AK-YOLO: a lightweight scale-adaptive model for traffic sign detection

Based on the double challenges of the multi-scale changing characteristics of traffic signs and the limited resources of on-board equipment, this study proposes a light-weight scale adaptive traffic sign detection model, AK-YOLO. The model uses a multi-tiered optimization architecture. In the feature extraction layer, we have designed a two-way feature extraction module, C3AK. The static path of this module reduces computational complexity by decoupling spatial filtering and channel fusion. Dynamic paths enable multi-scale feature extraction by adjusting the sensory field. The network architecture layer fuses the multi-scale feature extraction advantages of spatial pyramids with the efficient computational properties of cross-stage channel connectivity. The training optimization layer establishes a dynamic weight adjustment mechanism based on sample difficulty with adaptive threshold focus loss. Existing traffic sign datasets mostly follow the UNECE R68 standard, which is systematically different from China’s GB5768. We constructed a standardized annotation dataset (URS) based on functional characteristics, containing five major categories and fifteen subcategories of injunctions, warnings, directions, instructions, and aids. The output of the experiment indicated that the URS value of AK-YOLO reached 98.3% and the parameters of the model were reduced by 57.4% compared to the baseline model. The tested on the TT100k public dataset achieved 97.4%. AK-YOLO provides a reliable traffic sign detection solution for automated driving systems.