A Receptive Fields Enhancement YOLO V4 on Pedestrian Detection

Abstract

Since different illumination condition, viewpoint change, local occlusion, and scale effect make the appearance features of pedestrian change, field enhancement is essential for pedestrian detection using the deep learning method. Deep learning-based approach to further enhance pedestrian detection into a semantic segmentation problem, but distant pedestrians often disappear in the feature extraction stage because the target is too small. Also, distant pedestrian detection becomes increasingly tricky because similar images or marks can vary in shape and pose. A pedestrian detection model based on YOLO V4 is proposed to address the above problems by studying the similarity between people. The key idea of the model is to further improve the accuracy of YOLO V4 by marking full use of the relationship and distribution law of the appearance features between people. Specifically, more extensive kernel size convolution operation is used to enlarge the perceptual field and make the feature distinct. With a quantization method that can balance accuracy and speed, a module can enhance the convolution neural network's remote dependency modeling capability. Then, the accuracy of pedestrian detection can be improved combining proposed module with YOLO V4. Finally, comparative experiments on the COCO dataset are verified. The results show that the accuracy precision of the proposed model is up to 77.25% and the detection speed is up to 34.33 fps.