Research On Vision Based Road Recognition And Obstacle Detection For Intelligent Vehicles

With the development of world economy and the rapid progress of urbanization,vehicle has brought great convenience and become an indispensable transportation tool to our life and production activities.Because intelligent vehicle has the great significance of reducing traffic accidents and boosting transportation capability,extensive market prospect,and can lead the development of automobile industry in the future,it has received widespread attention.Due to excellent sampling speed and detection accuracy,lidar sensor plays a key role in current intelligent vehicle systems,and becomes an indispensable part for high accuracy positioning.However,to some extent,price is a major factor hindering its marketization.Since camera sensor is relatively fast on sampling,lightweight,power efficient and inexpensive compared to lidar scanner,it is now the research focus of many domestic and overseas institutes.Current vision based intelligent vehicle environmental perception technologies are easily affected by lighting,weather and road type changes,causing reduced detection accuracy and realtime performance.Therefore,this thesis makes the environmental perception for intelligent vehicles the target of study,and focuses on the aspects of road recognition and obstacle detection on the basis of existing research.The main research contents of this thesis are:(1)Research on the technique of detecting vanishing point(VP)from road images.As a basic characteristic of image,since VP can be used for steering control for straight and curved roads,and has outstanding performance under various kinds of road conditions,it has received extensive attention.However,existing VP detection approaches have the problem of high computational complexities in the texture extraction and voting processes,and are easily affected by the local strong texture features that are orientationally inconsistent with the road.Considering the above issues,an algorithm of detecting VP from road images is proposed,which is based on the joint activities of four Gabor filters and particle filtering technique.First,the algorithm exploits joint activities of four Gabor filters to boost the efficiency of texture computation.Meanwhile,in order to exclude the interference of weak texture pixels,confidence measure is integrated into the texture computation process for finding reliable pixels.Second,the algorithm utilizes sparse voting strategy and particle filter based tracking technique to improve the anti-interference ability and lower the computational complexity of the vote function.Finally,to improve the stability of VP detection,the distribution range of VP candidates is regulated according to the peakedness measure of the accumulation space of the vote function and the moving displacement of VPs between frames,which also makes the algorithm can recover from a wrong estimation.Through comparison with the most representative methods in the field,the experimental results demonstrate that the proposed algorithm has higher accuracy and efficiency,and stronger anti-interference capability.(2)Research on the problem of road image segmentation.In light of existing VP constraint road image segmentation approaches overemphasize boundary characteristics of the path,some non-road areas are easily misclassified as the path.To overcome this problem,a novel segmentation approach is proposed by fusing texture,road and non-road color features.The algorithm converts road segmentation into Bayesian posteriori estimation problem based on a straight road model with VP as a constraint.It utilizes the ratio of orientation consistency to describe the road texture feature,and nonlinear transform functions as well as self-supervised strategy to compute the similarity of image pixels to the “road” pixels to highlight the road area in the image,which is then used as the probability measurement model of the road and non-road color features.The algorithm takes the advantages of the texture,road and non-road color features,and segments the road surface through maximizing the Bayesian posteriori probability density estimation.Through comparison with the most representative road segmentation algorithms in unsupervised or self-supervised manner,the experimental results indicate that the proposed method has higher segmentation accuracy.(3)Research on the techniques of road scene 3D reconstruction and obstacle detection based on stereo vision.Illumination variation,matching ambiguity in disparity discontinuity,occlusion or weak texture regions,as well as real-time and resource consumption etc.are the key problems to existing stereo vision based obstacle detection algorithms for intelligent vehicles.Although there are plenty of approaches available for a specific issue,till now no a solution can tackle all the above issues comprehensively while guarantee the accuracy and real-time performance of the system.Therefore,this thesis utilizes: relative gradient census transform to overcome illumination variations;strategy of sparse matching cost aggregation,predictive coding and disparity up-sampling techniques to lower down the temporal and spatial complexities of the algorithm;and the technique of smoothness constraint as well as disparity refinement algorithm to eliminate the problem of matching ambiguity in weak texture regions.Through comparison with the most recent and representative local stereo matching algorithms,the experimental results demonstrate that the proposed algorithm has the merits of insensitive to illumination variations,high computational efficiency and accuracy,low memory cost.In addition,the algorithm uses reference grid approach to extract obstacles from the disparity map,experimental results indicate that it can be used for structured and unstructured roads,and has the ability of detecting positive,negative as well as overhanging obstacles.(4)Research on application of vision navigation algorithms in trajectory tracking control of intelligent vehicle.The problem of trajectory tracking and control of intelligent vehicle is investigated,which illustrates an application of the proposed three environment perception algorithms in intelligent vehicle systems,and analyzes how to combine with an automatic control algorithm to realize autonomous driving performance,such that the research contents of this dissertation can be systematically validated.The algorithm utilizes: vanishing point detection,road segmentation and obstacle detection algorithms to find the drivable orientation and obstacle-free regions;A* algorithm and cubic spline curve function for planning a smooth path;and sliding mode control algorithm with three hierarchical sliding surfaces to transform the visual navigation information into control input signal that can drive the vehicle moving along the reference trajectory.Experimental results verify the effectiveness of the proposed environment perception algorithms,and demonstrate that the presented control algorithm has the following merits: invariant to matched and mismatched disturbances;without obvious chatter phenomenon in the control input signal.Therefore,if being applied in an intelligent vehicle system,it has the effects of reducing heat loss of electrical components and wear of executing parts,and boosting trajectory tracking accuracy.