Improved YOLOv4-tiny network for pedestrian detection

Abstract

In the intelligent driving system, it is very important to identify pedestrians accurately and efficiently. However, when pedestrians are in a long distance, they are small in the field of vision and difficult to detect. This paper presents a pedestrian detection method based on YOLOv4-tiny network. According to the characteristics of pedestrians and the multi-level detection principle, we improved the anchor box and structure of YOLOv4-tiny network. The improved model was tested by using the collected multi-segment driving image data and the result shows that the performance of the model for pedestrian detection is significantly improved, especially for small pedestrians. In three of the test scenarios, the accuracy of pedestrian detection is improved from 54.5%, 68.2% and 67.8% to 86.7%, 90.4% and 91.3%, respectively. In addition, this method can also be used to detect other types of targets (such as vehicles) and has a certain versatility.