Research On Visual Odometer Based On Optical Flow Tracking And Feature Matching

Visual odometer,also known as visual positioning,refers to the use of image information collected by the camera during the robot's travel to estimate the position change of the robot.Compared with the traditional positioning method,visual positioning can overcome the shortcomings such as data loss and inaccurate positioning caused by wheel slip,which has become an important research direction in the field of robotics.However,the current visual odometer still has the following problems:1 point feature detection algorithm has poor real-time performance and robustness,which affects motion estimation.2 Poor performance in scenes such as texture loss and dynamics,ignoring structural line features.3 Closed loop detection is inefficient and difficult to apply to large-scale scenarios.In this paper,the visual mileage calculation method is studied for the above problems.Based on the visual positioning based on point features,a visual mileage calculation method with better environmental adaptability,real-time performance and stronger robustness is proposed.The main research contents are as follows:(1)Several common point feature extraction algorithms are analyzed and compared,and based on the ORB feature extraction,combined with the maximum stable region algorithm,the region is first divided and the point features are extracted to ensure the quality of the point features.Save on computing overhead.At the same time,the improved optical flow method is introduced to quickly track the movement of feature points between images,eliminating the time-consuming feature description and matching process.(2)Aiming at the problem that the features of extracting points in the under-texture environment are less and the pose estimation is difficult,a strategy for extracting the features of the dotted lines in the environment is proposed.Firstly,the LSD algorithm is used to extract the line features and improve the line feature matching process,so that the mismatch can be removed accurately and quickly.Then,by calculating the static weight of the line feature,the influence of the dynamic feature is eliminated.In solving the camera pose,the point feature is used to estimate the initial pose,and then the point line integrated graph model is constructed to obtain the camera motion between key frames,avoiding the problem that the positioning system is not robust due to the single feature.(3)Aiming at the problem that the closed-loop detection is inefficient and difficult to apply to large-scale scenes,a closed-loop detection algorithm based on visual dictionary is proposed.The visual dictionary constructs the feature points into a vocabulary tree by means of offline,so that the image similarity can be judged quickly.Through the closed-loop detection,the inter frame constraint is added,and then the pose map optimization is performed by using the general graph optimization tool g2 o to obtain a globally consistent camera pose.Finally,the Kinect camera was used to build a visual odometer system based on thedotted line feature.Multi-group comparison experiments under the TUM data set verify that the proposed algorithm can meet the real-time requirements and improve the robustness of the system.At the same time,the experimental analysis was carried out in the indoor scene environment,which verified that the system constructed in this paper has good positioning accuracy.