LAUSR

Traffic sign detection is an important research direction in computer vision, which is of great significance for autonomous driving and advanced assisted driving systems. Due to the complexity of traffic signs in natural scenes, existing traffic sign detection algorithms have disadvantages such as high false detection rates and poor robustness. To improve the accuracy of traffic sign detection, a feature expression enhanced SSD(ESSD) detection algorithm is proposed. ESSD extracts semantic information in a lightweight way, adds detailed information for fusion, and forms a new feature map through multiple convolution operations to enhance feature expression. Meanwhile, a new target default box was designed to increase the focus on traffic signs. The SSD and ESSD were retrained on TT100K and CCTSDB datasets. Experimental results show that the mAP of the improved ESSD is 81.26% and 90.52% and can improve AP up to 40%. The robustness of the ESSD model was verified using the PASCAL VOC data set, which showed better detection of small objects.