FDMB-YOLOv11: Traffic police command gesture recognition method under different lighting conditions

To address the challenges of false detection, insufficient feature representation, and limited real-time performance in image-based traffic police command gesture detection under different lighting conditions, this paper proposes a detection model named FDMB-YOLOv11 based on the YOLOv11n architecture. Firstly, the detection head was enhanced by introducing Frequency Adaptive Dilated Convolution (FADC), while the BiFormer attention mechanism was incorporated into the C2PSA module. Secondly, MobileNetV4 was adopted as the backbone network to reduce computational complexity. Finally, standard convolutions were replaced with Deformable Convolution to improve feature representation capability. Experimental results show that, based on the improvements described above, the precision, recall, and mean average precision (mAP@0.5) of the FDMB-YOLOv11 model under normal lighting conditions reach 97.91%, 93.32%, and 97.17% respectively, which are 23.1%, 13.63%, and 16.74% higher than those of the YOLOv11n model. The model’s parameter count is reduced from 2.58 million to 1.93 million, achieving lightweight optimization. Although the frame rate (FPS) is reduced to 65.789 fps, the model still achieves real-time inference performance and demonstrates potential for deployment on edge devices. The model outperforms target detection algorithms like SSD, Faster R-CNN, RTDETR, YOLOv12, and YOLOv13 by achieving the highest mAP@0.5. It also addresses the misclassification problem seen in the YOLOv11n model regarding left-turn and right-turn gestures while maintaining an optimal balance between accuracy and parameter count.