Design And Research On Binocular Stereo Vision Based3D Mapping System For Micro Aerial Vehicle In Indoor Environment

The binocular stereo vision technology has strong environmental awareness, and it has been widely applied into location estimation, obstacle detection,3D reconstruction and so on both in grounded vehicle and unmanned aerial vehicle. The objection of this article is to design a low-cost and good commonality binocular stereo vision system, and use this system to finish indoor3D reconstruction for unmanned aerial vehicle. Based on this starting point, following work has been done.First of all, on hardware platform, we have developed an embedded binocular stereo vision system with two consumer-grade USB2.0cameras, Beagleboard-xm board and other devices. This system is stable, reliable and good commonality, and can easily transport into outdoor UAV platform. Besides, this article has solved two engineering problems:bandwidth limitation and synchronization of two cameras.Then, in software, the objection is to improve algorithm’s reliability and real-time. After deeply analyze the principle of binocular stereo vision technology, we have finished system calibration, distortion rectifying, stereo rectification, stereo matching,3D coordinate recovery and so on, and get3D point cloud of the environment. For the point cloud is computational complexity and has poor robustness, we propose to develop octree-based3D occupancy map after having got3D point cloud. After deeply analyzing the structure of octree, this article proposes a sampling strategy that adopts a point among every ten points, and uses point cloud to initialize octree-based map.At last, the article has showed the result of3D environmental modeling both in indoor environment and corridor environment. Form the results we have found that the octree-based3D occupancy map is effectively modeling indoor environment, and this map can be used to path planning. Besides, we have analyzed those factors which affect the result of3D reconstruction in detail, and compare octree-based3D modeling and whole grid3D modeling in the perspective of real-time performance.