SLAM (Simultaneous Localization and Mapping)

This is the perception layer of SLAM, responsible for ingesting raw sensor data and extracting meaningful, trackable information. The choice of sensor defines the approach.

• Visual SLAM (vSLAM): Uses cameras (monocular, stereo, RGB-D) to detect and track visual features like corners (e.g., using FAST, ORB) or describe entire regions.
• LiDAR SLAM: Uses laser scanners to generate precise 3D point clouds. Features are often geometric primitives like planes, lines, or clusters.
• Feature vs. Direct Methods: Feature-based methods extract and match sparse landmarks. Direct methods (like DSO) optimize geometry using the intensity of all pixels, avoiding feature extraction but requiring more computation.

The output is a stream of observations (feature positions, point clouds) for the back-end.

This is the brain of the SLAM system. It takes the noisy observations from the front-end and estimates the most probable trajectory (path) of the robot and the map of landmarks.

• Probabilistic Framework: SLAM is fundamentally an inference problem. The system maintains a probability distribution over possible robot poses and landmark locations.
• Graph-Based Optimization: Modern SLAM often formulates this as a pose graph. Nodes represent robot poses at different times and landmark positions. Edges represent constraints from sensor measurements. The system solves for the node configuration that minimizes the error across all constraints (bundle adjustment).
• Filters vs. Optimizers: Early systems used Kalman Filters (EKF-SLAM) for online estimation. Today, non-linear optimization libraries (like g2o, Ceres Solver) are standard for their superior accuracy in handling loops and large datasets.

This critical component identifies when the robot has returned to a previously visited location, correcting the accumulated drift in its estimated pose and map.

• The Problem of Drift: Small errors in odometry and sensor noise compound over time, causing the map to become increasingly inaccurate. A loop closure provides a strong constraint that 'pulls' the map and trajectory into consistency.
• Visual Bag-of-Words: A common method where images are converted into a visual vocabulary of features. Re-observing a similar combination of features triggers a loop closure hypothesis.
• Verification: Candidate loop closures must be geometrically verified (e.g., using PnP or point cloud alignment) to avoid false positives that could catastrophically corrupt the map.

This defines the form in which the constructed environment model is stored and used. The representation is tightly coupled with the sensor type and the application's needs.

• Metric Maps: Precise geometric representations.
  • Point Cloud Map: A dense set of 3D points (from LiDAR or RGB-D cameras).
  • Voxel Grid: A 3D volumetric grid where each voxel stores occupancy probability (OctoMap).
  • Surfel Map: Uses small surface elements (surfels) for efficient, dense surface representation.
• Topological Maps: A graph-based representation where nodes are distinct places (like rooms) and edges denote connectivity. Less precise but useful for high-level planning.
• Hybrid Maps: Combine metric detail in local areas with a topological graph for large-scale environments.

Provides a high-frequency, short-term estimate of the robot's motion, which is essential for local tracking and as an initial guess for the back-end optimizer.

• Wheel Odometry: Uses encoder counts on wheels to estimate displacement. Prone to slip and drift.
• Visual Odometry (VO) / Visual-Inertial Odometry (VIO): Estimates motion from camera images. VIO fuses camera data with an Inertial Measurement Unit (IMU). The IMU's high-frequency accelerometer and gyroscope data provide robustness to rapid motions and visual blur, while the camera corrects the IMU's long-term drift.
• LiDAR Odometry: Estimates motion by aligning consecutive LiDAR scans (using algorithms like Iterative Closest Point (ICP) or Normal Distributions Transform (NDT)). This component is sometimes considered part of the SLAM front-end.

A practical engineering mechanism to control computational complexity. A SLAM system cannot afford to optimize every single sensor frame.

