Research On Simultaneous Localization And Mapping Based On Kinect

SLAM(simultaneous localization and mapping)technology has been a research hotspot of robotics and computer vision in recent years,and is widely used in the fields of driverless,home robots and so on.Vision sensors are main device of Visual SLAM technology,through the visual sensors such as monocular camera,binocular camera,and RGBD camera,we can perceive the surrounding environment,acquire scene features from image information and estimate its position,and establish kinds of environmental maps.By mounting visual sensors on mobile robots or smart cars,robots can explore and develop unfamiliar environments.Visual SLAM has been widely used in intelligent driving,defense,rescue,and service fields.In this paper,we focus on the study of environment perception,pose estimation,self-positioning and 3D construction in an unfamiliar environment based on mobile robot and RGBD camera.The visual SLAM system mainly includes visual Odometry,back-end optimization and loop closure detection,In the visual SLAM front-end solution,build a Kinect-based visual SLAM platform and construct a feature-based visual Odometry.the Kinect depth image restoration method and ORB feature extraction methods are explained.The inter-frame pose is estimated by extracting feature and feature matching.In addition,the visual SLAM algorithm based on graph optimization is analyzed and the loop closure detection algorithm based on deep learning features is proposed.Under the complicated conditions of environment,compared with Kalman filter,extended Kalman filter and particle filter optimization algorithm,the graph optimization algorithm is used to unify the pose and landmark in the nonlinear optimization framework such as Gauss-Newton and LM algorithm,and can get better global accuracy than Kalman filter.Loop closure detection is an essential part of visual SLAM.Loop closure detection aims at detecting the places that robots have visited through location identification method,and correlates with historical data to enhance constraints and repair errors,making SLAM system more accurate and robust.Compared with the traditional loop closure detection method,this paper proposes a location recognition algorithm based on deep learning features.By extracting appearance-invariant deep learning features,a better detection result is achieved especially in complex and variable external conditions.The main work and innovations of the paper are as follows:The experimental platform of visual SLAM algorithm based on Kinect and Kobuki is built.The Kinect depth image is repaired by filtering and smoothing.Based on the platform,the feature-based visual odometry model is analyzed.Compared with SIFT,SUFT,the performance of ORB is better.Then we can obtain the feature matching point pairs by feature extraction and matching methods.The improved graph transformation method is used to further filter the error matching and reduce the mismatch,which lays a foundation for more accurate pose estimation.Visual SLAM back-end optimization technology: The back-end schemes based on Kalman filtering,extended Kalman filtering and particle filtering are discussed.However,due to its Markov property,the current state has been only related with the previous moment.Aiming at this problem,the back-end algorithm based on graph optimization is to unify the data such as pose and features into the nonlinear optimization frameworks such as Gauss Newton and Levenberg-Marquart,which has better global optimization precision.Loop closure detection,which also called place recognition,is a critical part of simultaneous localization and mapping(SLAM).Some classical methods using handcrafted features to address this issue.Recently CNN features were also used to loop closure detection and achieved comparable or state-of-the-art performance.So we further studied the characteristic of feature maps and proposed a novel algorithm,which aims to understand which of the feature map is of great importance for place recognition and which is not,especially in complicated conditions.We presented extensive experiments on several challenging benchmark place recognition datasets and compared with the method using complete CNN features.The result demonstrates the proposed approach has a meaningful improvement on place recognition,especially for severe appearance change.