Research On Visual Slam For Dynamic Environment

Simultaneous localization and mapping(SLAM)is the key technology for robots moving in unfamiliar environment.Most of the current research on visual slam is based on the static environment.If there are a large range of dynamic objects in the environment,it will cause great interference to the SLAM system,which hinders the practical application of visual SLAM.Aiming at the problems of traditional visual slam in dynamic environment,combined with deep learning algorithm,this paper studies how to improve the robustness of visual slam in dynamic environment.In order to overcome the influence of dynamic objects,dynamic objects need to be recognized and eliminated,so as to locate and build the map with the remaining static scene.Dynamic objects not only affect the visual odometer,but also interfere with the accuracy of loop detection.With the help of deep learning convolution neural network,the structure flow suitable for dynamic object detection and loop detection is established,and the vision slam system based on depth camera in dynamic environment is built.Firstly,the method of detecting dynamic feature points is studied.By comparing the common methods of target detection,we choose Yolo v4 to recognize the dynamic objects.As the vision SLAM is easily affected by the large area of dynamic objects and the speed requirements,this paper improves the lightweight of Yolo v4,selects Mobile Net v3 as the backbone network to extract features,and compares the improved network model with the original method.Secondly,the specific process of eliminating feature points on dynamic objects is designed.Based on the semantic information provided by target detection,combined with RANSAC algorithm and epipolar geometry,the number of potential dynamic feature points is reduced.It improves the ability of association between the front and back frames when removing dynamic feature points,and enhances the fusion of depth learning and image geometric relationship.Then the loop detection based on Mobile Net v3 is studied.Mobile Net v3 is used to detect the loop and improve the loop detection process.The deep learning convolution neural network is used to replace the bag of words model.Finally,the experimental system is developed for comparative test.The loop detection data set is selected to test the performance of the loop detection module,and the different dynamic range of the tum data set is selected to test,which is compared with the classic ORB SLAM2.At the same time,the real environment test platform is built,and the real environment experiment method is designed to show the test effect compared with ORB SLAM2 method in real environment,which verifies the correctness of this method.