Deep learning has been widely used in environment perception of autonomous driving field，Its application on multi-object traffic sign (such as shield, overlap, incomplete, and small goals) detection on urban roads is current research emphasis. This paper proposed an improved detection method based on YOLOv3. Firstly, it introduced the depthwise separable convolution layer into the backbone network of YOLOv3 to optimize the parameters and reduce the quantity of calculation in convolution neural network. Secondly, the spatial channel attention mechanism (CBAM) was introduced after the residual module in the backbone network to, aiming to Enhance the ability of the network to extract weak feature information, and improve the detection accuracy of small target traffic signs. Finally, the optimized intersection ratio function of IOU combining with the CIOU function, can improve the target detection accuracy of the candidate box screening. The experiments were conducted with CSTSDB open source traffic sign dataset and partial self-built dataset. The experimental results show that the improved YOLOv3 network improves the accuracy rate has been improved by about 7% than the original YOLOv3 detection algorithm, and has a lower leakage detection rate, faster speed, which has some practical significance.