UR-YOLO: an urban road small object detection algorithm

The autonomous driving system heavily depends on perception algorithms to gather crucial information about the surrounding urban environment. However, detecting small objects on busy urban roads poses a significant challenge. To overcome this obstacle, we present UR-YOLO (Urban Roads-YOLO), a novel small object detection algorithm tailored for urban roads, which builds upon the enhanced YOLOv9 framework. UR-YOLO comprises three key enhancements. Firstly, to mitigate the high background ratio in small object datasets, we employ SCRConv to replace selected standard convolutions in the backbone network. The reduction in spatial redundancy sharpens the perception of vital features. Secondly, to address the sparse distribution of small objects, we incorporate SPPELANBRA, a refined version of SPPELAN, to enhance the model's sensitivity towards small objects, thereby improving its overall accuracy. Lastly, to address the issue of overlapping small objects, we upgrade the bounding box loss function by substituting the original SIoU loss with the Inner-MPDIoU loss. It not only improves the detection accuracy for small objects but also accelerates the convergence of the training process. To validate the effectiveness of UR-YOLO, we conducted comprehensive ablation and comparative experiments on the 2023 CICVAC dataset. The experimental results reveal that our proposed improvements have boosted the YOLOv9 model's mAP, precision, and recall by significant margins of 6.02%, 6.63%, and 4.81% respectively. Furthermore, when compared to prior YOLO series and two-stage detection models, UR-YOLO exhibits superior accuracy, higher frames per second, and greater robustness, making it a robust solution for diverse weather conditions on urban roads. Code is available at https://github.com/Ranghao/UR_YOLO.