Abstract:Traffic sign detection in complex environments is susceptible to interferences such as strong light and occlusions from rain or snow, leading to diminished detection accuracy and missed targets. To enhance model robustness in such scenarios, this study proposes an improved algorithm based on YOLOv8n, termed RDCM-YOLO. Firstly, the receptive-field dynamic convolution (RFDConv) is innovatively introduced into the backbone network. By combining receptive field attention with adaptive convolutional kernels, RFDConv aims to reduce redundant computations while simultaneously augmenting the model’s feature perception capabilities in complex environments. Secondly, an inverted context anchor attention (iCAA) module is designed by embedding the context anchor attention (CAA) into an inverted residual structure. This design guides the model to focus on critical regions and optimizes the interaction and fusion of local and global information. Furthermore, a mixed aggregation network (MAN) is incorporated, employing a multi-branch feature aggregation strategy to strengthen the model’s ability to integrate cross-scale semantic information. Experimental results on the augmented dataset TT100K_aug reveal that RDCM-YOLO achieves improvements of 3.9%, 5.4%, 5.4%, and 4.5% in precision, recall, mAP@0.5 and mAP@0.5:0.95 respectively, compared to the baseline YOLOv8n. Its overall performance is superior to that of current mainstream traffic sign detection algorithms designed for complex environments.