Research On Vision Based Environmental Perception Technology For Intelligent Vehicles

Intelligent Vehicle (IV) is an important constituent of the Intelligent Transportation System (ITS).The purpose of IV technology is to reduce the growing incidence of traffic accidents and to improve both the road traffic safety and the transport efficiency. To some extent, it can alleviate the energy consumption and environmental pollution. As a result, the IV technologies have been attracting worldwide attention of scholars increasingly.This paper conducts in-depth discussions on the vision based lane detection, vehicle detection and tracking, pedestrian detection in different road scenes for environmental perception of intelligent vehicles. Concrete content of the research has been summarized as follows:(1) Based on adaptive regularization architecture, an improved weighted iterative restoration algorithm is presented. It can select the regularization parameter and modify it automatically. In each iteration step, the algorithm updates the parameter and restores a part of the degraded image synchronously and finally gets the optimal image. Experimental results show that the algorithm is robustness under significant noise. A novel approach of the image de-noising using information measure and SVM is proposed. It can improve image quality by recovery corrupted pixels only, and keep good pixels unchanged. The proposed algorithm can achieve better performance both on de-noising and preserve more image detail.(2) A real-time structural road oriented approach in monocular camera based lane marking detection is presented. Firstly, Canny detection approach is used to obtain the edge map from a given road image. Secondly, a searching method based on orientation-priority is proposed, which reinforces those potential road lines while degrading otherwise edge features. Thirdly, Hough transform is employed to compute the linearity degree of every edge segment and filter the edges of intricate texture. Finally, the lane markings are identified by the pixel intensity of the image and its transform. Experiment results show that the proposed approach can achieve robust and effective localization of lane markings. The approach can run at an average speed of 13 frames per second on a P III 933 MHz CPU, and can meet the requirements of safety and real-time of vehicle driving.(3) A monocular camera vehicle detection and tracking approach which by fuse multi-cues is proposed. First, the horizontal symmetry of vehicle rear view is utilized to achieve the region of interest (ROI) so as to reduce search area of following process. Then, the sign of underneath shadow is employed to generate hypothetical positions on which potential vehicles maybe present. Following, both image intensity and figure information are combined to use to verify the vertical symmetry of the potential vehicle candidates. Meanwhile, Mean Shift procedure, based on the object feature model of combining color histogram and orientation histogram, is used to search the potential objects between two sequential image frames fast. More important, both detection and tracking work together under an interactive mechanism which can dramatically improve both detection efficiency and real-time. It shows that the proposed approach can achieve 96% true detection rate and run about average 24 frames per second, which validate the security and real-time requirements.(4) A novel pedestrian detection approach is presented. Firstly, Haar wavelet is employed to transform the input image into its sub patterns of wavelet region with different resolutions. And then, some relevant wavelet sub patterns are selected to compute the wavelet fractal signature in different scales. Next, this wavelet fractal signature is assembled to be a Wavelet Fractal Signature (WFS) vector, which is utilized to training Support Vector Machine classifier. To validate this approach, some experiments based on Daimler's Pedestrian Detection Benchmark are conducted; the experimental results show that the proposed approach has the advantages of compact feature expression form and higher detection rate than available approach.(5) An intelligent vehicle driving assistant device for the purpose of real world experimentation is developing. The hardware is designed including video acquire, image proceeding and display. Based on OpenCV (Open source Computer Vision platform), we designed a simple and friendly man-machine interface. The proposed algorithms have been realized by C/C++, and providing both lane departures reminds and prevents collision warning function. A test in practice has been conducted using this device, and the result validates the approaches proposed by this paper.In summary, this paper focus on the image de-noising, lane detection, obstacle detection and tracking, and based on the self-developed principle prototype of the driving assistance system, conducts a real vehicle tests of lane detection and vehicle detection and tracking, the results contribute practical and heuristic significance to both theory and engineering application in related fields.