Research On Lane Line Recognition Algorithms Based On Deep Learning Technology

Lane line recognition algorithms are important parts of the advanced assisted driving system and can provide reliable lane line data.This paper combines state of the art deep learning techniques with traditional image processing methods,and uses TensorFlow and OpenCV to design lane recognition algorithms to complete lane line feature extraction,lane line model establishment and so on.The algorithms presented in this work incorporate the advantages of high image recognition rate,the efficient image handling performance,and robustness that ensures recognition both in the shadow of the vehicles and in the environment with strong ambient illumination.Image datasets are used to train the proposed neural network model.In this paper,the original road image is labeled with LabelMe,and the lane line features in the original road image are manually extracted.Then,the image preprocessing methods are used to process the annotation image and the original road image,so that the data structure of the images conforms to the requirements of the model.Finally,the preprocessed images are encapsulated into image datasets by the data format of TFRecord.Lane line features are extracted by the neural network model.In this paper,TensorFlow is utilized to construct the neural network model.The linear weighted sum method is applied to calculate the input data of the deconvolution layer so as to weaken the effect of the resolution reduction of the output image caused by the pooling layer.In model’s training,the cross entropy function is defined as the loss function,and the gradient descent and moving average algorithms are used to update and optimize the parameters of the model,respectively.After 100,000 rounds of training,by analyzing the values of the learning rate,the decay of the moving average algorithm,the loss function and the weights,it can be known that the model tends to be stable and form a proprietary feature extraction capability.Finally,the method of extracting lane line features from the original road image using the trained model is introduced.The lane line model generates lane line data.The binarization method with variable thresholds uses F₄-measure as the binarization evaluation criterion,and obtains the relationship between the optimal binarization threshold and the sum of pixel values from the randomly selected image samples by statistical methods.Compared with the traditional binarization method,the binarization method with variable thresholds increases the growth value of F₄ by at least 1.68%.According to the shape characteristics of the lane line,the image scanning conditions are set.The interference items in the image can be effectively removed,and the lane line coordinate points obtained by the image scanning can accurately fall into the lane line area.The lane line function for fitting the shape of the lane line can dynamically adjust the highest degree according to the number of captured lane line coordinate points.The method of solving the linear equation by the inverse matrix can quickly obtain the values of unknown coefficients in the lane line function.In order to ensure the validity of the solution method,this paper proposes and proves the reversibility problem of the coefficient matrix under the restrictive condition.The lane line recognition system integrates lane line recognition algorithms.Firstly,this article describes the hardware and software environment on which the lane recognition system operates,and introduces the working principle of the system.Then,by testing the time taken by the system to complete the lane line recognition task for an original road image in the set hardware environment,it is calculated that the processing ability of the system is about 14 Hz.The average values of Precision,Recall and F₁obtained from the randomly selected 400 original road images under the F₁-measure evaluation method are 92.433%,96.439%and 94.210%,respectively.It indicates that the overall performance of the system is at an advanced level among similar systems.