SLAM System In Dynamic Environment

The existing SLAM system,which is based on static processing and application environment,needs feature points on stable static objects to extract key points to ensure the correctness of the matching relationship between the front and back frames.However static environment is still a minority in life.How to run robustly in the dynamic environment is what the thesis need to solve.In order to realize this system,this thesis makes innovations in the following three points:Firstly,the YOLOv5s-seg model is embedded in the system in this paper.It can provide more accurate initial value and position information for logical algorithm because of its advantages in dealing with illumination and rotation invariance.So,the YOLOv5s-seg model is used to preliminarily detect the position of potential dynamic objects.The mask is used to extract feature points from the static background and the static high-quality feature points are used to estimate the pose of the camera’s photometric reprojection error,so as to obtain a relatively stable camera pose.Secondly,the use of deep learning requires a large amount of computing power and its real-time performance is poor.In order to improve the efficiency of operation,it is not necessary to recognize every frame.The logical algorithm is used to predict and propagate the mask image information backward,and the information data is connected and transmitted through the algorithm.The pose of each potential dynamic object is obtained by perturbation of the camera pose.Then,the geometric relation in multiple views is used to judge whether the object is moving.According to the epipolar constrain between 2D and 2D points,the key point in the previous frame which is in different views will fall into the epipolar in the second frame.And a certain threshold is set to judge whether the object is moving.At the same time,the projection residual is used to judge by calculating the sum of all gray errors which is compared to thresholds.And transform the pose of the mask image by using the estimated pose of the dynamic object.It speeds up the instance segmentation mask provided by the front end of SLAM systemThe thesis realizes the robust operation of SLAM system in dynamic environment,significantly improves the accuracy compared with traditional SLAM system.And improving the operating efficiency of deep learning network model embedding to a certain extent.