| Vision based SLAM(Simultaneous Localization and Mapping)system is widely used in automatic driving,robotics,augmented reality and other fields because of its low price and wide application scenarios.Among them,the feature point method,as one of the common methods in visual slam system,estimates the camera motion and restores the scene by identifying and tracking the feature points between frames.The traditional feature point method is based on the assumption of static scene.Once a dynamic object appears in the scene,the dynamic points attached to it will interfere with the tracking process and ultimately affect the accuracy and robustness of the system.In order to solve the impact of dynamic feature points on the visual system,this thesis adopts the idea of semantic geometry fusion and proposes a dynamic point elimination algorithm based on instance segmentation and Delaunay triangulation.The specific innovations of the algorithm are as follows:In the semantic stage,this thesis proposes a method to classify common indoor objects according to the potential dynamics of objects,and divides the feature points into absolute static,dynamic and absolute dynamic.In the geometric stage,this thesis proposes an algorithm to judge the dynamics of feature points based on Delaunay triangulation.The algorithm uses the absolute static point as the benchmark to calculate the offset between the active and absolute active points and the benchmark point.Finally,its dynamics is judged by comparing the relationship between the offset value and the threshold.Through the trajectory error comparison experiment,the robustness of the algorithm in dynamic scene is proved.Finally,based on the real scene,this thesis designs and implements a SLAM system with client-server architecture,and introduces the inertial measurement unit to provide the motion information of the camera itself.The client of the system is based on Android system,which is mainly responsible for the collection and upload of visual inertia data and the rendering and interaction of the processing results of the server;The server is implemented based on Linux system,receives the data uploaded by the client,tracks and maps,and returns the operation results.Through the function and performance test of the whole system,the stability of the system is proved. |
