Research On Traffic Sign Detection And Recognition Algorithm Based On Neural Network

With the development of image sensor technology,people can easily obtain valuable image information.Therefore,a large number of scholars are engaged in scientific research on image processing technology,and image processing technology has been developed.The perception module of automatic driving system based on computer vision has been studied by a large number of scholars.In short,automatic driving does not require the driver's participation in the entire driving process,and its development can avoid driving accidents caused by human factors.Therefore,the development of this technology can really improve the level of road safety.The visual perception module in the automatic driving system can complete tasks including pedestrian detection,lane line detection,traffic sign recognition,vehicle detection and traffic signal detection.Lane line detection and traffic signal detection in traffic scenarios are the tasks to be completed in this thesis.By analyzing the advantages and disadvantages of traditional image processing methods and deep learning methods to achieve detection,this thesis designs a lane line detection algorithm and traffic signal detection algorithm based on convolutional neural network,and designed a network that can complete the above two functions at the same timeThe first to be designed is the lane line detection network which is essentially a task of semantic segmentation.According to the research on semantic segmentation network,Deeplabv3 plus is designed to be a dilated convolution with non-single expansion coefficient and a lane line detection network with dense connection concept.First of all,the Res Net101 network is designed as a feature extraction network,and is added to the dilated convolution of non-single expansion coefficients to achieve the purpose of improvement.Then,the extracted features are input into the improved ASPP module with dense connections.Finally,the lane line detection result is the deconvolution and upsampling of the features.It can be seen from the experimental results that the lane detection network designed in this thesis can improve the segmentation accuracy compared with the network structure before improvement.Secondly,the traffic signal detection network is designed.It is essentially a target detection task.Through the analysis of the target detection network,Faster RCNN has been improved into a traffic signal detection algorithm.The feature extraction network follows the lane detection network.The feature pyramid model was introduced to increase the perception of traffic signals.Experimental results show that the network designed in this thesis improves the recognition accuracy in traffic signal detection compared with the original Faster RCNN algorithm.Finally,this thesis designs a multi-task network that simultaneously implements lane line detection and traffic signal detection tasks.The design basis of the network is the feature extraction network designed above.The extracted features are input to the lane line and traffic signal decoder.Experimental results show that the multi-task neural network structure designed in this thesis can ensure the detection of lane lines and traffic lights in traffic scenes while ensuring that the detection accuracy is not reduced.