Research And Realization Of Vision-based Localization Algorithm For Unmanned Rotorcraft

In those years, UR (Unmanned Rotorcraft) has broad prospects for developmentin both military and civilian field, including reconnaissance and combat in militaryuse, search and rescue missions in dangerous environment (disasters such as fire orearthquake site) and agriculture in civilian use. Usually, unmanned rotorcraft use GPS（Global Positioning System）as its navigation mode, but sometimes GPS signal canbe disturbed or invalid, which causes failure in UR navigation. Besides, positioningaccuracy of GPS is about10meters, which is not satisfied in some occasions.Considering the high price of high accuracy GPS, this issue researched vision-basedlocalization algorithm for UR in flight, which used onboard HD camera and visionalgorithm instead of GPS to calculate the3D coordinates of UR to provide for itsposition control. In order to minimize localization time and increase its accuracy, animproved algorithm was proposed and was tested on a real testbed.Firstly, the vision-based localization algorithm was proposed. From thebeginning of its takeoff, UR used onboard HD camera to capture images of groundtargets in locus. At the same time, we used SIFT algorithm to extract feature points ofthe two successive images. Then KD-tree feature points searching algorithm andEuclidean distance were used to match the extracted feature points to calculate theirpixel locations in images. Our algorithm is an iterative algorithm. We should knowUR’s real-time positions of the first two moments when it took off. Then we can usecoordination transform algorithm Trf1to calculate the ground positions of thematched points in targets. Some of which were used to calculate UR’s next momentXYZ position with the help of coordination transform algorithm Trf2. So, thecalculated UR position was used in a new iteration. Here, the two coordinationtransformation algorithms and were proposed based on transformationsof UR body coordination system, ground coordination system, camera coordinationsystem and image coordination system. They should use UR’s yaw, pitch, and rollangles which were measured by IMU (Inertial Measurement Unit) and pixel locationsof the matched points mentioned above.Secondly, in order to reduce localization time and increase localization accuracy,we improved the localization algorithm mentioned above. The algorithm containsthree parts: image acquisition, image registration and position calculation. We improved the image registration part. First, dimension of SIFT algorithm featurepoints descriptor was reduced from128to64, then K-means clustering algorithm wasused to decrease the amount of feature points. So, only the feature points in andaround the targets were left, because they were what we need. Then we use KD-treethe nearest neighbor query improved algorithm BBF searching algorithm to match theselected feature points to further reduce the matching time.Finally, we designed experiment to realize the localization algorithm. We builtthe hardware platform which we used to test the algorithm. Experiment showed that,time of the algorithm reduced about500ms, the calculated UR’s XYZ positions’errors kept within6cm, the relative errors were around6%, which satisfied thelocalization demand.