved BiFPN structure to enhance multi-scale feature fusion without increasing computational complexity. Experimental results on the UA-DETRAC and Visdrone2019 datasets demonstrate that our proposed method significantly improves mAP50 accuracy by 9.8% and 1.8%, respectively. Meanwhile, the model maintains real-time detection performance with competitive FPS and achieves higher precision and recall compared to the baseline model. The detection effect graphs further validate the reliability and feasibility of our approach across various traffic scenarios.