Year: 2026
Type: article
Abstract: Detecting and tracking vehicles ahead during driving is a critical technology for ensuring driving safety and enhancing the driving experience. However, traditional detection and tracking models often suffer from high computational intensity and large parameter sizes, making them unsuitable for resource-constrained mobile devices. This paper introduces a lightweight vehicle tracking model that integrates YOLO and DeepSORT. By incorporating GhostConv and C3GhostV2 modules into the backbone and neck networks of YOLO, we enhance detection accuracy while reducing model size. The Focal-EIoU function is utilized to improve the recognition of small, distant target vehicles. Additionally, ShuffleNetV2 replaces DeepSORT’s feature extraction network, further reducing model parameters. Experimental results demonstrate that among the latest models in the YOLO series, YOLOv5 offers more stable and accurate performance, making it a more suitable baseline for vehicle detection and tracking tasks. Based on this foundation, experimental results show that, compared to the baseline model implemented using the Ultralytics framework, the proposed model reduces GFLOPs by 37.1%, decreases the number of parameters by 34.5%, and achieves [email protected] by 0.3%. The overall model size is only 13.62 MB, 44.8% smaller than the original. Validated on the SODA10M dataset, the model exhibits improved detection accuracy and maintains performance compared to other models, making it suitable for deployment on mobile devices during vehicle driving.
Source: Scientific Reports
Authors: Wei Xu, Jiacheng Liu, Ming Cheng, Yucheng Huang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7223339726