Research On Dynamic SLAM Based On Deep Learning And Optical Flow Acceleration

With the development of intelligent robot,unmanned driving and other fields,the moving subject needs to estimate the pose and realize the environment interaction in real time,and even need to move and estimate its position in the new environment like human.Visual simultaneous localization and mapping(VSLAM)uses the images taken by the camera to calculate the pose of the camera and build the environment map,which can be used in the above fields.However,if there are moving objects(human,animal or vehicle,etc.)in the surrounding environment,the calculation pose error is very large and can not meet the use requirements.Aiming at the problem of poor applicability in dynamic environment,this paper uses epipolar constraint and proposes a object detection network classification strategy to eliminate dynamic feature points.This paper optimizes the ORB-SLAM2 system.Aiming at the timeconsuming problem of extracting feature points,LK(Lucas Kanade)optical flow is used between ordinary frames to accelerate the operation of the system.The main research contents are as follows:(1)In order to solve the problem that dynamic feature points seriously affect the operation accuracy of VSLAM system,the epipolar constraint and object detection network are used to eliminate dynamic points.In this process,the motion trend of the detected target is classified into three categories: high,medium and low,and the different thresholds are set to 0.3,0.4 and 0.5respectively,so as to retain more static points and reduce the error.Through data set simulation,the root mean square error(RMSE)of the average trajectory error is reduced from 0.5 to 0.02,which proves that the algorithm in this paper meets the operating conditions in the dynamic environment.(2)In order to solve the problem that ORB-SLAM2 consumes too much computing power in ordinary frames,LK(Lucas Kanade)optical flow method and random sample consensus(RANSAC)method are used to track the pose of ordinary frames,optimize the running time,and change the key frame insertion conditions to make it related to the number of feature points matched by optical flow.Experiments show that after the change,the system can still insert key frames stably,and the error is basically unchanged.The average image of each frame of the original system is about 0.04 seconds.After using optical flow acceleration,the average image of each frame will be reduced by 0.017 seconds.It is reduced by about 15 seconds on the data set of 900 images,and the operation accuracy changes little.