Lane Line Detection Method Based On Fully Connected Conditional Random Fields

With the rapid development of the social economy and the improvement of the level of science and technology,the number of cars in China is also increasing,and the possibility of traffic accidents is also increasing accordingly.The issue of car driving safety has become a problem that must be studied and discussed.Autonomous driving systems can effectively reduce the incidence of traffic accidents,which has become a research hotspot in recent years.Lane line detection is a very important part of the automatic driving system.Real-time and accurate detection of lane lines is the most basic requirement for automatic driving.In recent years,due to the rapid development of deep learning technology in the application field of computer vision,the lane line detection algorithm based on deep learning has achieved remarkable results.However,due to the complexity of the traffic environment,the category information output by the deep learning algorithm will have some Misclassified points cause rough edges and noise in the detection results of lane lines.In response to this problem,this paper designs a lane line feature extraction algorithm that combines deep learning and post-processing methods,and fits the extracted feature results to obtain the final lane line model.The main research contents of this thesis are as follows:First,a lane line dataset is made based on real road scenes.Use the front-end camera of the self-driving car Sharing Van1.0+of Dongfeng USharing Technology Co.,Ltd to collect data,process and filter the collected data,and mark the lane lines on the filtered pictures by manual labeling,so as to obtain the lane Measured dataset of lines.Provide data support for the training and testing of subsequent algorithms.Second,feature extraction is performed on the lane lines.Aiming at some misclassified points in the category information output by the deep learning algorithm,this thesis designs a lane line feature extraction algorithm combining the ENet-SAD algorithm and the fully connected conditional random field.The algorithm is tested and verified on the CULane data set and the measured data set,compared with other typical algorithms,and the results are visualized.The results show that the algorithm designed in this thesis is an effective lane feature extraction algorithm.Finally,fit the extracted lane line feature data.In order to improve the fitting accuracy and efficiency,a feature point extraction method is proposed,which can screen the feature points of the lane line according to the position relationship and category relationship of the lane line pixels.The selected lane line feature points are fitted by a quadratic polynomial curve model based on the least squares method.The results of the point picking algorithm and the results of the lane line fitting algorithm are visualized,and the results show that the point picking algorithm proposed in this thesis can fit the lane line model well.