| With the rapid development of unmanned aerial vehicle(UAV)industry in recent years,civilian UAV platform is more and more toward miniaturization and intelligent direction.Intelligent development improves the computing power of UAV platform and provides a foundation for carrying advanced algorithms,miniaturization further expands the use space of civilian UAV,a growing number of small drones are launched in indoor environments that account for 70 percent of human activity.Based on indoor positioning and environmental perception tasks of UAV,this thesis presents feasible methods of positioning and map building of UAV in indoor environment based on RGB-D camera from three aspects:visual odometer between two frames,camera pose estimation optimization between multiple frames,and scene octree map construction.In view of the light reflection in indoor environment,the scene condition where the assumption of gray scale invariance is not valid,and the high real-time requirements of UAV platform positioning,a homogenized ORB feature extraction method based on strong real-time capability ORB feature is proposed.Homogenized ORB feature extraction ensures real-time performance and avoids the aggregation of feature points,on the basis of linear estimation of camera pose change,supplement an algorithm to cope with pose estimation when the depth information of feature points is lost.Finally,a two-frame visual odometer with good robustness is constructed.In terms of reducing the error accumulation of pose estimation,the limitations of traditional filter methods are analyzed,an optimization algorithm based on bundle adjustment is adopted.Through scale control of optimization,enables large-scale global optimization apply to UAV platforms with limited computing power.In addition,DBoW3 library was used to construct the scene word bag model to complete the closed-loop detection of the trajectory formed by the change of camera pose.Experimental results show that the accumulated errors can be corrected in time and the camera pose estimation is reliable and globally consistent.Finally,the advantage of real time depth measurement of RGB-D camera is fully utilized,based on the accurate estimation of camera pose,the dense point cloud reconstruction of indoor scene is completed with little computational cost.The outer point removal and sampling reduction were carried out to eliminate the redundant information.Finally,the reduced dense point cloud of the scene is transformed into an octree map that can be used as navigation.The experiment shows that the occupied space of the map is greatly reduced after the transformation,and the occupied map that can be applied to the three-dimensional space path planning is constructed at the same time,indicating that the method proposed in this thesis can well meet the requirements of UAV positioning and map construction in the indoor environment. |
