Traffic Sign Recognition Model Based on Small Object Detection

Fei Gao, Wanjun Huang, Xiuqi Chen, Libo Weng

Abstract: Traffic sign recognition is a popular task in the field of computer vision, but it faces challenges such as small object size, complex scenes, and real-time requirements. In this paper, a model framework is proposed to comprehensively improves the detection and recognition of traffic signs. To address small objects, FlexCut data augmentation is introduced, which generates non-repetitive sub-images through the strategy of maximizing sample region cropping. This approach enhances the detection ability of small objects. Additionally, PIoU loss function based on keypoints is also investigated, which accurately guides the position and shape of the bounding boxes by considering factors such as overlap area, distance, aspect ratio, and geometric size. Furthermore, the YOLOv5s network is enhanced by integrating the TransformerBlock module, SimAM attention mechanism, and Decoupled detection head to enhance the receptive field and feature extraction capability. In the experiments conducted on the TT100K dataset,, the proposed YOLOv5T achieves significant performance with an mAP@0.5 of 87.5% and an mAP@0.5:0.95 of 66.1%. These results validate the effectiveness of the proposed approach in addressing traffic sign recognition problems.

External IDs:dblp:conf/pricai/GaoHCW23