Real-time High Precision Localization Method Research With Monocular Vision

The power of processing unit has gained rapid growth in recent years.Simultaneous Localization and trajectory estimation based on visual sensor(Visual Simultaneous Localization and Mapping,Visual SLAM,Visual Odometry)has become one of the core technologies in artificial intelligence and automatic mobile robots.Since the cost of visual sensors,power consumption,volume has been significantly cut-down,the visual SLAM has been widely applied to human society and has attracted a lot of attention in research fields.However,many challenges remain for a realtime monocular visual SLAM like fast movement,low trackable information,uncorrect initial information.Therefore,the main purpose of this thesis is to improve the precision of localization without the scale information in monocular visual odometry,thus can also enhance the precision of localization with scale information in monocular visual-inertial SLAM in real-time constraints.To improve the precision of unscale localization by monocular visual odometry,this thesis proposes a monocular parallax visual odometry.Monocular parallax visual odometry utilizes parallax bundle adjustment with more convergency which is also proposed in this work.The thesis also analyzes the sparsity of monocular parallax visual odometry and design an efficient algorithm for fast computation.Experiments show that monocular parallax visual odometry archieves a real-time estimation for localization,with the precision of 15 cm.To enhance the precision in high frequency SLAM,this thesis presents an EKF based approach: RI-MONO-EKF.RI-MONO-EKF utilizes the proposed RI-IMU factor which can fully capture the covariance information among all variables in IMU sensor.Meanwhile,initialization technique is also given based on RI-IMU preintegration factor,which makes the SLAM can estimate its initial state information by robust estimation.Experiments show that initialization technique can notably fasten the convergence of state variable and improve the precision high computation efficiency.At last,to further archieve more precision in real-time location estimation,this thesis proposes a keyframe,graph optimization based monocular visual-inertial SLAM with RIIMU factors and initialization technique mentioned above.The SLAM utilizes multi-thread cooperation and can estimiate the localization information and build sparse structure in realtime only by CPU.Tremendous experiments also validate the fact that the SLAM can archieves the state of art localization precision with 10 cm,with the frequency of 50 Hz.