Research on Negative Obstacle Detection Method Based on Image Enhancement and Improved Anchor Box YOLO

The ability of environmental awareness is the premise for unmanned ground vehicles to navigate autonomously and avoid obstacles. Obstacle detection is an important part of environmental awareness technology. Compared with the positive obstacles above the ground, the negative obstacles located below the ground are more difficult to detect and cause more damage to the unmanned vehicle. Consequently, research on various types of negative obstacle detection technology has great significance and value. Aiming at the problem that the texture features of negative obstacles are not obvious, the features are easily blocked by shadows, and the negative obstacles are relatively small in the camera’s field of vision. This paper presents a negative obstacle detection method based on image enhancement and the improved anchor box YOLOv3. Firstly, AutoMSRCR is used to enhance the texture features of negative obstacles in the image to obtain the image enhancement data set. Then, k-means++ clustering and threshold scaling are performed on the data set to get the anchor boxes of the training network. The enhanced images are brought into the improved anchor box YOLOv3 as a training set. Finally, three sets of experiments are designed to verify the detection effect of the negative obstacle detection method, and the experimental results show that the network can quickly detect negative obstacles with high accuracy and a low missed detection rate.