PEDESTRIAN DETECTION IN SELF-DRIVING CARS IN COMPLEX LIGHTING SCENES

For pedestrian detection in low light or scenes with drastic changes in light conditions, the problems of missed detection and inaccurate positioning are likely to occur. A pedestrian detection method based on the joint learning of image enhancement (EnlightenGAN) and YOLOv5 is proposed. The feature extraction network in this method could learn the structural details and color features of the image reconstructed by the enhancement module of the adversarial neural network. The relative entropy loss function after smoothing was used as the confidence loss of YOLOv5 to improve the generalization performance of the network. The experimental results show that the method based on image enhancement and YOLOv5 joint learning can effectively reduce the sensitivity of different illumination to pedestrian detection on the self-made complex illumination scene pedestrian data set, and the precision reaches 87.58%. After improving the YOLOv5 loss on the Caltech data set, the convergence speeds up and the detection precision is increased by 1.73 percentage points.