Research On 3D Point Cloud Map Construction In Indoor Environment Based On Kinect Technology

Into the 21 st century,people on the intelligent robot requirements gradually increased,and use the intelligent robot to complete all kinds of work is the inevitable trend of technological development.Construct the 3D map is the premise and foundation of the mobile robot in unknown environment,so it has become a hotspot issue in the field of computer vision and robot.With the continuous development of all kinds of technology and requirements,the 3D map construction method in performance of real-time and accuracy is facing the new challenges.In view of this situation,we carried out extensive research for 3D map construction method in indoor environment based on Kinect.Our works are as follows:(1)The Kinect hardware structure and the imaging principle are described in detail.Use the cross-platform open source framework OpenNI to read the rgb images and depth images collected by Kinect,and preprocess the images with the associated components of OpenNI.Calibrate the Kinect and take the intrinsic parameters of the camera,thus generate the 3D point cloud based on rgb images and depth images collected by Kinect.(2)In order to obtain complete 3D map of surrounding environment,need to register the data collected by the camera at different viewpoints to the same coordinate systems,the procedure of calculating the coordinate system transformation parameters is called as frame-to-frame alignment.The features extraction and match algorithms are introduced simply,and the performance of three features extraction algorithm is compared by a features extraction experiment,the SURF is selected as the features extraction algorithm.In view of the defects of related algorithm,a 3D features ICP algorithm based on discrete selection mechanism is proposed.(3)Due to the results of frames registration process have error inevitable,the error will be gradually accumulated in 3D map reconstruct process and seriously affect the map precision finally.Thus in this thesis,construct the pose-gragh with the camera pose and relative motion,an loop detection algorithm based on tree structure visual bag of words is proposed to further increase the constraint of position,and build visual vocabulary tree,on the basis of traditional TF-IDF weight algorithm to detect loop,introduced the character of relative position invariant of the objects in three dimension space to reject the loop detection error,improved the perceived ambiguous of detection algorithm.Finally,global optimize pose-gragh with g2 o graph-based optimization framework,obtain the more accurate camera pose,improve the accuracy of the map.In order to verify the validity and feasibility of the 3D point cloud map construction method proposed in this thesis,tested in six groups of public datasets and the real environment.A large number of experimental results showed that the algorithm proposed in this thesis in complex environment with good performance of accuracy and robustness.