Abstract: Aiming at the problem of low accuracy of small-size traffic sign detection in intelligent driving scenarios, a traffic sign detection algorithm based on Tiny YOLOv3 network was proposed. By using the depth separable convolutional reconstruction feature extraction network and increasing the feature fusion between the shallow and deep feature layers, the model's attention to small targets could be improved; at the same time, the accuracy of the prediction box was also improved by modifying the size of anchor boxes. The test results on the TT100K traffic sign data set showed that the mean average precision（mAP）of the proposed algorithm was 19.3% higher than that of Tiny YOLOv3, and it is more robust to the detection of small-size traffic signs.

Key words: traffic sign detection, depth separable convolution, feature fusion, Tiny YOLOv3