Mahalanobis Distance

Unlike Euclidean distance, the Mahalanobis distance accounts for the scale and correlation of the data. It automatically adjusts for variables with different units and variances, and for relationships between variables.

• Example: In a 2D sensor measurement (e.g., position x, velocity v), if velocity is measured in m/s and has high variance, a Euclidean distance would be dominated by velocity differences. Mahalanobis distance uses the covariance matrix to normalize these differences, giving a true statistical distance.

For a point x from a multivariate normal distribution with mean μ and covariance Σ, the squared Mahalanobis distance follows a chi-squared distribution.

• The formula is: D² = (x - μ)ᵀ Σ⁻¹ (x - μ)
• If x is truly from the distribution, D² is distributed as χ² with k degrees of freedom, where k is the dimensionality.
• This provides a rigorous probabilistic threshold for outlier rejection (gating). A point with D² > χ²_k(0.95) can be rejected as an outlier with 95% confidence.

Points with the same Mahalanobis distance from the mean lie on an ellipsoid centered at the mean. The shape and orientation of this ellipsoid are defined by the covariance matrix Σ.

• The eigenvectors of Σ define the ellipsoid's axes.
• The eigenvalues determine the lengths of these axes.
• This visually represents the distribution's uncertainty: directions of high variance have longer ellipsoid axes, meaning points can be farther away in that direction while still being statistically 'close'.

In multi-sensor tracking and data association, Mahalanobis distance is the standard metric for validation gating. It determines if a new sensor measurement is a plausible match for an existing track.

• Process: A Kalman filter predicts a track's state (mean μ) and its uncertainty (covariance P). An incoming measurement is tested by computing its Mahalanobis distance to the predicted state.
• Gating: If the distance is below a χ² threshold, the measurement is considered a candidate for association. This efficiently prunes improbable matches before running more expensive association algorithms like the Global Nearest Neighbor.

The distance measure is inherently tied to the Gaussian assumption. It is most meaningful and optimal when the underlying data is multivariate normal.

• Limitation: For highly non-Gaussian or multi-modal distributions, the Mahalanobis distance can be misleading, as a single covariance matrix cannot capture complex shapes.
• Mitigation: In such cases, techniques like mixture models (e.g., Gaussian Mixture Models) are used, where distance is computed to the nearest mixture component, or non-parametric methods are employed.

The core computation requires inverting the covariance matrix Σ⁻¹, which has implications for stability and speed.

• Singularity: If Σ is singular (not invertible), due to collinear features or insufficient data, the distance is undefined. This is addressed via regularization (adding a small value to the diagonal) or using the pseudo-inverse.
• Efficiency: For real-time systems, the inverse is often pre-computed and updated recursively alongside the state estimate in filters like the Kalman Filter.
• Robustness: It is sensitive to outliers in the estimation of μ and Σ. Using robust covariance estimators (e.g., Minimum Covariance Determinant) can improve performance in contaminated data.

This is the most common application. Before a new sensor measurement is incorporated into a state estimate (e.g., in a Kalman filter), its Mahalanobis distance from the predicted measurement is calculated.

• A measurement is accepted if its distance falls within a probabilistic threshold (e.g., a chi-squared gate).
• Measurements with a large distance are statistically improbable given the current estimate and its uncertainty, and are rejected as outliers or noise.
• This prevents spurious data from corrupting the filter's state, which is crucial in cluttered environments like urban driving or dense sensor networks.

In Multi-Object Tracking (MOT), the system must decide which new sensor detection corresponds to which existing tracked object. The Mahalanobis distance provides a statistically sound metric for this assignment.

• For each track, a gating region is defined in measurement space based on the track's predicted position and covariance.
• Detections are associated with the track whose predicted measurement they are closest to in Mahalanobis distance.
• This method, often used in the Global Nearest Neighbor (GNN) algorithm, is more reliable than Euclidean distance because it accounts for the direction and shape of a track's uncertainty ellipse.

The Mahalanobis distance can monitor the consistency of a sensor's output relative to the fused system estimate or other sensors.

• A persistent, high Mahalanobis distance for measurements from a specific sensor indicates it may be faulty, misaligned, or experiencing interference.
• This forms the basis for Fault Detection and Isolation (FDI) subsystems. By comparing distances across sensors, the system can identify and potentially exclude the malfunctioning unit.
• Example: In an inertial navigation system, a sudden drift in the IMU's reported acceleration, measured via its distance from the GPS-corrected estimate, can trigger a fault alert.

