Mobile Robot Visual Odometry Algorithm Reserch

As an important part of the mobile robot Simultaneous Localization And Mapping(SLAM),visual odometry is very important for mobile robot autonomous positioning and mapping navigation tasks.Due to the error of the mobile robot's own to sensor and the factors such as blurred image,too fast motion and missing features,the accuracy of the visual odometry cannot be guaranteed.Therefore,this paper analyzes the existing problems of the visual odometry calculation method.The main research contents are as follows:Firstly,we summarize the significance and background of the research,analyze the research status of visual odometry calculation method about some key problems,and introduce the existing open source algorithms of mobile robot visual odometry.The key problems of visual odometry are analyzed according to different algorithm schemes,which provide theoretical models and basis for subsequent research.Secondly,aiming at the disadvantage of poor positioning accuracy caused due to limited scene information obtained by sensors,a visual odometry with multiple pose constraints is proposed.The singular value decomposition method is used to solve the initial pose,and the geometric error or re-projection error is established according to if with or without depth information of the registration image points.The Gauss-Newton nonlinear optimization is used to simulate the experiment,and the improved algorithm improves the positioning accuracy.Thirdly,in order to improve the accuracy of visual odometry in fast motion environment,a semi-dense visual odometry calculation method of mobile robot mixing is proposed.Using the feature point method to track the camera pose,obtain the initial pose and key frames,the Sobel convolution kernel is used to obtain the pixels with obvious gradients,register the key frames by semi-dense direct method,estimate the relative motion between key frames.Finally,the key frame's pose and the transformation between adjacent frames are optimized locally by the pose map.Experimental shows that the improved algorithm can meet the requirements of autonomous positioning in fast motion scenarios.Finally,in order to overcome the limitations of single sensor measurement and further improve the accuracy of visual odometry positioning,a visual inertial multi-sensor fusion algorithm based on line features is introduced.In the front end,the Inertial measurement unit(IMU),is used to measure the state information and calculate the initial pose.In the back end,the tightly coupled method is adopted.The line feature extraction re-projection error and the IMU pre-integration state error constraint are added to the non-linear optimization to improve the accuracy of the algorithm.