Occluded pedestrian detection based on improved YOLOv5n

Abstract

Aiming at the problem of pedestrian targets occlusion and multi-scale error and missed detection in pedestrian detection, a lightweight pedestrian detection algorithm based on improved EA-YOLOv5n is proposed. This method introduces the ECA attention module into the backbone feature extraction network, and learns the channels of pedestrian images by learning Information, improve the accuracy of pedestrian object detection in the case of occlusion, improve the calculation method of Bounding box loss function for the disadvantages of loss function calculation, adopt EIoU Loss and introduce power transformation to obtain higher bounding box regression accuracy. The experimental results show that using the improved model to conduct experiments on the Widerperson dataset reaches 69.6% mAP, which is 2.0% higher than the original algorithm, and the detection speed reaches 65FPS.