Simultaneous Location And Mapping Of Mobile Robots In Dynamic Scenes Based On Vision

Robotics and computer vision technology began to be increasingly applied in actual production and life following the development of science and technology these years.Among multitudinous technologies,Simultaneous Location and Mapping(SLAM)is particularly important for robot autonomy.Since robots are unable to use GPS positioning and navigation indoors,one solution is to make robots self-mapping there and to use such map to navigate.In a simple static scene,the use of distance sensors can be commendable to achieve this goal.But in complex dynamic scenarios,robots need to better understand the environment,at this time the use of visual sensors can complete the task in a better way.Thereby,this paper discusses the visual SLAM technologies in dynamic scenes.This paper proposes and obtains a solution for dynamic scenarios through stereo computer vision technologies,the main research contents are as follows.Firstly,we model the binocular camera,complete the camera calibration as well as image correction,and preprocess the image collected by the camera,so that the prepared picture frame can be better used in the algorithm.Secondly,we improve the existing open source algorithm,so that it possesses a good adaptability to the dynamic scenes.The traditional robust estimation algorithm is used to eliminate the influence of dynamic targets,which account for a small proportion of the image.For a large proportion of dynamic targets,by training an instance segmentation network to complete the tracking of a dynamic target in pixel level to eliminate the target with transcendental dynamics.For the stability of the detection,we remove the dynamic target by calculating the sparse scene flow and combining with the instance segmentation.For the problem of feature extraction and matching instability in dynamic scenes,we manage to extract the features that can be relied on constantly in picture frames and combine with a robust matching algorithm,so that the extracted features can be tracked for a longer period of time,and can prevent the mismatch caused by dynamic object occlusion.Finally,the effectiveness of the verification system is analyzed through designing experiment.By comparing with ORB-SLAM under different data sets and after he systematic performance is quantitatively analyzed,the experimental results show that the system designed in this paper is significantly better than ORB-SLAM in the dynamic scenes.By testing the environmental adaptability of verification system in different practical scenarios,the experimental results show that the algorithm proposed in this paper can run effectively in different environments.