Research On Traffic Sign Recognition Based On Improved YOLOv3 Network In Natural Environment

Traffic sign detection and recognition is one of the key technologies in assisted driving,unmanned driving,intelligent navigation and other fields.Accurately obtaining traffic sign information in front of the road can provide decision support for the driver or intelligent vehicle to perform correct actions,reduce operational errors,and reduce the frequency of traffic accidents.For the safety of people,the detection and identification of traffic signs must be accurate and fast.Based on the research and analysis of traffic sign detection and recognition technology at home and aboard,this thesis proposes a method of traffic sign detection and recognition based on improved YOLOv3 network in natural environment.The main researches are as follows:1.The imgaug library was used to enhance the GTSDB dataset,and simulate the generation of data samples in the case of rainy,snowy,hazy,cloudy,sunny and foggy weather.Some traffic sign pictures with a small amount in that data set had been zoomed by 80%～180%.The network model trained on the enhanced dataset could be adapted to detect and recognize traffic signs in real road scenes and in bad weather environments accurately.2.The K-means algorithm was used by to cluster 9 anchor points on the COCO dataset in YOLOv3,and some of them were not applicable on the GTSDB dataset.It was proposed that using the improved K-means++algorithm to re-cluster 9 new anchor points on the GTSDB dataset.It could improve the detection rate of bounding boxes in the model.3.When training a deep learning neural network model,the BN layer could accelerate network convergence and prevent overfitting.After the BN layer normalized the data,it could solve the problem of gradient disappearance and gradient explosion effectively.But in the forward inference of the network,there were some extra operations of some layers,which affected the performance of the model.It was proposed that merging the parameters of the BN layer into the convolutional layer.In that case,the speed of the model’s forward inference would be increased.4.For the small target of traffic signs recognition task in this thesis,the low-level features rich in detailed information were not fully utilized just using the three scale features of the YOLOv3 network.That was not conducive to the detection of small targets.The improved multi-scale prediction of YOLOv3 network was proposed.This thesis proposes 2x upsampling for the feature map down-sampled by 8x of the previous network,and the feature map upsampled by 2x is concatenated with the output of the second ResNet block unit.A feature fusion target detection layer,whose feature map is down-sampled by 4x is established.The model’s detection and recognition of small target traffic signs was improved.The traffic sign recognition method based on the improved YOLOv3 network proposed in this paper was trained and tested on the GTSDB dataset.Experimental results show that the method proposed in this paper had a good recognition effect on 43 types of traffic signs,with an average recognition accuracy rate of 98.1%,an FPS index of 28.64,and a single picture recognition speed of 34.92ms.The model in.this paper was tested on pictures and videos which were not in the GTSDB dataset,and a good recognition effect was also obtained.It indicated that the model of this paper has robustness.