| Taking a panoramic view of the history of the vehicle development, the study of vehicle safety has always been the top priority. In particular, with the development of economy and the improvement of society recently, problems such as road passing ability, traffic safety, energy consumption and environment pollution have been becoming increasingly severe. The most hazardous issue to the modern society is traffic accidents. Thus how to improve the traffic safety is a social problem that has to be solved urgently. Research and development of intelligent vehicle are the basic approach to reduce traffic accidents, to enhance transportation efficiency and to relieve drivers'burden, which has attracted more and more interest of institute and automobile manufactures and becomes one aspect of key technologies of the Intelligent Transportation System. The major characteristic of intelligent vehicle is able to perceive, recognize, acknowledge and adapt the environment and is an automatic driving and control system integrating with environment perception, planning, making decision, operation and control systems. Although various sensors can be used to perceive the environment for intelligent vehicle, computer vision is proved to be the most effective perceiving manner after many years'research, because it can work with much information, effectively and more cheap.. Therefore the technical route mainly based on computer vision is adopted in the design of most of intelligent vehicles in the world. This paper is a part of the High-speed Vehicle Intelligent Assistant Driving System research subject of the State Key Laboratory of Automobile Dynamic Simulation, Jilin University. In the context of application of computer vision to intelligent vehicle, the lane detection of the structured road, which is the key technologies of visual perception system of intelligent vehicle, is the main task. The aim is to develop a more perfect travelable area detection algorithm which is fit for most of structured road conditions.The work of lane recognition is normally made up of two parts, followed by road binary image generating and lane fitting. The aim of road binary image generating is that the lane image part can be extracted from the image and the interference of the other image information can be excluded with the image processing method. The purpose of lane fitting is that a (group of) curve is designed to cross the image part of lane, and then the lane position can be described in a road image. If necessary, the results of lane fitting can also be projected back to 3D space from 2D plane with the camera model and perspective projection relationship, and the decision model of the intelligent vehicle will be served with the projection results.The purpose of binary road image generating is that through the analysis of the road image in front gathered from the vision system on vehicle, the target image can be extracted and can be used as the import of the decision model and the control model follow-up. The target image here may be the lane in the structural road, the edge of the unstructured road, the other cars or passerby in road and so on with the different needs. The problems faced are that due to the complexity of the road image that there are all kinds of information there in the road image such as the sky, the surroundings and the buildings besides the road and the lane, and the changing of the illumination and the weather, and also the shadow, the smutch, and seeper and firn in road, it is difficult to guarantee that the lane image can be saved accurately in binary road image, so the stability of the algorithm is poor. In the paper the method of adaptive binary road image generating with each window is raised. The method is that: At first, the road image is divided into a number of rectangle windows; and then, the binary image in each window own is generated adaptively. So the influence of the changing of the illumination and the weather can be resisted. And then, to take into account the characteristics of the road image that influence of perspective projection relationship of camera model is great, the means to improve on is raised later. The means is that the work of dividing road image into rectangle windows is finished in 3D space, and then project back to 2D plane with perspective projection relationship. So the perspective characteristic that the window near is big and the window far is small can be materialized with the rectangle windows gathered, and the characteristic is according with the transmutation rule of the main scenery in image, so the probability that the unsatisfactory binary road image will be produced is reduced. At last, the binary image is generated in each rectangle window with OTSU, and the results are fitted together, then the integrated binary image is gathered, which can be the import of the work follow up.In the structural road image, lane exists just in some certain regions. According to the method to generate the binary image in the whole area in before-mentioned content, it is difficult that the lane image can be saved accurately in the binary image. If with the physics and continuity restriction of the image stream, the regions which contain the lane image or neighbor with the lane image can be found, which always be defined as ROI(Region of Interest), the binary image generation work can be done in ROI, so the workload of image processing will be reduced, and the other information in image except lane will be filtered further, and then the real-time, accuracy and anti-jamming capability of the binary image algorithm can be improved thereby. On the basis of the method of adaptive binary road image generating with each window in the last chapter, the two methods are raised in the paper that a group of windows of interest(WOI) are selected which contain the lane image in them, and then the ROI can be generated with serial linking WOIs. The two methods are fit for the two phase of lane recognition respectively. The first method is that the complexity of the image in WOI is scaled with the contrast of the gray image in WOI, and then to forecast there are lane in WOI or not, so the decision that the WOI should be held or not can be done. The result of the binary image will become more adaptive, robust