Direct error formulation for real-time SLAM applications across diverse sensors

This thesis investigates direct error minimization methods for real-time Simultaneous Localization and Mapping (SLAM), a fundamental problem in robotics and autonomous systems. SLAM enables sensor-equipped systems to construct representations of unknown environments while simultaneously estimating their own pose, relying exclusively on onboard sensory data. Traditional SLAM approaches often depend on explicit data association and indirect optimization techniques; in contrast, direct methods optimize error functions defined directly on sensor measurements, improving robustness and versatility across various sensor modalities. Our research begins with the formulation of a direct error minimization pipeline for LiDAR, as realized in the MD-SLAM framework. This approach highlights the key advantage of direct methods, namely, the avoidance of explicit data association, and its modular design permits straightforward adaptation to additional sensor types such as RGB-D cameras. To further support the development and benchmarking of state estimation algorithms, we present a novel dataset featuring high-precision ground truth derived from a sensor fusion pipeline that combines GNSS-inertial measurements with multi-view LiDAR optimization. Within this framework, the LiDAR optimization problem minimizes a composite error function, integrating both geometric and photometric (direct) terms. The use of bilinear interpolation in the projection function enables sub-pixel accuracy in the alignment process. Expanding this paradigm to radar sensors, we formalize a direct approach to radar odometry based on correlation maximization between a local map and incoming radar scans for ego-motion estimation. While the field of radar odometry has historically been dominated by indirect approaches, considered faster and more amenable to real-time operation, our work demonstrates that direct methods can achieve competitive performance when equipped with a GPU-parallelized implementation. This enables real-time operation without sacrificing the accuracy and versatility inherent to direct approaches, setting a new state of the art in radar odometry. Collectively, this thesis demonstrates that direct error formulation is a powerful and adaptable framework for SLAM across diverse sensing modalities, providing advancements in accuracy, versatility and robustness, and, when accelerated with GPU computation, can achieve real-time capability.