Small object detection algorithm based on improved YOLOv10 for traffic sign

Abstract

Traffic sign detection (TSD) is a critical task in intelligent transportation systems (ITS) and autonomous driving, facing challenges such as complex backgrounds and small-scale objects. Existing methods often suffer from high miss and false alarm rates, particularly in dynamic or cluttered environments, limiting their practical applicability. To address these issues, we propose LTS-YOLOv10, an improved version of YOLOv10, designed to enhance small object detection accuracy and overall performance in complex real-world conditions. Our approach introduces Omni-Dimensional Dynamic Convolution (ODConv), which utilizes a four-dimensional dynamic convolution mechanism to improve the capture of multi-scale and complex background features. Additionally, we integrate an attention-guided bidirectional feature pyramid network (EMA-BiFPN) to enhance feature fusion, further improving the detection accuracy for small objects. The MPDIoU loss function is employed during bounding box regression to optimize precision and recall for irregularly shaped targets. Experimental results on three public datasets demonstrate that LTS-YOLOv10 achieves a 3.8% improvement in mAP on the CCTSDB dataset compared to the original YOLOv10, with notable gains on the TT100K and DFG datasets as well. These improvements are achieved with only a slight increase in parameters, demonstrating the model’s superiority in terms of accuracy, robustness, and real-time performance. LTS-YOLOv10 provides a promising solution for practical traffic sign detection, with future work focusing on further enhancing the model’s real-time capabilities and optimizing its application in edge computing environments.