A lightweight road pedestrian target detection method based on the improved YOLOv7

Abstract

Aiming at the issues of large model parameters and high computational complexity of the existing YOLOv7 algorithm in the task of road pedestrian detection, this paper proposes a light-weight and improved YOLOv7 object detection method. Firstly, the backbone feature extraction module is reconstructed by using the GhostNet lightweight network, and the model parameters are reduced through depth-separable convolution and feature reuse strategies. Secondly, the ECA attention mechanism is added into the feature fusion layer in order to enhance the model's ability focusing on fine-grained features of pedestrian targets and improve the detection accuracy of small-scale targets. Meanwhile, the EIoU loss function was designed to optimize the bounding box regression process and alleviate the positioning deviation caused by pedestrian occlusion and scale variation in complex scenes. The experimental outcome shows that the improved model performs well on the public dataset PASCAL The average detection accuracy (mAP) on VOC and KITTI reached 89.7% and 81.2% respectively. Compared to the original YOLOv7, the number of parameters was reduced by 42.3% and the computational load was decreased by 37.6%. Not only the detection accuracy is ensured, but also the real-time performance was significantly improved. This method can be effectively deployed in vehicle-mounted terminals and edge computing devices, providing a lightweight solution for pedestrian detection tasks in intelligent transportation systems.