Based on improved YOLOv8 and Bot SORT surveillance video traffic statistics

Abstract

Aiming at the problems of leakage detection and low detection accuracy of existing deep learning based surveillance video traffic flow detection algorithms, a traffic flow counting system combining improved YOLOv8 detection and Bot SORT tracking is proposed. First, the backbone network is used to incorporate the SPD-Conv convolutional layer to improve the network's ability to detect small targets. Then, the attention mechanism CoTAttention is introduced into the neck network to further improve the model generalization ability. Finally, the improved YOLOv8 model and the Bot SORT algorithm are combined to design and implement a traffic counting system capable of monitoring video traffic in real time, and trained and tested on the open-source UA-DETRAC vehicle detection dataset. The experimental results show that the improved YOLOv8 algorithm improves F1, P, mAP50, and mAP50-95 by 0.36, 2.2, 1.8, and 2.1 percentage points, respectively, compared with the original algorithm. Combined with the Bot SORT tracking, it achieves more accurate and reliable results in the task of traffic counting, which provides a strong support for the vehicle detection and counting in the monitoring system.