Road Detection Under Complicated Environment

Road detection based on computer vision is a key technology in autonomousdriving system of intelligent vehicles. Although many scholars have made greatefforts on the research of this field, the accurate detection of road is still faced withmany problems because of the variety and complexity of the environment, such asshadows, illumination variance, pavement dilapidation, vehicle occlusion andbackground interference. In order to overcome these above problems, in this thesisdeep research is carried out for structured and unstructured road to realize reliableroad detection under complex background. The main contents of this thesis are asfollows:(1) A lane detection algorithm based on linear-hyperbola model is proposed,which is to overcome the difficulties that simple road model can’t descript differentshapes of road, and complex road model is vulnerable to be disturbed by background.Based on the FIR filter, OSTU method is adopted to solve the problem that traditionaledge detection is liable to be influenced by high brightness area. And finally theproposed algorithm can complete reliably structured road detection of differentshapes.(2) In order to solve the problems that the existing structured lanes detectionalgorithm is susceptible to occlusion and background interference under complicatedroad environment, a novel structured road detection algorithm based on gradient-pairsand parallel perspective model constraint is proposed. Based on further extracting thegeometric features of structural road, the parameters of the parallel perspectiveprojection model and linear-hyperbola model are estimated by using twice Houghtransform. And then the single-lane and multi-lane detection are realized under allkinds of complicated situations.(3) Aiming at the shortcomings that the existing unstructured road detectionalgorithm based on Gabor texture suffers high computation cost and bad real timeperformance, a novel texture extraction method based on Haar features is proposed.By designing the real part and imaginary part of Haar template respectively, thecomplex response of the texture features is obtained. And the texture features ofunstructured road are extracted rapidly by using integral image.(4) Concerning the disadvantage that the finite Haar template can’t extract roadtextures accurately, an orthogonal rectification algorithm of the road texture extraction is proposed. And by using weighted operation of angle and distance information, thevanishing point of unstructured road is detected. Then the unstructured road issegmented and detected accurately by using orientation consistency ratio feature andcolor information.In this paper, the proposed algorithms are tested on a variety of public testlibraries of road detection and are compared with some existing main algorithms.Experimental results show that the proposed methods in this thesis can achievereliable and effective road detection in many complicated situations.