Research On RGB-D Visual SLAM In Indoor Dynamic Scenes

The core technology of mobile robots is real-time simultaneous localization and mapping to realize autonomous movement of robots in the environment.Traditional SLAM technology builds a map based on a static environment assumption,but in the actual environment,the movement of pedestrians,vehicles and objects will cause dynamic changes in the environment,making the map created by the SLAM system unable to maintain consistency for a long time.In order to make SLAM work properly in dynamic environments,this paper conducts research and exploration in the following aspects in terms of vision and semantics,respectively.The target detection algorithm in deep learning is added to the SLAM system,the priori dynamic objects in the scene are eliminated by the target detection,and the actual dynamic objects are further filtered out by the multi-view geometry technology,so as to improve the positioning of the system Accuracy and robustness.(1)The YOLACT++ network(Semantic segmentation network)is used to remove dynamic objects in indoor scenes to solve the influence of dynamic objects on scene positioning and mapping in dynamic scenes.This paper adopts the YOLACT++ algorithm,which meets the requirements of the SLAM system for real-time performance and accuracy.YOLACT++ adds the Mask Re-scoring module and the variable convolution module,which further improves the detection accuracy of objects in the scene based on the high real-time performance of YOLACT.(2)In the feature point extraction stage of ORB-SLAM2,since the feature points extracted by the ORB algorithm are too uniform,many low-quality feature points are extracted.In this paper,the quadtree-based feature point extraction algorithm in the ORB-SLAM2 is improved,and the division depth of the quadtree is limited based on the number of feature points required for each layer of the image pyramid,reducing the number of iterations of quadtree division and reducing Extraction of low-quality feature points.(3)The two methods based on semantic information and geometric constraints are combined to reduce the impact of dynamic objects on the accuracy and stability of the visual SLAM system.The target detection algorithm YOLACT++ is used to detect a priori dynamic objects in the environment,filter out the a priori marked dynamic objects,and then detect the truly dynamic objects through multi-view geometry.The two methods are combined to optimize the accuracy and stability of visual SLAM.