Research On Traffic Sign Detection And Recognition Based On Deep Learning

Traffic sign recognition can deliver effective highway information to drivers and ensure smooth traffic and safe driving.Traffic sign detection and recognition based on computer vision is always the key and difficult problem in the field of target detection.In the real environment,the road situation is complex and changeable,which is susceptible to the influence of natural environmental factors.In addition,since the traffic signs themselves occupy a relatively small area in the whole image,they are easily diluted or interfered by the surrounding background or other small objects,affecting their detection and recognition on the road.In addition,current traffic sign detection and recognition systems are generally installed in intelligent mobile platforms such as intelligent cars,which have high requirements on the running speed of algorithms.However,existing algorithms are difficult to achieve a balance between speed and accuracy.Therefore,in view of the above problems,this paper studies the domestic traffic sign detection and recognition algorithm.Through a series of improvements to the network model,traffic signs can be quickly and accurately detected and recognized.The main work of this paper includes the following aspects:(1)A traffic sign detection and recognition algorithm based on improved YOLOv4 is proposed.As the current mainstream target detection algorithm,YOLOv4 has very strong detection ability,but the complex network structure leads to unsatisfactory reasoning speed.Therefore,in order to optimize the model and speed up the network operation speed,The main network of YOLOv4,CSPDarknet53,is replaced by Ghost Net.Ghost Net is a lightweight network,which uses linear operation to replace part of convolution operation and reduce the amount of network computation.The attention mechanism module inside the network strengthens the feature extraction ability of the main network for useful information.In addition,in order to further lightweight the network model,this paper simplifies the path aggregation network and fuses the convolution layer and the normalization layer,replaces the previous common convolution with the depth-separable convolution,and further reduces the number of parameters in the network.The experimental results show that the detection accuracy of the improved YOLOv4 algorithm is slightly reduced by 0.8%,and the model inference time is greatly shortened from 163 ms to 57 ms.The improved network model has a faster detection and recognition speed,and its accuracy is slightly lower than that of YOLOv4,but it is also better than most current target detection algorithms.(2)A traffic sign detection and recognition algorithm based on improved YOLOv5 is proposed.YOLOv5 network runs at a fast speed and meets the real-time requirements of the current lightweight traffic sign detection and recognition system.However,the detection accuracy of the model needs to be improved.The characteristics of the pyramids,multi-scale feature fusion and prior box office was improved,proposed circulation features before pyramid structure to replace the traditional pyramid structure,make the backbone network in feature extraction can self repeatedly study reinforces model perception of information at the same time can improve the model of small target detection ability.Secondly,an adaptive feature fusion module is added in front of the detection layer to improve network stability and information processing ability.Finally,combining with the appearance characteristics of traffic signs,k-means algorithm is used to re-cluster the prior frames suitable for detecting and identifying traffic signs,which helps the model to locate and identify traffic signs more quickly and accurately.The experimental results show that the detection accuracy of the improved YOLOv5 algorithm is improved by 3%,the model reasoning time is slightly increased by 19 ms,the reasoning speed of the improved YOLOv5 algorithm is slightly increased compared with the original algorithm,and the detection and recognition ability is significantly enhanced.