Robust And Efficient 3D Registration And Structure Recovery For Challenging Environment

In recent years,artificial intelligence is advancing at unprecedented speed,and has widely applied in kinds of areas.In applications like mobile robots and augmented re-ality,an intelligent device must have the ability to recover its 3D motion in the physical 3D space,and the 3D structure of the scene.The video-based camera tracking and 3D perception,only requires a common camera,and has become an indispensable and es-sential technique for the future smart mobile devices.This technique can be divided into the off-line Structure from Motion(SFM)and on-line Visual Simultaneous Localization and Mapping(V-SLAM)according to whether or not real-time performance is required.As the application scenarios become more and more complex and diverse,current SFM and V-SLAM methods face the challenges in terms of robustness and efficiency.On the one hand,vision based SFM and V-SLAM easily fail in a complex environment.For example,current methods have difficulties in handling large-scale scenes and fast motion with strong rotation,and the stability largely relies on the richness of image texture.On the other hand,due to high redundancy of image sequence in both spatial and temporal domain,current methods heavily consume computational resources,even need to leverage the parallelism power of GPU,thus cannot be applied on mobile devices with low computing power.In addition,as the scene becomes larger and larger,current methods easily meet the bottleneck of memory and efficiency.In order to solve the above issues,this thesis deeply studies the problems of SFM and V-SLAM in challenging environment,and proposed a series of rubust and efficient SFM/V-SLAM methods/systems,which not only significantly improves the robustness and efficiency compared to previous methods,but also can nicely satisfy the demand of practical applications.Specifically,the main contributions include:? We propose a novel structure from motion framework for large scale scenes.By leveraging a loop detection and closing method based on non-consecutive feature matching,and a segment-based bundle adjustment,the proposed method performs efficient global optimization to eliminate accumulated error for large scale scenes with limited memory space,and achieves efficient and accurate 3D registration for multiple video sequences and real-time simultaneous localization and mapping for monocular video sequence.? We propose a robust and efficient keyframe-based monocular SLAM method.By leveraging a multi-homography based feature tracking method,and an efficient lo-cal map expansion and optimization strategy,we effectively tackle the difficulty of robust camera tracking under strong rotation and fast motion by current keyframe-based monocular SLAM methods,and achieve much higher efficiency than state-of-the-arts like ORB-SLAM and LSD-SLAM.Besides,we proposed to simulate IMU measurements by utilizing visual information and incorporate them into op-timization function to further improve the robustness to motion blur or featureless scenes.? We propose an efficient and accurate SLAM method with an RGB-D camera.By combining the dense alignment of low-resolution RGB-D images and the sparse feature based tracking,and further leveraging a novel incremental bundle adjust-ment,the proposed RGB-D SLAM system is more efficient and achieves better accuracy than other methods.Even using only CPU,our method still can achieve strong real-time performance.