Lane Detection And Recognition Based On Vision In The Complex Illumination Condition

Lane detection and recognition is an important sub-problem of the vehicle driverassistance technology, it plays an essential role in many application field such as lanedeparture warning system, vehicle collision avoidance system. At present, large numbertheories of vision-based vehicle assistance driving technology have been proposed, andsome vehicles equipped with the corresponding products are available on the market.However, the detection and recognition of lane needs to deal with varieties of roadconditions, such as visibility change caused by weather changes, at twilight or at night,the road scene change (tunnel, urban), shadow result from the multiple backgroundobjects or its dynamic changes, as well as another variation lighting conditions, itsreduce the quality of the video or image of traffic, thus making the lane detection andrecognition more difficult. Therefore, study of the technology aimed at handling andadapting to complex scenes in lane detection and recognition is of great significance.In this thesis, we study the lane detection and recognition based on vision, and wefocus on the research under complex lighting conditions. We will be use the videos orimages of road condition collected by the CCD camera placed on the windshield of avehicle as test data. After processing the light and noise of the test images, we study thecontext3D features, including the horizon line, the vanishing point of lane,3D scenelayout and road geometry features, and discuss the correlation characteristics betweenthe continuous image sequences, finally, these features are fused to build a monocularvision-based lane detection and recognition system. experiments on data with a varietyof lighting conditions show that our system that combines multiple3D features iseffective，of high accuracy, robustness, and it has a good real-time performance.The main work of the thesis is as follows:(1) We studied the vertical gray_scale step search method, and then used it toextract the region-of-interest (ROI) in lane marking line detection and recognition, thusremoved out a vast background area effectively.(2) We studied a method combining the Hough transform and a random sampleconsensus (RANSAC) algorithm to implement lane detection and recognition, which can eliminate most of the "outlier" data points from the main data group during the lanemarking line fitting procedure.(3) We studied a weighted moving average method to execute time sequencesanalysis to the input data, which can be controlled flexibly by adjusting its algorithmparameters, and then reduce the interference caused by the dynamic changes of thevarious background objects.(4) Through the procedure of the lane feature detection and the lane recognition,we studied a method that integrated a variety of low-level features and context3Dinformation to support the final recognition output.