Study On UAV Navigation Algorithm With Onboard Stereo Camera

Autonomous navigation is a very popular research field in UAV research.At present,most of the unmanned aerial vehicle navigation technologies in the world are based GPS-based positioning systems.Although there are few ways to use visually aided navigation,most of the visual positioning is based on the optical flow method,where only the relative speed can be estimated for hovering.When the UAV is indoors or the GPS signal is invalid or poor,it is challenging to achieve autonomous navigation of the UAV.To address the above problems,this dissertation presents a UAV visual navigation algorithm based on airborne cameras to realize UAV autonomous navigation in unknown environments without relying on any external sensor.Autonomous navigation needs to address three basis problems,i.e.,UAV should know their own location information in the world,have a certain understanding of the surrounding environment to know the location of obstacles,and be able to plan a feasible path to reach a target point.Correspondingly,we have developed three modules,including state estimator to achieve visual positioning,environmental perception to establish 3D dense map,and planner to plan a path to a target with obstacle avoidance.First,the state estimation of UAV is obtained by fusing the information of vision images and inertial measurement unit with extended Kalman Filter.The method is designed based on the technologies of state quantity of the current airborne coordinate system,the inverse depth of the feature points,the representation of the feature points on the Lie algebraic space,the alignment of the pyramid of the multi-layer image blocks,and the Bundle-Adjustment optimization.The proposed state estimator does not require any special initialization process,and it is of low computational complexity and good accuracy.Second,a real-time binocular stereo matching algorithm based on the Delaunay triangulation method is proposed for the perception of the surrounding environment.Combined with the aforementioned pose estimation information,a 3D dense map of the environment is established.Because many points in the binocular image are markedly easy to get the correct matches during the binocular vision matching process,the Delaunay triangulation is performed by these correct matching points,for those points that have not been matched yet.With the triangulation,the depth estimation can be obtained and thus narrow the scope of parallax searching,reducing the amount of computation and increasing the matching accuracy.Based on the built dense map and state estimation of the aircraft,a path planner is further developed with real-time performance.By considering the limited computational resources and the large number of system stat e space dimensions,a heuristic fast exploration random tree(RRT)algorithm is proposed.Its sampling range is heuristic,and the algorithm performs heuristic sampling based on an initial solution.The final path is obtained by iteratively optimization.T he algorithm needs less time and computational resources to get the global optimal solution compared to the original RRT algorithm.Finally,in order to verify the proposed scheme,a six-axis unmanned aerial vehicle for experimental verification is built.The actual data from the motion capture system are used to illustrate the effectiveness of the proposed algorithm.The obstacles in the environment appears accurately in the 3d dense map which is constructed in real time.The dissertation also demonstrates the feasibility of the path planning algorithm for different starting points and target points.The proposed approaches will play an important role in the autonomous navigation applications of robots under the GPS-denied situations.