Traffic Signs Detection Based on Enhanced YOLOv5 Network Model

In this paper, in order to improve the accuracy of YOLOv5s6 in traffic sign detection, we propose two types of methods to improve the problem of missing detection and low accuracy in small traffic signs. Our type 1 method is based on the stacking of residual blocks to increase the depth of the network, and we add channel attention Squeeze-and-Excitation (SE) blocks to increase the learning ability of the network. In type 2 method, we halve the channel of type 1, reducing computation and parameters, and add an additional 3.3 convolution to improve the learning ability of the model. We choose the traffic signs dataset Tsinghua-Tencent 100K (TT100K) to train the model. TT100K contains 221 categories. The large categories make detection difficult. In addition, we collected Taiwan traffic signs as a customized dataset to validate our method. We also test the performance of the proposed methods on 3 public datasets. The experimental results show that in the TT100K dataset, the mAP of type 1 method is increased by 1.9% and the mAP of type 2 method is increased by 3.2%. In the customized Taiwan traffic signs dataset, and the mAP are increased by 4.9% and 5.7% for type 1 method and type 2 method, respectively.