A traffic flow statistical detection method based on computer vision technology

Abstract

This article addresses the issues of high deployment costs, low accuracy in small object detection, and easy ID switching in multi-target tracking of existing traffic flow statistics methods. Based on the visual sensing mechanism of visible light imaging, this article designs and implements a traffic flow statistics system based on YOLOv8 object detection algorithm and ByteTrack multi-target tracking algorithm. The system first constructed a multi scene traffic dataset containing approximately 9000 images; Then, the YOLOv8n model is used for training, and the model's generalization ability is improved through hyperparameter tuning and data augmentation strategies; Subsequently, the TensorRT inference framework was used to perform FP16 semi precision acceleration on the model, increasing the inference frame rate from 54.62 FPS to 63.37 FPS, with an increase of 16.01%; Finally, the ByteTrack algorithm is integrated to achieve multi-target tracking, and a virtual line based cross line counting method is designed to achieve bidirectional traffic flow statistics. The experimental results show that the counting accuracy of the system reaches 96%, 97%, and 94% respectively in three typical scenarios: daytime intersections, nighttime highways, and daytime congestion. It has good practical value and prospects for promotion and application, and also provides a technical basis for the subsequent introduction of infrared thermal imaging multimodal fusion to enhance low illumination robustness.