Research On Key Issues Of Monocular Vision Based MAV State Estimation

Vision plays an important role for the creature to understand the environment.With the development of the computer vision technique,vision based perception is becoming more and more popular in the application of the autonomous operation system such as unmanned aerial vehicle(UAV)to realize the autonomous perception of the surrounding environment.One of the important application goal of the vision-based perception is to realize the state estimation of the UAV,including ego-motion estimation of UAV and 3D reconstruction of the environment.To apply the vision perception technique on the micro aerial vehicle(MAV),researchers has put significant attention on the monocular vision based technique in recent years because of the small size of the monocular vision sensor.However,due to the lack of the distinct baseline,monocular vision based state estimation cannot directly get the metric depth information of the environment,which makes monocular vision based state estimation a particular challenging technique in actual application.Though in recent years some research on monocular vision based state estimation methods have show great performance,there are still some issues in this challenging research focus,which limits the further application of the monocular vision based state estimation.In this thesis,we did investigation and research on the optimization based sliding windowed monocular visual-inertial navigation system(mono-VINS),and put forward some improvement and innovative methods on the topic of state estimation robustness and observability analysis,towards improving the performance of state estimation formulation and robustness of its application.For improving the robustness of the state estimation formulation,we did research on three key issues of the mono-VINS,analyze the drawback of the existed methods and put forward improvements.Firstly,for the rotation estimation in the linear initialization,we propose an optimization method of rotation matrix to improve the linear initialization accuracy.The proposed method integrate the rotation information in image frames and the IMU measurements,formulate a linear equation constrained by the epipolar geometry and system dynamics to yield an optimized rotation estimation.Secondly,for informative selection of key frames used in the sliding window,we propose a multi-level threshold method to determine the key frames.Three inspects are taken into account by the multi-level threshold,including the time interval,translational parallax and overlapped feature number,which ensures full-rank condition for the linear initialization process and informative and tight connection between the graph nodes for fast and precise convergence in the nonlinear optimization process.Thirdly,we designed a precise synchronization method to realize the data synchronization from the camera and IMU,based on sensor triggering by hardware and data acquisition by software.Based on the above improvements,we implemented a complete monocular visual-inertial state estimation system to realize the ego-motion estimation and 3D reconstruction of the environment.The performance of the estimator is demonstrated by simulation and experimental evaluation.For maintaining the observability of the state estimation system,we proposed a series of methods to realize the online observability quantification,degradation direction prediction,local observability prediction and active motion direction suggestion,based on the EOG and motion primitive techniques.A EOG-EVD based observability analysis method is proposed to realize the degradation detection and direction prediction.An EOG-trace based system observability quantification method is proposed for fast observability evaluation,and motion primitive technique is incorporated to realize the local observability prediction to avoiding the potential observability-deficient perceptual environment.A local trajectory sampling based observability constrained motion direction suggesting strategy is proposed for informative motion guidance,by which the trajectory is well selected to ensure the system observability and state estimation performance.Based on the self-implemented monocular vision-based state estimator,this thesis puts forward some innovation points through investigation and research on the topics of state estimation robustness and system observability analysis,which greatly improves the state estimation performance and is of great significance for the promotion of its practical application.