Research Of Vehicle Detection And Motion Estimation Methods Based On Stereo Vision

Globally, automobile industry is experiencing the intelligent technological revolution. The study of safety drive assist system and unmanned vehicle technology is aiming at the perception of vehicle collision in advance to reduce the driving accident and to ensure the safety of drivers and pedestrians. In fact, enabling vehicles to cognize the outside environment is the key to improve the intelligent level of Intelligent Vehicle and Advanced Driver Assistance Systems (ADAS). Thus, in this paper, based on the stereo vision technology, an on-road vehicle detection method utilizing point clouds segmentation and Mean Shift Clustering is proposed to realize the vehicle environmental perception. In addition, the motion estimation method for the recognized vehicle is also studied.Firstly, the camera calibration and stereo image rectification technology is studied, based on the configuration of the hardware and software for the stereo vision system. The semi-global stereo matching method is used to compute the disparity map, and edge-preserving image smoothing method is utilized to filter the disparity guided by the original left image. These disparity data could provide reliable 3D reconstruction data for the implementation of ground filtering and vehicle detection.Secondly, it is found that the ground plane cannot be accurately depicted by the V-disparity line model, while applying the traditional V-disparity ground detection method to a non-horizontal camera. Thus, a new method utilizing path features is proposed based on the path boundary features extracted from the disparity map, the least square method is used to fit the ground plane by considering the horizontal tilt, and the ground detection accuracy is improved. The experiment is also implemented to verify the effectiveness and the application scope of the two methods.Thirdly, a vehicle detection method based on the point clouds segmentation and the 3D reconstruction data from disparity map is proposed. However, in the experiment, this method is found tending to miss or duplicate the object detection when the segmentation step size is not reasonable. Therefore, Mean Shift clustering detection method is further developed, which utilizing the Mean Shift clustering to replace the fixed step depth image segmentation. The improved method do not rely on the reasonable step settings and got a better detection result.At last, the moving target candidate regions and the static background regions are determined by the vehicle detection results. The ego and target vehicle motion estimation method is the proposed based on RANSAC to remove the outliers and the unit quaternion method to directly solve the coordinate system transformation problem, and the vehicle motion estimate experiment and the error comparison analysis are also implemented.