Research On Monocular Visual Odometry For Mobile Robot

Visual odometry is the process of estimating self-motion by using one or more cameras bounded to moving objects(such as mobile phones,vehicles,robots).At present,VO is often used as a building block of a full SLAM algorithm.Because of its different sensors,it can be divide into single,stereo and RGBD odometry.Traditional odometry generally take advantage of inertial sensors IMU ? laser measurement instruments ?ultrasonic sensors and infrared ranging sensors to achieve the localization of the robot.But there are many disadvantages such as vehicle slippage,large measurement error and so on,and it is easily affected by terrain,weather and other environmental factors.However,the visual odometry only needs to obtain the image information to be able to solve the problem of robot precise localization.Because of its low cost,fast processing speed,simplicity,high accuracy,monocular visual odometry has gradually become an important choice in visual navigation.Firstly,the method of optical flow can also find the corresponding feature points of adjacent frames,and need not to describe the features.In addition the method of optical flow has the high localization accuracy in small range of angle rotation between the adjacent frames and the method of feature matching is more robust to the large range of visual field changes.In order to raise the accuracy,an improved 3D-2D motion estimation algorithm is proposed which needs to optimize 3D points twice,and combines feature match and optical flow to select the best result.Experiment results imply that the proposed algorithm improves the accuracy of the estimation effectively.After that,For long distance odometry,the error is inevitable,and with accumulation of camera pose error,the cumulative error will greater and greater,so wo need to optimize the estimation results.This thesis introduced the graph optimization and optimization tool g2 o in detail.The camera pose and transform constraint of adjacent frame estimated by the motion estimation will be constructed pose graph,and then optimize the estimated results to obtain the optimal results.At the same time,in order tospeed up the operation and avoid redundant data,this thesis propose a key-frame selection strategy which based on the visual contents and inter frame motion.Finally,in order to verify the feasibility of the algorithm,we carried out experiments in public data sets and in a natural environment,compared with estimated results with the groundtruth,and analyzed the errors of rotation and translation.It is proved that our method can finish the accurate localization of mobile robot.