Research On Motion Blur SLAM System Of Mobile Robot Based On RGBD

Motion blur caused by high-speed camera motion is a major problem in the field of visual SLAM research.Due to the blur of the image,the feature points with the same name become fewer,and effective feature information cannot be extracted from the image,so there is no way to get the correct match,which affects the camera Pose estimations and reduces the robot's positioning accuracy.In severe cases,it will directly lead to positioning failure and the system's robustness is poor.Once the camera movement speed is too fast,the situation of motion blur cannot be avoided.For this visual problem,this article improves on the basis of the classic visual SLAM algorithm,and uses the D435 camera to build an RGBD-SLAM system that responds to motion blur.The research contents are as follows:First,the RGBD-SLAM system framework based on D435 camera is constructed,which mainly includes three parts: image preprocessing,front end and back end.The image preprocessing part involves the acquisition of motion blurred images and the processing of image noise.The front end mainly involves feature extraction and matching and camera pose estimation.The back end mainly deals with pose optimization and map construction.Second,the design of visual odometer based on point features.First extracting the ORB feature points and calculating their binary descriptors,and then using the Hamming distance to perform pairwise matching between images.Based on the improvement of the traditional RANSAC algorithm,the model establishment time is shortened,the feature point matching quality and matching efficiency are improved,and finally the robot pose is estimated based on the feature matching result.Third,research on the back-end pose optimization algorithm based on ORB features.Under the premise of satisfying the real-time performance of the system,a key frame screening method based on Brenner gradient is proposed,which screens out the blurry frames and does not participate in the back-end optimization to improve the positioning accuracy of the system.Introduce the update strategy of the map,construct the bag-of-words model for loop detection,and finally construct a three-dimensional point cloud map.Fourth,using D435 to build a motion blur SLAM software and hardware system.According to the experimental data comparison provided by TUM,the accuracy of the visual SLAM system proposed in this thesis is verified by calculating the error between the camera trajectory provided by the data set and the ground real trajectory.Use the Turtlebot2 robot equipped with the D435 camera to perform real-time experiments in the actual environment,change the robot's movement speed,observe the system relocation and map reconstruction effects,and verify the robustness of the vision system proposed in this thesis in the case of image blur caused by rapid camera movement.