and accurate with the method. This method are fitting for the phase of lane searching, in which the lane recognition work should be done in the whole road image area because there is no lane information in last frame of image, so the accuracy of the image processing algorithm should be guaranteed here. The second method is that: The initial position of the WOIs can be fixed on with the result of lane recognition in the last frame of road image, and then the best position can be gathered with the regular sliding of the WOI. The workload and the difficulty of the road image processing can be both reduced, and at the same time the accuracy of processing result can be improved. Thereby the method is fitting for the phase of lane tracking, which is the main state of the lane recognition system working. And the influence of the lane discontinuous part can be effectively reduced, so the robustness of the initial algorithm of lane recognition can be improved further.Lane is a very important integral component of the traffic management, and plays a positive role in the safe, smooth and order road traffic. Each kind of lane has its own characteristics, and expresses the different signification. It can bring large effect on high-speed vehicle intelligent assistant system to determine the lane style correctly, such as to assist or guide the driver to control vehicle, to alarm or even forbid the driver for possible vehicles traveling on non-performing line caused be driver's action. In the paper a method for lane style recognition is raised, which is working in two ways, color and line type. In the color recognition of lane, there are two kinds of useful color, white and yellow. In the line type recognition of lane, the situation is more complex. Aim to this situation, a method for the recognition of the broken points of broken line in the paper, with which the line type can be recognized for real line or broken line. With this method, not only the lane style can be recognized finally, but also the accuracy of the lane fitting result can be improved combined with the lane fitting method mentioned in next chapter.Lane is the establishment base of the travelable area of intelligent vehicle, and also one of the key problems of the vision system of the high-speed vehicle intelligent assistant driving system. There is a situation that the accuracy of lane fitting results is poor when is described with the simple mathematical lane models, and the real-time request of lane fitting results is difficult to satisfy when is described with the complex mathematical lane models. To resolve the problem above, a new method to describe the lane position with a group of line, named recursive half-dividing broken line fitting method, is raised in the paper. The main idea of the method is that the road image region will be divided into several regions with some beeline paralleled to the image horizon, and then the beeline is used to fit the points belong to lane image in each region, so a group of beelines can be gathered to describe the lane in the whole road image region, finally. With the results of the experiments, it can be seen that the advantage of the method is flexible, which can describe various shapes of lane more accurately, and is less time-consuming, which can basically satisfy the real-time requirement of lane recognition of the intelligent vehicle vision system.Vehicle lane changing running is a special situation which is different from vehicle running in lane, so in the paper the lane tracking in the process of lane changing is made some preliminary exploration, and then a method which can describe the process smoothly is raised. At first, it can be estimated with the vehicle steady preview model that the distance between the vehicle and the lane is safe or not, and then the purpose that the vehicle will change lane or not can be ascertained. When it is ascertained that the vehicle will change lane, the temporary WOIs is introduced to recognize the new lane appeared in the lane changing process, so the normal two-lane tracking state is translated to the three-lane tracking state at the same time. With the method the process can be described smoothly that the new lane in lane changing process becomes from inexistence to appearing, and one of the original two lanes becomes from existing to vanishing, so the lane tracking process in lane changing can proceeding smoothly, too. At the same time, it can be switched flexibly between the three-lane tracking state and the two-lane tracking state with the lane changing state estimating.At the end of the paper, many kinds of typical road experiments with real car for all the algorithms mentioned above are done to prove their effect. In the phase of lane searching, the contrast-regional homogeneity and OTSU image segmentation algorithm is used to process the road image, and the LMedSquare algorithm is used to fit the lane curve. In the phase of lane tracking, the WOI-sliding and regional OTSU image segmentation algorithm is used to process the road image, and the recursive half-dividing and broken line fitting algorithm is used to fit the lane. In the phase of lane changing, the WOI-locating and regional OTSU image segmentation algorithm is used to process the road image, and the beeline fitting algorithm is used to fit the lane. It is shown with the experiment results that all of the algorithms raised in the paper can pertinently resolve some practical issues, and can achieve good effects on practical applications.The creative points of the paper is that a variety of lane recognition conditions of intelligent vehicle has been studied systematically, and the technology roadmap of lane recognition has been analyzed, then a new method of lane recognition is explored which is adaptive, and can satisfy the request of real-time and accuracy. In the process of lane recognition, several new methods are proposed to the problems in the binary road image generation phase, the lane ROI in road image establishment phase, the lane style recognition phase, the lane fitting phase and the lane tracking phase in lane changing process. At last, the road experiments with real car is done, and the effect of all the methods above are validated according to the experiments results.
|
PDF Full Text Request