Research on Enhanced Multi-Task Traffic Scene Object Detection and Classification Method Based on YOLOv5
Jie Wu, Zhian Zhang · 2023
A YOLOv5 based enhanced multitask traffic scene object detection and classification method, is proposed to address the issues of weak ability for multi-scale small object detection in complex traffic scenes. By improving the feature fusion network, the model can fully utilize feature information and increase the accuracy of target localization and boundary regression; Design multiple detection head structures and perform structural lightweighting, can preserve the original feature information and fuse multi-scale feature information. Enhancing the global perception ability of the model, to achieve simultaneous learning and detection classification of multi-scale and different category attribute samples. The experimental results show that, the improved method achieved the highest mAP of 97.2 % and inference speed of 105.62 frames/s on the Chinese urban traffic signal dataset, which were 2.1 % and 40.9 frames/s higher than the original YOLOv5. The model has significantly improved the detection ability and robustness in various complex environments, effectively improving the real-time detection accuracy of multi-scale targets in general complex scenarios.