Motion vehicle detection for traffic video streams
Jianhao Song, Hua Ding, Lai Liu, Juntao Chen · The Journal of Engineering · 2025
Abstract This research offers an enhanced version of the YOLOv5s+MobileNetV3+BiFPN method to address a number of issues in moving vehicle recognition in traffic video streams, including lightweight network model, motion blur, scene delay etc. By substituting YOLOv5s BackBone with the lightweight MobileNetV3 backbone network and employing a parameter‐free attention mechanism, the network model's parameters are reduced. To improve the ability of features to learn, use SimAM instead of SENet structure; Integrating the BiFPN network will increase the computing cost while improving the detection accuracy; Utilize the K‐means++ method to improve the anchor frame choice and boost the detection performance for targets of various scales; In addition, the dataset is arbitrarily disregarded to address the issue of extreme information redundancy. Results are obtained by applying the enhanced algorithm to the processed video detection dataset. The experimental findings demonstrate that this method, when compared to the original YOLOv5s algorithm, has a faster detection speed and a higher detection accuracy in the detection of vehicle targets in traffic video streams. The [email protected]:0.95 improved by 11.23%, and the FPS increased from 15.87 to 47.62.