Traffic sign detection algorithm based on deep learning

In this paper, an improved algorithm based on YOLOv8s is proposed to solve the problems of small target, low precision and high computation in traffic sign detection. Combining Dual and HetConv ideas, the CSPHet structure is proposed to replace BottleNeck in C2f module, reduce parameters and computation load, and improve accuracy. Depth separable Convolution (DWConv) is used to accelerate inference, and 160×160 scale detection head is added to improve detection accuracy. UIOU loss function is introduced to solve the problem of sample imbalance, and can be combined with any bounding box regression loss function. Experimental results on the TT-100K dataset, mAP50/% increased by 4.6%, the number of parameters decreased by 48%, and the amount of computation decreased by 11%.