Research On Stereo Vision And Target Detection Technology For Robot Navigation

This paper concentrates on the research of stereo vision system for Robot navigation,including camera calibration,stereo rectification,stereo matching,as well as road segmentation and vehicle detection based on stereo vision.We compared several current popular methods on camera calibration,distortion rectification and compact stereo rectification in detail,and an experiment was conducted following Zhang's method.Stereo matching does take a big place in the area of computer vision.Recently the adaptive local support weight method developed from the global method achieves a state-of-art performance.However,One of major shortcomings of these algorithms is that they are computational expensive and the complexity increases quadratically as the window size increases.For that reason,we studied a stereo matching method based on Bi-Exponential Edge-Preserving Smoother(BEEPS),which is non-local and has a constant time consumption independent of input data,filter parameters or even the degree of smoothing.Compared to other adaptive weight methods,the studied method is more efficiency at a similar precision.A stereo vision based road scene segmentation and vehicle detection method was also studied in this thesis.In this method,we first did stereo matching by dynamic programming,and removed the mismatched pixels via left-right correspondence.Secondly,we built the V-disparity map,and a projection based method was used to detect lines in it.Then,the road surface and the horizon were extracted,and the intersection points between the road and the vehicles could be computed after that.Ultimately,a U-disparity map was set up to detect the horizontal positions of the vehicles.The experiment result was satisfying,and it validated the studied method on road segmentation and vehicle detection.