Visual and lidar-based SLAM by variational bayesian methods

• Main Author: Jiang, Xiaoyue
• Other Authors: Justin Dauwels
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: https://hdl.handle.net/10356/139813
• Institution: Nanyang Technological University

Autonomous vehicles, such as self-driving cars, unmanned aerial vehicles, planetary rovers, and autonomous underwater vehicles, have found applications in a wide variety of domains in recent years, from transportation to searching, rescue, and military. A major challenge for designing autonomous vehicles lies in understanding the surroundings via their onboard sensors. A technology to bridge over this difficulty is Simultaneous Localization and Mapping (SLAM). SLAM continuously scans and ``learns'' about the operating environment, allowing vehicles to augment the physical world with digital content depending on their locations within this space. SLAM has been extensively studied over the past decades resulting in a variety of solutions based on different types of sensors, including cameras, Global Positioning System (GPS), Light Detection and Ranging (Lidar), Inertial Measurement Unit (IMU), Radio Detection and Ranging (Radar), and Sound Navigation and Ranging (Sonar). The most widely applied two types are Visual SLAM (vSLAM) and Lidar-based SLAM. Techniques based on vSLAM and Lidar-based SLAM fundamentally investigate a computational problem of localizing a vehicle while simultaneously constructing or updating the environmental map. This estimation can be addressed via two forms: online and offline. Online updates the most recent pose and map, whereas offline approximates the entire path and map. Both forms are of high importance and wide applications in the real world. Existing online algorithms are referred to as filtering. Popular approaches like Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), Extended Information Filter (EIF), and Rao-Blackwellized conditional particle filter (FastSLAM) belong to this category. These algorithms are of high computational complexity, can hardly be used in scenarios where the number of features is large. Conversely, smoothing approaches consist in the estimation of the offline problem, and typically produce a maximum a posterior probability (MAP) estimate by least square error minimization. Smoothing outputs more accurate estimates, but requires higher computational cost if there is a large number of frames. Moreover, both filtering and smoothing are sensitive to noise priors as they are fixed during updates. On the other hand, Lidar-based SLAM registers neighboring frames to predict an initial guess before optimization. Prevalent registration algorithms, Iterative Closest Point (ICP), and Normal Distributions Transform (NDT) are either computational demanding or sensitive to outliers, noise, and a portion of missing points. In recent years, variational Bayesian methods have shown great potential in numerous complex statistical models. Their fast speed, improved scalability, and accuracy motivate us to apply variational Bayesian methods to the SLAM community. In this thesis, we first develop online and offline SLAM algorithms based on stochastic variational Bayesian inference (SVBI), namely, online SVBI and offline SVBI. In both formulations, we model SLAM as a posterior estimation problem, and then employ a Gaussian distribution to approximate this posterior by maximizing the evidence lower bound (ELBO). We restrict the uncertainty matrix of the Gaussian distribution to be diagonal and tri-diagonal in online SVBI and offline SVBI, respectively. Consequently, both have linear computational complexity. In addition, we jointly update the noise priors together with the posterior. Therefore, online and offline SVBI are robust. The map variable is encoded into a state vector for online SVBI, however, shared across all frames in offline SVBI. We have evaluated online and offline SVBI on synthetic and real-life data, obtaining more consistent estimations. This type of SLAM has achieved a superior performance of a 7% error reduction than the traditional algorithms in both online and offline scenarios. Apart from the fundamental estimation, we propose two types of variational Bayesian point set registration for Lidar-based SLAM: variational Bayesian point registration based on Gaussian Mixture Models (GMMs-VB-PSR) and variational Bayesian point registration based on kernel density (KD-VB-PSR). The former is appropriate for clustering data, and the latter is widely applicable. The registration is described as a state space model (SSM), with the transition being the hidden variable. GMMs-VB-PSR and KD-VB-PSR represent the sets as Gaussian Mixture Models (GMMs) and Gaussian kernel densities, respectively. The L2 distance between two GMMs or densities is then interpreted as the ``observation''. We employ SVBI to solve the formulated SSM in both methods, resulting in linear complexity and robust performance. Furthermore, GMMs-VB-PSR and KD-VB-PSR adopt quadrature and randomized quasi-Monte Carlo (RQMC) points, respectively, to ease the computational burden. In addition, KD-VB-PSR relies on stochastic variance reduced gradient (SVRG) method and variance reduction techniques to improve the convergence speed further. The validity of GMMs-VB-PSR and KD-VB-PSR is demonstrated through extensive experimental results, illustrating the effectiveness, robustness, and efficiency. To summarize, this thesis demonstrates the superiority of the proposed variational Bayesian approaches for visual and Lidar-based SLAM problems, including online and offline SVBI, GMMs-VB-PSR, and KD-VB-PSR. The proposed method could be further employed in a variety of applications such as real robotic platforms, sensor data fusion, data association, and point shape completion.