Road Recognition And Moving Object Tracking In Infrared Image

With the fast development of computer and sensor technology, and the urgent demand from variant realms, more and more interests are focused on research of Autonomous Land Vehicle (ALV), especially the two subsystems: all-weather navigation and environment surveillance. Because the common monochrome or color camera cannot work well at night or in ill-weather condition, more and more attentions are attracted by infrared (IR) camera, which is not influenced by illumination. Thereupon, the research of road following and object detection based on infrared image becomes an important topic.In this dissertation, several key techniques in road recognition and moving object detection based on IR image are discussed.The edges of unstructured road are usually blurred in IR image, which makes the normal methods based on gradient disabled. A new road segmentation method based on watershed transformation was then presented. Firstly, the new watershed transformation using extended chain code was introduced. Then a new region merging method for solving over-segmentation was discussed. Experiments show the new algorithm can segment the road regions correctly.The structured road edges can be represented by straight line or segmented lines. So the lane detection method based on segmented line model was proposed. Firstly, the new chain code based line detect algorithm was introduced. Then combined with the rule of relation stable-state, the contour-sum image could be created. With the concept of multi-scale morphology gradient, another mask image was made. Finally, by validating the two mask images derived from different methods, the description of lane could be produced.For detecting moving objects in IR image, this paper presented some algorithms. For instance, this paper proposed a new method for generating background reference images using multiple stable/unstable Gauss distributions. As the moving objects are usually too small and have no distinct difference from background, they could hardly be detected from one single frame. So the strategy of track-before-detect was introduced to detect moving objects. In outdoor surveillance, the merging and splitting among multiple objects could not be avoided. A new method based on object similarity was presented to solve such problem, which made the result more useful.In the end, the two whole subsystems, road recognition and environment surveillance...