ORO-YOLO: An Improved YOLO Algorithm for On-Road Object Detection

Environment perception and object detection are the foundation for unmanned ground vehicles (UGV) to autonomously complete various tasks, and also one of the important research topics in the field of unmanned systems. On-Road object detectors based on deep learning have made significant progress, but it is still difficult to balance the detection accuracy and speed. In this paper, an improved fast on-road object detection algorithm ORO-YOLO based on YOLOX network is proposed to solve this problem. First, we improve the backbone network by the reparameterization method to enhance the network feature extraction capability. Second, a feature enhancement module is presented for the actual road target aspect ratio, which enables the network to be more suitable for on-road object detection tasks. Moreover, an adaptive feature fusion module is introduced to replace the original feature pyramid module to further compress the network size. Comparative experiments on KITTI and BDD100k datasets demonstrate the effectiveness of the proposed algorithm, with up to 2% higher accuracy than the original network and about 10% reduction in the amount of network parameters.