Panorama-based Detection And Recognition Of Road Markings

It is the road, of which the information is increasingly distinct that isdefinitely the skeleton of the city. Whether the assistant driving, GPS basedvehicle navigation or the prospective driverless vehicles, depends on thecognition of the road. In order to extend the existing map and enrich the roadinformation. This paper analyzes various information on the road surface,including the lane markings, traffic signs, vision based ego-motion estimationand so on.To deal with the limited field of view of traditional vision-based systemsand their occlusion problems, this paper designs a novel vision-based sensingsystem, which consists of an around-view system and a panoramic system.The around-view system is composed of four fisheye cameras distributedaround the vehicle, and is used for the current lane sensing. The panoramicsystem is mainly composed of one camera that can take a360panoramicmap, and is used for multi-lane sensing. The two sub-systems, which arecalibrated and fused can complete coordination and informationcomplementation in order to maximize the road information.In order to detect and recognize the current lane, around-view basedmethods are proposed. In this paper, region of interest based lane lineextraction method is used to detect the lane line, which can realize filteringeffectively. In order to analyze the type and color of the traffic lane,histogram statistics of Y direction and color space threshold value divisionare completed for the lane line in the region of interest and the results are sentto the decision-making filter. Experimental results prove the feasibility of theproposed method. In order to detect and recognize multi-lane, panorama based methods areproposed. Based on the present lane width and traffic lane position offered bythe around-view system, the width is assumed to be constant and then pixelsearching and extrapolation are used to do cluster analysis for lane line pixel.At last, conic fitting is used to analyze the data. Experimental results showthat this method can detect multi-lane with high accuracy rate under differentconditions.In order to detect and identify traffic signs on the road surface, slidingwindow based detection method is proposed in this paper. Template matchingmethod and SVM are used to recognize the traffic signs. For SVM, Zernikemoments, wavelet moments and Fourier Descriptors, are compared andeventually the combination of Fourier Descriptors, one with better effect, andtemplate matching method is used to realize mark detection. Experimentsshow that the methods, which relies on the multi-lane detection results, canhandle many kinds of road sign with high accuracy rate.To deal with vehicles disturbance in real traffic, this paper proposesvision based ego-motion estimation to realize vehicle detection andinterference removal. As around-view images contain less dynamicinformation and gray scales, improved ICP algorithm is introduced to achievethe registration of edge point sets so that the ego-motion parameters areconsequently obtained. For panoramic images, optical flow algorithm is used.The results from the around-view system help to filtrate the optical flow andoptimize the ego-motion parameters. Till the end, according to the opticalflow feature the detection of other vehicles on other roads is completed.