Research on traffic sign detection method based on multi-scale feature fusion

SSD algorithm, as one of the mainstream detection frameworks, is a one-stage algorithm to obtain the target category and location by direct regression, which can act on the output feature map of the convolutional network for prediction and complete the target localization and classification at one time, but it is poorly applied in practical problems due to low accuracy of small target detection, and low accuracy of traffic sign recognition and weak generalization ability. There are countless types of traffic signs, and there are more than one hundred kinds of traffic signs in China, which can be divided into warning signs with black borders on yellow background, prohibition signs with red circles on white background, and instruction signs with white characters on blue background by category, and the shapes are mainly triangular, circular and rectangular. By loading the pre-trained model to initialize the network model parameters and adjusting the parameters in the aspect ratio set to reduce the number of default boxes generated, the training of the network model and the convergence speed are improved. Finally, after experimental comparison and analysis, the feasibility of the improved algorithm proposed in this paper is proved.