Research On Lane Detection Method Based On Monocular Vision

At present, with the widespread use of the car, it brings lots of convenience to people.At the same time, it also brings some potential threat. People have been gradually paidattention to the security problem of cars. To improve the security problem of cars duringdriving, scholars at home and abroad have done lots of research driving assistance systemsof cars. With the development of machine vision, digital image processing and patternrecognition, as an important part of safety driving assistance system, the lane detectionbased on vision is becoming a hot research area of driving assistance systems.Visual navigation sensor is sensitive to illumination changes. When the light is toostrong or too weak, it will make the lane edge blur and the characteristic points extracted toolittle to recognize the lane line. This article mainly does some works about the lanedetection based on monocular vision under some kinds of road condition and the availability,real-time, adaptability of algorithms. In order to improve the adaptability of lane detection,the influence of illumination change and climate is considered. We did some experiments inthe modes of normal, strong light, weak light, night and snow respectively. Since the lanecan be divided in straight and curve, we used the linear model and the curve modelrespectively to fit lane lines. What’s more, we compared the algorithms in the aspectsof effectiveness and real-time. Some works about the problems in this article have done asfollows:(1) The algorithms of Naive Bayes and BP neural network are used to recognize the imagecategory.Images under different conditions need different image processing methods, so theimage category should be recognized. Firstly, through the experienced feature selectionmethod, extract the partial gray value of sky and road as feature vector. Then the images areclassified by the algorithm of Naive Bayes and BP neural network. At last, the identificationperformance and real-time of the two algorithms are compared.(2) For the non-standard conditions such as strong light and night road images, HistogramEqualization, Linear Transformation, Homomorphic System are used to enhance thecontrast.For the severe conditions of road images, in order to improve image’s quality, imageprocessing methods including Histogram Equalization, Linear Transformation,Homomorphic System enhancement are adopted to enhance the contrast of lane and road.(3) Image edge detection and feature point tracking search.Edge tracking. Firstly, detect the edge of image by Roberts-operator, Prewitt-operator,Sobel-operator,LOG-operator and compared the detection effect of them. Then the method of special directional search is used to track the boundary to filter out interference fringe.(4) An improved Hough transform and RANSAC algorithm are used to identify the modelparameters to fit the straight lane.Because of a special range constraints to the lane angle, we applied the improvedHough transform and the algorithm of RANSAC for the parameters identification of linearmodel to fit the straight lane and compared the two methods from the identificationperformance and real-time.(5) The model of parabola and B-spline curve is respectively established to fit curve lanemark.For the recognition of curve lane, parabolic model and B-spline curve mode areestablished. The particle swarm optimization (PSO) algorithm is used to identify theparameters of parabola model and the binary particle swarm optimization (BPSO) algorithmis used to optimize the number of control points of B-spline curve. Some experiments havebeen done to contrast the two fitting methods.