A Deep Learning Based Vehicle Detection Method for Road Scenes

Currently, there are numerous vehicle detection algorithms available. However, in real-world road scenes, factors such as difficult-to-identify features, low resolution, and complex environments often lead to situations of false and missed detections, making it challenging to achieving high accuracy while maintaining high detection speed in practical applications. This paper proposes a multi-scale vehicle detector based on one-stage network YOLOv7, incorporating the Multi-level Feature Pyramid Network (MFPN) into the network to enhance direct interaction between non-adjacent layers. Since the localization accuracy of vehicle detection largely depends on the bounding box loss function, the focal loss combined with the SIoU loss ——Focaler-SIoU is introduced to replace the original bounding box regression loss. The improved network achieved an average precision of 73.9% on the UADETRAC dataset.