Research On Monocular Visual SLAM Technology For UAV Autonomous Navigation

Unmanned aerial vehicle(UAV)autonomous navigation is an inevitable trend of development.Autonomous navigation includes three steps: localization,mapping,and path planning.Simultaneous localization and mapping(SLAM)technology solves the first two problems.Vision sensors have the advantages of low cost and miniaturization.Therefore,visual SLAM is regarded as the key technology for UAV autonomous navigation in an unknown environment.In this paper,we choose a monocular camera with a wider range of scenes as the front-end sensor.UAV visual SLAM are analyzed in terms of real-time,robustness and positioning accuracy.Under the existing theoretical framework,the work is divided into two parts.First,a monocular visual odometry(VO)based on the semi-direct method is studied.In view of the problems in semi-direct motion estimation,two improvements are proposed in the local map tracking step.Firstly,by incorporating the more robust ORB point feature matching on the local area of the projection point,the fact that motion estimation tends to fall into local optimum is improved;Secondly,a density-based tracking strategy is proposed to improve the disadvantages of local map tracking which is insufficiently robust in scenes with concentrated features.The experimental results on the TUM dataset show that the correct matching rate of the ORB point feature is 98.6%,and the average pixel error can be optimized by 5.20.The experimental results on the ICL-NUIM dataset show that the density tracking method has better robustness than uniform tracking strategy.Then,keyframe-based mapping and back-end graph optimization is studied.The key frame strategy based on time sampling or spatial transformation is not suitable for the flight modes of hover,pure rotation and forward motion.To overcome this,we take the visual changes of the scene into account and propose a key frame strategy combining visual and spatial transformation,which can save the scene information more completely.In terms of feature extraction,this paper uses the ORB point feature to replace the original FAST corner point of the semi-direct method.To solve the problem of insufficient number of ORB point feature extraction in low-texture scenes,gradient points and line features are increased for improvement,and the way of gradient point features participating in motion estimation is analyzed.For the line features,the paper establishes the connection between the two points to complete the representation of the line features by extracting the features of the gradient points at both ends of the line,which eliminates the parameterization of the spatial line,and then participates in the motion estimation as a point feature.The advantage of this method is that even if the line is detected as pseudo or occluded,the gradient points at both ends are independent.Aiming at the problem of cumulative error of front-end VO,the back-end graph optimization strategy is proposed in this paper.The experimental results show that the density tracking method combined with the gradient features has best robustness.And the map reconstruction based on key frames can accurately represent the scene details.The line features enrich the structure information of the scene.The reconstruction of the map based on points and lines is more intuitive and facilitates further obstacle avoidance and path planning.Finally,the SLAM system of this paper is constructed and a comprehensive evaluation experiment was carried out.In terms of real-time,our SLAM can achieve a frame rate of 75 fps.In terms of positioning accuracy,the experimental results on the TUM dataset show that the VO without backend optimization in this paper is 80.95% and 69.01% higher than that of SVO and PLSVO,respectively.Our SLAM is increased by 56.11% compared with LSD-SLAM,and is more robust than PTAM and LSD-SLAM.ORB-SLAM is 51.07% higher than our SLAM.The experimental results on the UAV dataset EUROC show that our SLAM is 65.35% higher than our VO.Our VO is 38.72% higher than SVO(edgelets)in parts of the results,and is 52.55% higher than LSD-SLAM without closed loop.