作者
Jianping Gao,Yingyu Li,Haotian Li,Jiaxu Yang,Zhe Li
摘要
Abstract With the development of autonomous driving vehicle-infrastructure cooperative systems, roadside object detection has become a key component in enhancing environmental perception accuracy due to its global perspective advantage. However, roadside traffic scenarios are characterized by dense pedestrians, frequent bicycle crossings, and significant shadow and light variations, leading to considerable challenges in detecting small objects, handling occlusions, and adapting to complex lighting conditions. To address these issues, this paper proposes an improved YOLOv8n model. First, a P2 small-object detection head is added to enhance the feature extraction capability for tiny objects. Second, deformable convolution DCNv3 is introduced to improve feature extraction and localization accuracy for overlapping objects by adjusting the dynamic receptive field. Finally, the CBAM dual-attention mechanism is integrated to enhance feature robustness in scenarios with varying illumination. To validate the model’s generalization ability, experiments were conducted simultaneously on a self-built campus dataset and the public DAIR-V2X-I dataset. Experimental results show that compared to the baseline YOLOv8n model, the improved model achieved accuracy increases of 3.9%, 9.5%, and 26% for small objects such as pedestrians, bicycles, and tricycles, respectively, along with a 7.6% improvement in recall and a 6.4% increase in mAP50. On the DAIR-V2X-I dataset, accuracy for the pedestrian, cyclist, and motorcyclist categories improved by 8.5%, 4%, and 4.7%, respectively, while recall increased by 5.6% and mAP50 by 5.3%. These results fully demonstrate the effectiveness and generalization capability of the proposed algorithm in complex roadside scenarios.