Study On Visual Information Processing Technology Of ALV On Structural Road

The real-time processing of vision information is a critical technology in ALV(Autonomous Land Vehicle)guidance and the basis of ALV navigation. Its main task is to recognize and understand the road scene rapidly, avoid obstacle, and obtain the road on which ALV can drive. There are two main contents in the processing, one is to recognize road surface, the other is to recognize obstacle.Most existing algorithms on structural road are only applied a certain circumstance, therefore they are not robust. To satisfy real-time and robust requirement, the processes in this paper are based on the fusion of space and time information.In order to avoid the disturbance from surroundings, landmarks and road edges are first recognized, and then used to restrict road areas, in these areas road surface is extracted by histogram. Most algorithms detect landmarks and road edges under the perfect circumstance, and can not deal with pollution, occlusion and discontinuity. In this paper, landmarks and road edges are firstly detected by fan-projection which is fast and can detect consecutive lines, and secondly compensated based on the principle that their positions have little changes in consecutive frames when they are unable to be detected because of pollution, occlusion and discontinuity.The main problem in road surface segmentation is shadow and water-dirt, which are commonly extracted according to color information because they are of no certain shapes. Most existing methods base on color information for shadow and water-dirt extraction only applied a certain circumstance, so they are very limited. Color comparing system is used in this paper, which takes the color feature of non-shadow and non-water-dirt areas as a reference, therefore it is adaptive to all kinds of road.Obstacles have changeable shapes. Those methods which can recognize shape-changeable obstacles are of high complexity, such as optical flow and stereo vision methods. Though reprojection method is easier than stereo vision method, it is also of high complexity when it comes to process the whole image. This paper firstly regards the foregrounds in road surface recognize result as suspicious obstacles, and secondly uses perspective method to check whether the suspicious obstacles are approaching, and then thirdly uses reprojection method based on feature points to check whether the approaching suspicious obstacles are higher than road plane, and then finally decides which are real obstacles. This method avoids reprojection processing on a large area, and can recognize shape-changeable obstacles, therefore it is robust.Experiments showed that those methods mentioned above can process a single frame within 40 ms, which can satisfy the requirement of real-time in ALV vision information processing.