Real-Time Obstacle Detection Method in the Driving Process of Driverless Rail Locomotives Based on DeblurGANv2 and Improved YOLOv4

In order to improve the detection accuracy of an algorithm in complex environment a coal mine, including low-illumination, motion-blur, occlusions, small-targets, and background-interference conditions; reduce number model parameters; speed algorithm; make it meet real-time requirements edge equipment, obstacle method driving driverless rail locomotives based on DeblurGANv2 improved YOLOv4 is proposed this study. A blurred image was deblurred using DeblurGANv2. The design YOLOv4, lightweight feature extraction network MobileNetv2 used replace original CSPDarknet53 algorithm. There high amount background interference target mine scene. strengthen attention paid target, SANet module embedded Neck under target-occlusion, small-target, other conditions. To further algorithm, K-means++ adopted cluster prior frames, focal loss function introduced increase weight small-target samples. experimental results show that deblurring motion-blurred can effectively obstacles missed detections. Compared with increases by 65.85% 68 FPS 0.68% 98.02%.