Detection And Classification Of Traffic Sign Targets In Natura Scenes

The detection and classification of traffic sign targets are an important part of achieving intelligent driving,which not only can assist the driver to drive safely,but also an important part of the realization of unmanned driving.In natural scenes,the detection and classification of traffic sign targets are very challenging due to factors such as different weather,light intensity,sign deformation,and sign deflection.In addition,in order to consider the influence of various background interference factors on the road,the generally collected pictures are high-resolution distant pictures,so that the traffic sign targets occupy a small proportion in the pictures.Existing traffic sign target detection and classification recognition algorithms cannot solve these problems well.This thesis combines deep learning and image processing algorithms to achieve rapid and accurate detection and classification of traffic sign targets for the above problems.The main research content has the following aspects:(1)The characteristics of traffic signs are analyzed,and several data sets commonly used to detect and identify traffic signs are introduced.The common algorithms of traffic sign detection and recognition are expounded,and the difficulties of traffic sign target detection and classification recognition are summarized.Then introduced the specific content of deep learning YOLOv3 network model and transfer learning algorithm.(2)A method of traffic sign target detection and classification recognition based on improved YOLOv3 is proposed.Aiming at the problem of small traffic sign target in the collected long-distance pictures,a TSR-YOLOv3 deep neural network is proposed.By improving the previous feature extraction residual network,and merging the shallow network features in the small-scale detection layer to achieve accurate identification of small-scale.In addition,the data set under various severe weather conditions is used to display a variety of road environments in natural scenes,the dark channel prior defogging algorithm is used to enhance the training image,and the TSR-YOLOv3 network is first cropped for low resolution Training on the rate picture,and then transfer learning and retraining on the high-definition vision picture,in order to accelerate the convergence of the network.(3)A method for detecting and classifying traffic sign targets based on color and shape is proposed.In order to overcome the problem of different light intensities,the HSV threshold segmentation method that is less affected by light combined with morphological image processing method is used to detect traffic signs.Then the convex hull is detected to fit the polygon area,and the radian value of the polygon and the rotation angle approximately parallel to the horizontal line are solved.The target area is set within a certain range,which is uncontrollable for the sign deflection,sign deformation and shooting angle,etc.The influence of factors is very robust.Finally,we use the improved Le Net-5 neural network combined with the SVM classifier to classify the obtained target area,and solve the network overfitting problem by using the square hinge loss instead of the cross entropy loss.