Research on Road Object Detection Algorithm Based on Improved YOLOX

Abstract

Target detection is one of the important branches of autonomous driving, where road targets are detected to achieve awareness of the surrounding environment and an efficient and accurate detection model is required to improve driving safety. The current target detection algorithm can have the problem of missing detection for obscured targets and small targets. In this paper, the YOLOX target detection algorithm is improved to enhance the recall rate of road targets. In this paper, DecIoU loss function is proposed to optimise the bounding box regression process, and Push Loss is used to improve the target occlusion problem. Finally, in order to improve the accuracy of the labels and optimise the model training process, this paper uses a dynamic anchor frame mechanism to adjust the confidence label values. Experiments are conducted with the KITTI autonomous driving dataset, and the experiments show that the method in this paper provides an effective improvement in the detection accuracy of road targets, with the mAP and mAR of YOLOX-s reaching 88.7% and 90.1% respectively, an improvement of 2.54% and 3.21%. The performance on other versions of YOLOX is also consistent, validating the effectiveness of the method in this paper.