Object detection algorithm based on improved Yolov3-tiny network in traffic scenes

Abstract

The object detection based on deep learning is an important application in the field of vehicle environment perception, which has been a hot topic in recent years. We propose a novel improved Yolov3-tiny to implement more accurate object detection for the objects in traffic scenes. We employ K-means algorithm to cluster the common objects in traffic scenes to obtain a suitable size and numbers of anchor box. In addition, we modify modifying detection scale and the backbone network structure of Yolov3-tiny, improving the detection accuracy for small object. The stereo vision is also introduced to improve the accuracy of boundary location. Experiments results demonstrate that the improved yolo-tiny has higher accuracy than the original algorithm and it also meet the requirement of real-time performance.