• Keyframes: Are selectively chosen frames that represent a significant change in viewpoint. They become the nodes in the pose graph. Frames between keyframes are discarded after their data is used for tracking.
• Selection Criteria: A new keyframe is typically created when:
  • The robot has moved a minimum distance or rotational angle from the last keyframe.
  • A significant number of new map landmarks become visible.
  • Tracking quality drops, indicating a need for a new reference.
• Landmark Sparsification: Similarly, only a manageable subset of stable, high-quality features are added to the map as persistent landmarks for long-term optimization. This prevents the optimization problem from growing unbounded.

Visual Odometry is the process of estimating an agent's ego-motion (position and orientation) by analyzing a sequence of camera images. It is a core component of visual SLAM systems but focuses solely on tracking the camera's path, often without building a persistent global map.

• Key Difference from SLAM: VO is primarily concerned with local, incremental pose estimation. Full SLAM systems integrate VO with loop closure and global map optimization to correct accumulated drift.
• Sensor Input: Typically uses monocular, stereo, or RGB-D cameras.
• Applications: Essential for drones, augmented reality, and any system requiring real-time motion tracking from visual data.

Occupancy Grid Mapping is a probabilistic approach to representing an environment as a discrete grid, where each cell stores the probability that it is occupied by an obstacle. This is a common map representation used in many SLAM implementations, particularly for 2D navigation.

• Representation: The environment is discretized into cells (e.g., 5cm x 5cm). Each cell holds a value between 0 (free) and 1 (occupied).
• Sensor Fusion: Algorithms like the log-odds update rule integrate successive sensor readings (e.g., from LiDAR) to build a consistent map.
• Use Case: The foundational map for robot path planning and collision avoidance in warehouse robots and autonomous vehicles.

Loop Closure Detection is the critical process in SLAM of recognizing when an agent has returned to a previously visited location. This corrects the accumulated drift in the estimated trajectory and map, leading to a globally consistent model.

• Mechanism: Compares current sensor observations (e.g., visual features, LiDAR point clouds) against a database of past observations to find a match.
• Algorithms: Often uses bag-of-words models for visual features or scan matching for LiDAR.
• Impact: Successful loop closure triggers pose graph optimization, which distributes the correction error across the entire estimated path.

Pose Graph Optimization is a backend optimization technique in SLAM that refines the estimated poses (positions and orientations) of the robot over time. It represents the trajectory as a graph of nodes (poses) connected by edges (constraints from odometry or loop closures).

• Graph Structure: Nodes are robot poses. Edges represent relative transformations between poses, with an associated uncertainty (covariance).
• Optimization Goal: Minimize the error between the predicted relative transformations (from the graph) and the measured ones (from sensors).
• Solver: Typically solved using non-linear least squares libraries like g2o or Ceres Solver. This is the 'Smoothing' part of algorithms like GraphSLAM.

Feature-Based SLAM is a class of SLAM algorithms that rely on detecting, describing, and matching distinctive keypoints (features) in sensor data, typically from cameras. ORB-SLAM is a prominent, open-source example of this paradigm.

• Pipeline: Extracts features (e.g., ORB, SIFT) from images, matches them across frames to track the camera and map 3D points, and performs bundle adjustment.
• Advantages: Computationally efficient, provides sparse but accurate maps suitable for localization.
• Sparsity vs. Density: Contrasts with dense or direct SLAM methods (like DTAM or DVO) that use pixel intensity information directly, yielding denser maps at higher computational cost.

LiDAR SLAM uses Light Detection and Ranging sensors as the primary perception modality. It is the dominant approach for high-precision mapping and localization in autonomous vehicles and large-scale outdoor robotics.

• Core Algorithm: Iterative Closest Point (ICP) is a fundamental algorithm for aligning two LiDAR point clouds to estimate motion.
• Modern Methods: LOAM (Lidar Odometry and Mapping) and its successors (LeGO-LOAM, LIO-SAM) are state-of-the-art pipelines that combine feature extraction from point clouds with inertial measurement unit (IMU) data for robust odometry.
• Output: Creates highly accurate 3D point cloud maps of the environment, which can be converted into occupancy grids or semantic maps.