Pedestrian Detection for Vehicle-borne Image Based on Two-level YOLOv3

Abstract

In this paper, I mainly focus on real-time pedestrian detection, which is a critical part of robot vision and autonomous driving cars. In recent, convolutional neural networks and deep learning have received so many reputations due to their enormous ability and wide use. For example, image classification, understanding climate, analyzing documents, advertising, etc. Object detection is different from image classification, which is a relatively new area where are waiting for more researchers to dedicate themselves. In the first part, I introduce the appliance of real-time object detection, and in the second part, I introduce some related works of real-time object detection, In the third part, where my work is, I indicate the method to increase the performance of pedestrian detection. I delete the y1 layer of the output of YOLOv3 and magnify the upsampling rate. At the last, I regulate the anchors to achieve more accuracy and better performance. Finally, I explain my experiments and give my research conclusion.