Intelligent Mobile Robot Visual Navigation Technology

In recent years, more and more interests are focused on the researcl1 of vision-based mobile robot,which can be widely used in both military and civilian appIications. They can accomplish somedangerous military missions such as batt1efield recon and sweeping mines. And they can be sent toaccomplish science exp1orations in dangerous or distant places, such as other planets where it isvery hard for human beings to arrive and survive. And the technologies applied in these robots canalso be applied in many otl1er cases, such as Computer-Aid-Driving, intelligent vehicles andintelligent wheel chairs. The mobile robots can be classified as indoor and outdoor robots. Andthey are very different with each other because their working environments are very different.Indoor environment often has constant lighting conditions and simple scenes and the range is oftenvery s1nall. On the other hand, outdoor environment often has varying lighting conditions andmuch more complicated scenes and the range is also very large. So the outdoor robots need to besmarter and more robust tl1an the indoor robots.ln this paper, effOrts are focused on the researches of some pivotal technologies of the vision-based navigation of outdoor mobile robots such as: road scene understanding, stereo-vision forobstacle detection, and tl1e generation and fusion of local map. The navigation modu1e of mobilerobots has tWo main tasks: to keep the robot running on the road and avoiding the obstac1es infront of the robots. So the vision module needs to recognize what is the road and to detect whetherthere are obstacIes in front of the robot and where they are. The algorithm should keep highveracity in different cases. And the processing speed shouId be fast enough to direct the robot intime. So the robustness and pTOcessing speed are two main criterions of a vision-based navigationa l gorithm.ln this paper, a real-time knowledge-based road scene understanding algorith1n is presented at first.Considering the co111plexity of the road e11viron1nent, it is very difficult to develop a versatilea1gorithm that can l1andle every case. So our a1goritl1m used severaI special-designed filters toprocess some specia1 cases such as tl1e shadou', strong light and waters. And then a fusion n1odu1ebased On tl1e evidence reasoning tl1eo1y was app1ied to fuse the results of these filters. Althoug11this fusion result can give a coarse region of the road, it is not accurate el1ough because it usedlocal infon11ation only. In order to obtain a more accurate result, the fusion result was then appliedto direct the road edge detection module to detect tl1e edges of the road. And the road shapeknow1edge is then used to organize tl1ese road edges in Order to obtain a fine region of the road.Using Intel's MMX technology with the aids of a DSP-based processing board, this aIgorithm canbe realized in real-time.In order to measure the height and distance of the obstacles in front of the robot, stereo vision iswidely used by many research groups. But there are sti1l some problems' the first is the largecomputation, which makes that the reaI-time realization is very difficult; the second is thematching, and a match error will cause the error in tl1e measure of distance. AIthough CMU'sexpenments proved that multi-baseline stereo can reduce the error rate in matching, but oneadditional camera means much more information need to be processed and more conlputationresources wi11 be occupied. So in our experiments, we use three cameras: one color camera andtWo B/W cameras. And the difference within different cameras shouId be considered whenmatching.At last the local map generation is also presented in this paper ln order to give the local-path-pIanning moduIe a good result. the result of the road scene understanding and obstacle detectionare fused together And then the previous results are also fused with current result. And the methodpresented can be also applied in othe...