The Research on Improving YOLOv10 Algorithm for Object Detection in Autonomous Driving Scenarios
Year: 2025
Type: conference-paper
Abstract: This study investigates the challenging problem of target detection in autonomous driving scenarios, illustrating the new baseline modular improvements of the YOLOv10 algorithm. Our collaborative team made significant enhancements to the model's feature extraction ability and general performance by using a Swin Transformer, RefConv, and HWD module. RefConv performs particularly well at detecting small objects, the Swin Transformer showed great robustness when discriminating difficult-to-distinguish objects in complex backgrounds, and the HWD module was capable of compressing resolution without loss of key information. Experimental results reveal that the modified model achieves an mAP of 92.32%, a precision of 91.73%, and recall of 90.14% representing improvements of 92.32%,3.02% and 3.86% from the baseline model. The ablation experiments confirmed the performance of each module, properties which were fully substantiated and contextualized for use in autonomous driving perception tasks.
Authors: Hanjin Wen, Junhua Xiao, Keming Guan, Chenrui Wang
Institution: Beijing Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4415074070