The invention discloses a live panoramic traffic sign picture preprocessing method. Statistical analysis is performed on traffic signs in a real-time panoramic traffic image training set by means of ageometric perspective principle. The number distribution rule of traffic signs in a panoramic image is mainly counted; preprocessing the panoramic traffic image to be detected according to the statistical result; sending the processed image blocks into an inference network for prediction; the prediction result is processed through the NMS algorithm, the optimal prediction frame is selected and mapped to the original panoramic traffic image to be output, and the problems that in the live panoramic traffic sign image prediction stage, due to image compression, the prediction result is not ideal, sliding window processing is directly conducted on the original image, and calculation resources are excessively consumed are solved.