Improvement in On-road Object Detection Using YOLOv7 with High-order Spatial Interaction and Attention Mechanism.

Object recognition on the road has always been an important research area within deep learning-based object detection. Fast and efficient object detection algorithms can effectively prevent traffic accidents during the autonomous driving process. In this paper, we propose several enhanced versions of YOLOv7 that incorporate attention mechanisms and recursive gated convolutions into the YOLOv7 object detection algorithm to improve its detection accuracy. Specifically, we employ attention mechanisms to enable the network to focus more on important features and use the recursive gated convolution to enhance the network’s ability to extract object features, thus achieving better detection results. We evaluate the effectiveness of the proposed methods on the KITTI and BDD100K datasets, which contain images under varying weather, lighting, and traffic conditions. Experimental results and analysis demonstrate that our proposed methods improve the detection accuracy while maintaining the fast processing speed characteristic of conventional YOLOv7. These results suggest that the proposed methods can improve the detection accuracy of conventional methods under challenging visibility conditions, such as nighttime and/or adverse weather. Therefore, they can be considered as contributing to the enhancement of safety and reliability in autonomous driving technology.