Research On Lane Detection On Structured Road In Various Scenes

Lane detection from complex environment is a key factor for advanced driver assistant technologies and research on lane detection on structured road in various scenes is important and full of meaning.This thesis analyzed the current solution methods of lane detection both at home and abroad and found that hand-crafted visual cues were needed in the traditional lane detection methods.Traditional lane detection algorithms used special filters to detect lane area,and then combined with Hough transform,RANSAC algorithms to find the lane markings.These algorithms always needed to adjust filter operators manually,set adjustment parameters according to the street scene characteristics,the process was work load and algorithms had poor detection results when driving environment changed significantly.What's more,traditional lane detection methods did not focus on the identification of multi-functions of lane markings.Based on the current lane detection research,an efficient lane detection approach on structured road in various scenes based on deep learning method was proposed in this thesis.The main contents in this presentation are as follows:(1)Road image preprocessing: The classification standard of lane quality level was established,and the multi-scene of lane markings was defined.A lane marking label method was proposed to extract the feature for the lane detection network.The lane marking label method was not only beneficial to the detection of lane marking,but also provided line type information.In addition,Image preprocessing was performed on the marked images,including Region of Interest(ROI),Inverse Perspective Mapping(IPM),image grayscale and image enhancement.The results showed that the image preprocessing method in this thesis preserved the lane markings information maximally and provided good data information for the lane marking recognition algorithm.(2)Lane Detection and Tracking: An algorithm of lane detection and tracking based on deep learning was proposed.The method included two parts,lane detection based on Convolutional Neural Network(CNN)and lane tracking based on Recurrent Neural Network(RNN).By training the convolution neural network,the lane markings features were extracted firstly,then the extracted lane markings characteristics were transformed to the convolutional neural network to further extract the advanced features of the inter-lane markings in the time dimension.According to lane detection and tracking neural network,the lane markings feature points extraction and lane multi-scene functions recognition were achieved.(3)Lane fitting and post-optimization method: According to the characteristic points of the lane,the dynamic programming algorithm was proposed to find the optimal path of the lane and the least square method was used to fit the quadratic polynomial of the lane.The effectiveness of the proposed lane detection algorithm in multi-scene was validated and contrasted.The results showed that the lane detection algorithm proposed in this thesis achieved recognition accuracy of more than 90% for lane multi-purpose identification,and the accuracy of lane detection can reach more than 95%,which is much higher than the traditional algorithm.