GRFS-YOLOv8: an efficient traffic sign detection algorithm based on multiscale features and enhanced path aggregation

Traffic sign detection is a crucial element of advanced driver assistance systems (ADAS) for environmental perception. However, challenges persist in the detection of small-scale targets and susceptibility to adverse weather, varying light conditions, and occlusions. To address this issue, a novel traffic sign detection algorithm, GRFS-YOLOv8, is proposed. GRFS-YOLOv8 introduces an enhanced greater receptive field-SPPF (GRF-SPPF) module to replace the original SPPF module, enabling the capture of richer multiscale features from image feature maps. Additionally, a new SPAnet architecture is designed by introducing two “shortcut” paths and additional smaller target detection layers to enhance path aggregation capabilities. This architecture propagates complete semantic information, alleviating the reduction in resolution for small targets and enhancing the model’s capability to detect them. Finally, by employing GhostConv and C2fGhost to replace multiple CBS and C2f modules in Backbone and Neck, cost-effective linear operations are utilized to obtain more feature maps, thus reducing the computational cost of the model. Experimental validation across multiple datasets demonstrates the efficacy and adaptability of GRFS-YOLOv8, achieving an 80.3% mAP and 72.4% Recall in CCTSDB 2021, a 71.2% mAP and 95% Recall in TT100K, and a 94.0% mAP and 96.0% Recall in GTSDB, surpassing mainstream detectors and comparative methods.