| Visual SLAM is the key technology for robot localization and three-dimensional map construction.Aiming at the problem that SLAM can be easily interfered by moving objects in the environment,this thesis focuses on the study of robust visual location in dynamic scenes.The main research contents are as follows:(1)A Cam Shift target tracking and recognition algorithm based on instance segmentation is proposed.In the given algorithm,Cam Shift is used to track the instance.When the instance segmentation fails,the result of Cam Shift will be regarded as instance.When Cam Shift tracking fails,it will be initialized with the result of YOLACT.Experiments show that the algorithm has higher object tracking accuracy than Cam Shift,and is more robust than YOLACT.(2)A visual SLAM algorithm based on distinguishing moving object identification is studied.the objects are divided into three categories as the prior conditions for dynamic object recognition,and optical flow method is used to track the feature points of non-dynamic object regions,and then establish the judgment condition is established to judge dynamic and static feature points,so as to.a Avoiding the moving objects interfere SLAM system,and improve the accuracy of positioning.Experiments show that the algorithm can is able to improve the positioning accuracy by up to 53.3% in high dynamic environments.(3)A loop detection technology to eliminate interference from potential moving objects is studied.Aiming at the problem that potential moving objects often cause loop detection failures,semantic information is used to remove the features from potential moving objects,and then calculate the feature vector by Bo W.Experiments show that this algorithm has a higher loop recall rate and can improve the positioning accuracy of the SLAM system in dynamic environments. |
