Visual-Inertial Odometry (VIO)

VIO exploits the complementary strengths of a camera and an Inertial Measurement Unit (IMU). The camera provides rich, absolute orientation and scale information but suffers during motion blur, low texture, or rapid motion. The IMU provides high-frequency, metric-scale acceleration and angular velocity measurements that are immune to visual conditions but drift rapidly due to integration error. By fusing them, VIO maintains accuracy when either sensor alone would fail.

High-performance VIO systems typically employ a tightly-coupled fusion approach. This means raw or pre-processed sensor data (e.g., IMU pre-integration factors and visual feature reprojection errors) are combined within a single probabilistic optimization framework, such as a factor graph. This is contrasted with loosely-coupled fusion, which would fuse separate visual odometry and inertial navigation poses. Tight coupling provides higher accuracy by directly modeling correlations and constraints between all sensor measurements.

A key benefit of VIO is its improved observability compared to a standalone INS. While an INS alone cannot observe its global orientation (yaw) or position without external aid, the camera provides periodic absolute measurements (from feature parallax) that make these states observable. This allows the estimator to correct for the IMU's inherent drift in position, velocity, and attitude, especially the unbounded drift in yaw, without requiring GPS or magnetometers.

VIO is engineered for robustness in challenging visual environments. The IMU acts as a high-frequency motion prior, enabling the system to:

• Bridge temporary visual outages caused by sudden lighting changes, occlusions, or pure white walls.
• Reject visual outliers more effectively by checking feature motion consistency against the inertial prediction.
• Handle aggressive motion and blur, as the IMU provides a reliable short-term motion estimate even when feature tracking fails. This makes VIO suitable for drones, AR/VR headsets, and legged robots operating dynamically.

Pure monocular visual odometry can only recover trajectory and map structure up to an unknown scale factor. VIO solves this scale ambiguity intrinsically. The IMU's accelerometers measure specific force in meters per second squared, providing an absolute metric reference. By fusing accelerometer data with visual observations of feature motion, the estimator can directly infer the true metric scale of the environment, which is critical for physical robot interaction and control.

VIO is fundamentally a recursive state estimation problem, often solved using filtering (e.g., Extended Kalman Filter, Error State Kalman Filter) or fixed-lag smoothing. It maintains a running estimate of the robot's state (pose, velocity, sensor biases) and updates it at the rate of the fastest sensor (typically the IMU at 100-1000 Hz), with visual updates incorporated asynchronously. This provides a low-latency, continuous pose output essential for closed-loop robotic control and augmented reality.

The algorithmic process of combining data from multiple disparate sensors to produce a more accurate, complete, and reliable estimate of a system's state than is possible from any single sensor. In VIO, this is the overarching principle that governs the combination of camera and IMU data. Key architectures include:

• Tightly-Coupled Fusion: Raw sensor data (e.g., pixel features, raw IMU readings) are fused in a single estimator.
• Loosely-Coupled Fusion: Independent estimates from each sensor (e.g., a pose from visual odometry) are fused at the state level.

The process of inferring the internal, often unmeasured, state variables of a dynamic system from a sequence of noisy sensor measurements. For a mobile robot, the state typically includes position, orientation (attitude), and velocity. VIO is a specific state estimation technique. The problem is formalized within the Bayesian filtering framework, where the goal is to compute the posterior probability distribution of the state given all measurements.

An optimal recursive algorithm for state estimation in linear dynamic systems with Gaussian noise. It operates in a two-step cycle:

1. Prediction: Projects the current state and its uncertainty forward using a process model.
2. Update (Correction): Refines the prediction using a new sensor measurement and a measurement model. The filter maintains a covariance matrix representing state uncertainty. The Extended Kalman Filter (EKF) linearizes nonlinear models to apply this framework, forming the basis for many early VIO systems.

A bipartite graphical model used to represent the factorization of a complex probability distribution. In robotics, it is the dominant paradigm for structuring large-scale, nonlinear optimization problems like Simultaneous Localization and Mapping (SLAM) and modern VIO. Factors represent probabilistic constraints (from measurements or motion models) connecting variable nodes (states to be estimated). Solving the graph via Maximum A Posteriori (MAP) estimation provides the optimal state trajectory. This approach is more flexible than filter-based methods for incorporating loop closures.

A self-contained navigation system that uses accelerometers and gyroscopes (an IMU) to track a vehicle's position, orientation, and velocity through dead reckoning. By integrating acceleration to get velocity, and integrating velocity to get position, INS provides high-frequency, short-term motion data. However, it suffers from unbounded drift due to the integration of sensor bias and noise. VIO uses the camera to provide absolute visual constraints that correct this IMU drift, creating a far more stable system.

The computational problem of constructing or updating a map of an unknown environment while simultaneously tracking the agent's location within it. VIO is often considered the front-end of a visual SLAM system, providing robust, high-frequency odometry. The full SLAM back-end performs global optimization (e.g., bundle adjustment) over the entire trajectory and map when loop closures are detected, correcting accumulated drift that VIO alone cannot fix.