Research On Lane Line Detection And Traffic Sign Recognition Based On Deep Learning

With the continuous development of artificial intelligence,people's demand for automation equipment is getting higher and higher,and autonomous driving technology has gradually attracted people's attention.Lane line detection and recognition of traffic sign are very important components in environmental perception.Due to the complexities of scenarios,lane line and traffic sign detection based on traditional methods can no longer meet people's needs for autonomous driving technology.With the continuous rise of deep learning in recent years,the algorithm can be used in complex scenarios by introducing deep learning in target detection.Besides,It can also have excellent features.Based on deep learning theories,this thesis studies the lane line detection and traffic sign recognition technology.Following is the specific contents:First of all,addressing the problem that the existing YOLO v3 network object detection algorithms cannot classify traffic signs in detail,this thesis retrains the data set and readjusts the training parameters so that the accuracy of the detected YOLO v3 model will be improved and detailed classification of traffic signs can be achieved.Secondly,for solving a variable number of lane line detection problems,a method is first applied to turn the lane line detection problem into an instance segmentation problem,which is different from the existing methods that rely on a fixed and predefined perspective transformation matrix.In the line fitting,the depth separable convolution is introduced into the VGG16 network to generate a variable conversion coefficient.Experiments show that the method has a detection accuracy of 93.2%,which can handle a variable number of lanes and can respond to lane changes with strong robustness.Finally,in order to improve the accuracy and speed of lane line detection,this thesis proposes a method of combining a convolutional neural network with an automatic encoder to construct,construct a convolutional auto-encoder neural network(CAENN)with fewer layers.This method used the encoding and decoding characteristics of the encoder.The experimental results show that the detection speed can reach 46 FPS and the detection accuracy is 94.1%.This method can handle a variable number of lanes,achieving the advantages of fast detection speed and high accuracy.