Detection And Recognition Of Traffic Signs In Natural Scenes

The detection and recognition system of traffic signs is a key subsystem for realizing intelligent transportation systems,and it has a wide range of applications in the fields of assisted driving,intelligent navigation,automatic driving,smart transportation,and maintenance of traffic signs^[1].At the same time,for safety reasons,the traffic sign detection and recognition system needs to be both accurate and fast.The existing traditional traffic sign detection and recognition algorithms are difficult to adapt to various lighting conditions,complex backgrounds,and under-targets in natural scenes;the existing traffic sign detection and recognition algorithms based on deep learning methods are difficult to have both algorithm performance and speed in vehicle scenes.Therefore,this paper attempts to apply deep learning methods to the detection and recognition of traffic signs,and accelerate the process,which is both accurate and fast in natural scenes.This article aims to detect and recognize traffic signs in natural scenes,and the algorithm needs to be both performance and real-time.In order to ensure the accuracy of the algorithm,this paper proposes a YOLOv3-SPP detection network with channel attention;in order to ensure the real-time performance of the algorithm,this paper prunes the detection model for the channel and residual block,while using a light recognition model.Among them,the main work and innovations are as follows:?1?A YOLOv3-SPP detection network with channel attention mechanism is proposed.This paper combines the YOLOv3-SPP detection network and the SE structure,uses the Focal Loss in the confidence loss and classification loss,and uses the GIo U Loss in the bounding boxes,and customizes the anchor for traffic sign detection.Experiments prove that the model can detect traffic signs in natural scenes well.?2?The model compression method based on channel pruning and residual block pruning is used.In this paper,the gamma scale factor of the BN layer is used to induce sparseness of the model.The merging pruning channel mask is used to prune the convolutional layer which is connected with the residual layer,and the threshold pruning mask is used to prune the ordinary convolutional layer;through evaluating all the residual blocks uniformly,the unimportant residual blocks are then pruned.On the premise of ensuring the performance of the model,this hybrid pruning strategy can greatly compress model parameters,reduce the amount of FLOPs required by the model,and reduce the time consumed by the traffic sign detection algorithm.?3?Light traffic signs classification network based on Mobile Netv3 is used.The configuration parameters of the classification network used in this paper are obtained through neural architecture search,inheriting depthwise separable convolution,inverse residual structure,and linear bottleneck structure.The network can take into account both the accuracy and speed of classification to meet all requirements in the traffic sign recognition stage.The experimental results show that the pruning detection model in this paper can achieve a recall of 91.8%,a precision of 87.6%,and the fastest speed of more than 100FPS in the test set;the accuracy of the classification model used in this paper is 94.6%.After the detection and recognition model is cascaded,it can accurately and quickly detect and recognize the traffic signs in natural scenes.