Long-Tailed Traffic Sign Detection Using Local Cross-Channel Interaction for Complex Scenes

Traffic sign detection technology has broad application prospect in the area of intelligent driving and has received extensive attention in recent years. However, in reality, traffic signs usually obey a long tail distribution, and the dimensions of traffic symbol in images are comparatively less, which makes it extremely difficult for traffic symbol sign detection. This paper proposes a traffic symbol detection model based on local cross-channel interaction specifically to resolve these issues. Firstly, this paper proposes a C2f_EC module using the Efficient Channel Attention Structure (ECA), which improves the detection of tail class and small object by maintaining inter-channel correlations through localized cross-channel interaction information and by performing inter-channel interactions more efficiently. Secondly, this paper adds a small object detection layer and replaces the CIoU (Wise Intersection over Union) with the AlphaInnerIoU, which improve the accuracy of detection small parts and show stronger robustness and better generalization performance. The mAP50 (mean Average Precision), mAP<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50:95</sub>, mAPt and mAPs metrics on the self-made TT100K_2021LT dataset are improved by 4.1%, 4.2%, 4.2% and 5.1% compared to the base network. The P (precision) and R (recall) on the GTSDB dataset are improved by 1.1% and 8% compared with the base network. The experimental results show that this method maintains the optimal detection speed for traffic sign detection in complex scenes while effectively improving the detection performance of tail class targets and small targets.