Research On Visual SLAM Key Technologies In Complex Indoor Environments

SLAM is the core technology of intelligent robot positioning in unknown environment and complex area.The efficiency of SLAM algorithm directly affects the accuracy,robustness,and real-time of the whole navigation system of the intelligent robot.Visual sensor has become one of the most widely used sensors in SLAM technology because of its rich information,wide application range and expandability.However,visual SLAM does not perform well in complex indoor environments where dynamic objects interfere,weak texture scenes exist,and continuous rotation and fuzzy motion exist.The main reasons are:the existence of dynamic objects makes the feature association of the system unreliable or incorrect;weak texture scenes result in insufficient number and uneven distribution of point features extracted by the system;when the camera rotates continuously or preforms fuzzy motion,the system visual odometer deteriorates and cannot track the feature points of adjacent frames.These problems all lead to many kinds of positioning and composition defects,such as offset distortion and incomplete mapping of the system estimating track integrity.In order to improve the positioning and mapping accuracy,robustness and real-time performance of the visual SLAM in complex indoor environments,this paper studies the visual SLAM algorithm from the perspective of multi-feature and multi-sensor fusion.In this paper,the indoor dynamic object detection algorithm is used as a starting point to improve the accuracy and speed of indoor dynamic object extraction from the perspective of image segmentation.On this basis,a robust SLAM algorithm for indoor dynamic environment is proposed to improve the robustness and accuracy of visual SLAM positioning in indoor dynamic scenes.Furthermore,in view of the stability and localization defects of visual SLAM algorithm in weak texture environment,a multi-feature visual SLAM algorithm in indoor weak texture environment is proposed,considering the multi-feature visual SLAM solution first.Then,considering the SLAM scheme of multi-sensor fusion,a SLAM algorithm with indoor inertial/visual tight coupling is further proposed to improve the performance of the system in case of continuous rotation and fuzzy motion.The main content of this paper is around the following four aspects:(1)An indoor dynamic object detection and extraction algorithm is proposed to overcome the problems of misrecognition and inaccuracy of existing object detection and extraction algorithms.To solve the problem that the existing YOLOv3 algorithm cannot accurately detect and identify indoor objects,the ability of YOLOv3 detecting indoor objects is improved through the transfer learning of small sample target data domain.To solve the computational complexity caused by local image segmentation in the object extraction algorithm,an improved global K-means image segmentation algorithm combined with YOLOv3 is proposed,and the uncertain number of segments and the slow speed of the improved K-means algorithm with maximum extremum method and pre-classification technology are also presented.On this basis,a multi-parameter extraction mechanism is proposed to further improve the accuracy and anti-interference of object extraction.The experimental results show that the proposed algorithm is superior to other traditional target extraction methods in terms of accuracy and speed,and provides a stable feature selection strategy for the subsequent study of SLAM algorithm in indoor dynamic environment.(2)A robust SLAM algorithm for indoor dynamic environment is presented to solve the unreliable feature extraction and inaccurate matching pairs in the indoor dynamic object interference scene of a general visual SLAM system.Considering that people can cause unstable and mismatched feature points,which can lead to errors in camera pose estimation and incomplete mapping,this paper presents a dynamic region detection method based on dynamic target semantic information.On this basis,a dynamic target feature point culling algorithm based on geometric relationship is presented,which ensures that the real dynamic target and its related feature points can be identified and culled.The experimental results show that the proposed algorithm has better positioning accuracy and real-time performance than the original ORB-SLAM2 algorithm and other dynamic environment visual SLAM algorithms in indoor dynamic environment.(3)To solve the problem that the number and distribution of feature points extracted by point feature visual SLAM algorithm in indoor weak texture environment are insufficient,and feature point tracking is easy to lose,this paper presents a multi-feature SLAM algorithm for indoor weak texture environment.To solve the problem that traditional line detectors can not detect and extract line features in real time,an Edlines algorithm with fast line detection speed is introduced.To solve the problem that the uncertainty of line features cannot be measured by geometric models like point features,this paper presents a method of modeling and measuring the uncertainty of line features based on the scale of entropy to eliminate the unreliable line features.Furthermore,in order to improve the accuracy of system looping detection,a word bag model combining point and line features is proposed to solve the looping error detection problem of point feature visual SLAM algorithm in weak texture environment.The experimental results show that the positioning accuracy and real-time performance of the proposed algorithm in indoor weak texture environment are better than other multi-feature SLAM algorithms.(4)To solve the problem of inadequate positioning accuracy and robustness of indoor visual SLAM system in weak texture environment,pure rotation or motion blurring,an indoor inertia/visual tight coupling SLAM algorithm is proposed.Firstly,an adaptive filter is proposed to reduce the noise of MEMS IMU.Secondly,the IMU pre-integral model is proposed and constructed to solve the complex calculation problem introduced by the repeated integral in the system back-end optimization process.In the aspect of inertia/visual fusion,aiming at the problem of insufficient real-time inertia/visual initialization,a joint initialization model of inertia/visual loose combination is proposed and gravity amplitude is introduced to improve system visibility.In addition,a tightly coupled inertial/visual combination model is proposed from the optimization point of view,and the local optimization information constraints of the algorithm are solved by adaptive local window and fixed window methods.Finally,the experimental results show that the proposed algorithm is superior to(3)and other inertial/visual localization algorithms in the continuous rotation and motion blurring of indoor unmanned aerial vehicles,handheld inertial/visual sensors,small robots and weak texture environment experiments.The effectiveness and applicability of the algorithm are verified.