Environmental Perception Algorithm for Multi-Target Autonomous Driving Based on YOLOv7-CBAM

Abstract

In this study, we proposed an improved YOLOv7 network for autonomous driving multi-target detection, YOLOv7-CBAM. Based on YOLOv7, a tiny target detection structure was added to the backbone network for detecting tiny-scale targets. The ResNet-CBAM structure was utilized to reconstruct the YOLOv7 backbone network for enhancing the image feature extraction capability of network model, thereby obtaining more nonlinear features information from different feature layers. Besides, the Kernel-means algorithm was used for obtain the anchor instead of the K-means. Experiments showed that the proposed YOLOv7-CBAM network demonstrates faster convergence and reaches a mean average precision value of 82.0% on the Udacity Self Driving Car dataset, superior to the original YOLOv7 model.