Research On The Key Technology Of Vision-based Navigation For Outdoor Mobile Robots

Outdoor Mobile Robot (OMR) is an intelligent mobile robot, which can run autonomously and continuously on road or crosscountry on field. Its research involves theories and technologies of many disciplines, and embodies the latest achievements of information and machine intelligence. Of all the OMR key technologies, vision-based navigation is a difficult and comprehensive subject that is involved in almost every aspect of computer vision researches. The fundamental tasks of visual-based navigation are identifying the road region and the non-road region correctly according to the image information and then planning an appropriate path to the local target. Due to the complexity and inconstancy of outdoor environment and huge information of colorful image, it is difficult to satisfy the real-time requirement and robustness for vision-based system. In order to overcome these shortages, three pivotal technologies, such as road identifying, stereo matching for obstacle detection, and obstacle-avoiding path planning, are researched thoroughly and some novel and effective algorithms are proposed in this dissertation.Firstly, a real-time and robust road-identifying algorithm is presented, which is based on normalized RGB value. Considering the complexity of the road environment, it is very difficult to develop a versatile algorithm that can handle every case. So the algorithm aims at those roads with neuter gray color and uses an active window in the road image to find the intensity and normalized RGB value of the road model. Then the current region of the robot can be confirmed by the characteristics of the normalized RGB value. With the useful information of the current region, the real road region can be identified according to the region growth. Otherwise, obstacles in the road region can be detected by finding the pixels of road region at the same row. Experiment results show that the method succeeds in detecting shadows and is robust to those sceneries with neuter gray color.In order to make the vision-based navigation robust to more scenery, another road-identifying algorithm is presented, which is based on image division and the characteristic of image color. According to the analyses of image color under sun and shadow, some intrinsic rules are elicited to make vision-based navigation robust to all kinds of sceneries. Experiment results show that the process of image division not only can accelerate the performance of road identifying by avoiding analyzing each pixel of the image, but also can resist the effects of small watermarks and dust region.Secondly, area-based stereo matching algorithms are studied thoroughly to improve their real-time requirement and robustness. With qualitative and quantitative analyses of different matching criteria and correlate window sizes, the principles of choosing the two pivotal factors for vision-based navigation are presented. Furthermore, four accelerating techniques are presented. The first is using the road's flatness constraint to reduce the dubious obstacles region. The second is using efficient incremental computation technique to avoid redundant calculation. The third is using the constraints of the epipolar line and obstacle region to reduce the search area of the disparity value. The last is using the constraints of the uniqueness, distinctiveness and sharpness to eliminate the efforts of repetitive patterns and less texture region.Finally, three novel obstacle-avoiding algorithms are presented to overcome the drawbacks of traditional obstacle-avoiding path planning algorithms, i.e. poor adaptability to complex environment where stationary and moving obstacles exist. One is based on improved potential field, which uses collision time instead of relative position and introduces the constraint of collision angle, inherits the advantages of general artificial potential field and mends its shortcomings in dynamic environment. Additionally, a component force is used to escape free-path local minimum due to the repulsive forces generated by the obstacles nearby the target, and a local fluctuant function is used to avoid the local minimum problem. The second including three stages is based on obstacle's encasing box and relative velocity. In the first step, the obstacle's encasing box is presented to achieve the planar shape of the obstacle. In the second step, geometrical knowledge is used to analyze the orientation of the collision and relative velocity between the robot and obstacles. In the third step, two principles are used to realize obstacle avoidance. The third is based on risk-degree of collision and fuzzy logic, where a MIMO fuzzy controller for obstacle avoidance is designed. The risk-degree and state of collision are regarded as input values, acceleration and rotation angle as output value of the fuzzy controller. Due to the utilization of distance and relative velocity between the robot and obstacles, the risk-degree of collision is more efficient than other known model. The state of collision also makes the avoidance decision reasonable.