Implementation of Pedestrian Detection Algorithm Based on Improved Yolov3-Tiny in ROS Framework

Abstract

In order to improve the low detection accuracy and recall rate of Yolov3-Tiny on small or obscured pedestrian target, this paper proposes an improved Yolov3-Tiny pedestrian detection algorithm. On the basis of data set (head and shoulders) making, we add 52×52 feature maps into the network to improve the multi-scale fusion structure, use a data augmentation method to enhance the generalization ability of the model, and adjust the learning-rate dynamically. Based on this algorithm, we implement real-time pedestrian detection experiments for mobile robots in ROS framework, and the results show that the detection accuracy, recall rate and IOU have been improved.