An object detection algorithm combining self-attention and YOLOv4 in traffic scene

Abstract

Automobile intelligence is the trend for modern automobiles, of which environment perception is the key technology of intelligent automobile research. For autonomous vehicles, the detection of object information, such as vehicles and pedestrians in traffic scenes is crucial to improving driving safety. However, in the actual traffic scene, there are many special conditions such as object occlusion, small objects, and bad weather, which will affect the accuracy of object detection. In this research, the SwinT-YOLOv4 algorithm is proposed for detecting objects in traffic scenes, which is based on the YOLOv4 algorithm. Compared with a Convolutional neural network (CNN), the vision transformer is more powerful at extracting vision features of objects in the image. The CNN-based backbone in YOLOv4 is replaced by the Swin Transformer in the proposed algorithm. The feature-fusing neck and predicting head of YOLOv4 is remained. The proposed model was trained and evaluated in the COCO dataset. Experiments show that our method can significantly improve the accuracy of object detection under special conditions. Equipped with our method, the object detection precision for cars and person is improved by 1.75%, and the detection precision for car and person reach 89.04% and 94.16%, respectively.