Research On Traffic Sign Detection And Recognition Algorithms Based On Machine Vision

Traffic sign detection and recognition is the core technology of the assisted driving system and the unmanned driving field.So it has wide application value in vehicle safety warning,car navigation and sign maintenance.However,the current traffic environment is complex and changeable.Unfavorable factors such as illumination and weather changes,sign occlusion and deformation,fading and blurring make the research of this technology face many difficulties.Therefore,this paper summarizes the research achievements of related technologies at home and abroad.And we conduct further research on traffic sign detection and recognition methods by analyzing the problems of it.The main research contents of this paper are as follows:(1)In the traffic sign detection stage,a kind of multi-feature fusion traffic sign detection algorithm is proposed by making full use of the color,shape,gradient and texture features of the traffic sign itself in this paper.Firstly,aiming at the problem that traditional color segmentation methods are susceptible to illumination,according to the distribution characteristics of image brightness histogram,single-scale Retinex algorithm is used to improve the image quality of abnormal illumination.And based on this,a hybrid segmentation strategy combining multiple color spaces is proposed,which effectively improves the color segmentation effect of low-brightness images.Then based on the shape features of the traffic signs,the morphological operations and geometric feature are used to screen out the rough candidate regions.Finally,in order to improve the accuracy and speed of detection,the local HOG features and adaptive ULBP features of candidate regions are fused,and an efficient SVM classifier is used to make linear judgments on fusion features.Thus,the accurate detection of traffic signs is realized.The experimental results show that the detection effect of this method is greatly improved in terms of accuracy and real-time.(2)In the traffic sign recognition stage,mainstream deep learning algorithm is applied for the purpose of avoiding the limitations of manual design features.For the problems of high computation cost and poor real-time performance in current convolutional neural networks,a lightweight convolutional neural network algorithm for traffic sign recognition is proposed.First,redundant characterization of standard convolution is removed by using deep separable convolution,which can reduce computational complexity of the model.Then,a linear bottleneck inverse residues structure is used to deepen the network layer to fully extract image features.And on this basis,a feature calibration mechanism is added to selectively enhance some useful feature maps information.Finally,In order to improve the generalization ability of the neural network,the focal loss function is introduced to reduce the influence of sample size imbalance on the model accuracy,which especially improve the recognition ability of difficult samples.Experimental results show that the computational complexity of this method is significantly lower than that of AlexNet,and its classification accuracy and real-time performance are effectively improved.(3)A more realistic low-resolution traffic sign data set based on TT-100 K data set is made in this paper.We also develop the system software for algorithm verification on the QT platform.The system can realize real-time detection and recognition of traffic signs,which has certain application value.