Research On Vision-based SLAM Method In Dynamic Environments

SLAM(simultaneous localization and mapping)is the key technology for mobile robots to realize autonomy and intelligence.Compared with inertial navigation,lidar and other sensors,the cost of vision sensor is lower,and the collected image can provide more environmental information.SLAM algorithm based on various vision sensors has achieved a lot of research results.However,the existing visual SLAM algorithms all assume that the camera works in a static environment.In actual application scene,moving objects will inevitably appear,which will become interference elements in the image and seriously affect the pose estimation of the robot and the environment map established by the robot.To solve this problem,this paper studies a robust visual SLAM algorithm in dynamic environment,uses a method based on monocular vision to estimate the camera’s pose information,and establishes a static semi-dense map without motion elements.The main work includes:The visual SLAM system under dynamic environment is analyzed,and the monocular camera is modeled,and the transformation relationship between the two-dimensional coordinates of the image and the actual three-dimensional coordinates is deduced.The implementation principle and existing problems of ORB-SLAM2 algorithm are further studied,and on this basis,an overall framework suitable for robust visual SLAM system in dynamic environment is designed.Aiming at the problem that the localization accuracy of visual SLAM system is declining in dynamic environment,a new monocular visual SLAM localization algorithm based on dynamic feature point elimination is proposed.Using the optical flow dynamic detection algorithm,the feature points are divided into static points and dynamic points,and the static points are used to estimate the camera pose,so as to complete the monocular visual SLAM localization function in dynamic environments.Experimental results show that the proposed localization algorithm can significantly improve the localization accuracy of SLAM system.Aiming at the problem that the sparse point cloud map constructed by monocular visual SLAM system can not meet the complex tasks such as navigation obstacle avoidance and intelligent capture,a new static semi-dense map construction method is proposed.Based on the localization algorithm,the precise pose transformation between key frames in dynamic environment is obtained,and the inverse depth estimation of pixel points is obtained by searching along the polar line between key frames Then deep smoothing processing is carried out on the whole key frame,and noise removal is carried out on the recovered map points on the dynamic object,so as to construct the globally consistent static semi-dense map.Experimental results show that the proposed map construction method can build a relatively clear semi-dense map without moving elements,which can meet more requirements in the future.