In decentralized or distributed fusion architectures, different nodes maintain local estimates. Before fusing two estimates, their consistency must be checked.

• The Mahalanobis distance between the two state estimates (or their innovations) is computed.
• A small distance indicates the estimates are statistically consistent and can be safely fused (e.g., using Covariance Intersection).
• A large distance signals a fundamental disagreement, potentially due to unmodeled errors, data association mistakes, or faults, preventing an incorrect fusion that would reduce overall confidence.

Beyond filtering, the Mahalanobis distance defines statistically meaningful clusters in multi-dimensional feature spaces derived from sensor data.

• In perception systems, features (e.g., visual descriptors, lidar shape signatures) from objects can be modeled with a mean and covariance.
• New observations are classified based on their minimal Mahalanobis distance to these class distributions.
• This is effectively a Gaussian discriminant analysis, useful for terrain classification from multispectral data or material identification from hyperspectral sensors.

The core strength of the Mahalanobis distance (D_M) for a point (\mathbf{x}) relative to a distribution with mean (\mathbf{\mu}) and covariance (\mathbf{\Sigma}) is its scale and correlation invariance:

[ D_M(\mathbf{x}) = \sqrt{(\mathbf{x} - \mathbf{\mu})^T \mathbf{\Sigma}^{-1} (\mathbf{x} - \mathbf{\mu})} ]

• It measures distance in units of standard deviation, automatically weighting directions of high uncertainty less heavily.
• Points lying on the same probability density contour (ellipsoid) have the same Mahalanobis distance.
• This makes it the theoretically optimal distance metric for Gaussian-distributed errors, which is a foundational assumption in many sensor fusion algorithms like the Kalman filter.

A covariance matrix is a square matrix that quantifies the uncertainty of a state vector and the correlations between its components. It is the foundational element scaled by the Mahalanobis distance.

• Key Role: In the Mahalanobis formula, the inverse of the covariance matrix transforms and scales the Euclidean distance, accounting for variable correlations and different units of measurement.
• In Estimation: Algorithms like the Kalman filter propagate and update this matrix to represent the evolving uncertainty of the system state estimate.
• Practical Implication: A diagonal covariance matrix assumes uncorrelated variables, reducing the Mahalanobis distance to a normalized Euclidean distance.

A Kalman filter is a recursive algorithm for optimal state estimation in linear dynamic systems with Gaussian noise. It is a primary application context for Mahalanobis distance.

• Prediction & Update: The filter alternates between predicting the next state and updating it with new sensor measurements.
• Innovation Gate: The Mahalanobis distance is used to compute the innovation, the difference between predicted and actual measurements. This distance gates whether a measurement is likely a valid observation or an outlier.
• Assumptions: It assumes known linear process and observation models, and that all uncertainties are Gaussian, making the Mahalanobis distance the statistically correct metric for measurement validation.

Outlier rejection is the process of identifying and discarding sensor measurements that are statistically inconsistent with the current system model. Mahalanobis distance is a direct mechanism for this.

• Gating: A measurement is considered an outlier if its Mahalanobis distance from the predicted measurement distribution exceeds a threshold (e.g., based on the chi-squared distribution).
• Robustness: This is critical in sensor fusion where faulty sensors, environmental noise, or false detections (clutter) can corrupt the state estimate.
• Example: In multi-object tracking, gating with Mahalanobis distance prevents incorrect associations between tracks and detections.

State estimation is the process of inferring the hidden variables (state) of a dynamic system from a sequence of noisy sensor observations. Mahalanobis distance provides a probabilistic measure of fit.

• Core Problem: Determining position, velocity, orientation, etc., from imperfect sensors like IMUs, lidar, and cameras.
• Probabilistic Framework: Within Bayesian filtering, the state is represented as a probability distribution. The Mahalanobis distance measures how "far" a new observation is from the center of the predicted distribution.
• Algorithms: This underpins Kalman filters, particle filters, and graph-based SLAM, where it helps evaluate measurement likelihoods.

The multivariate normal distribution is the generalization of the one-dimensional normal distribution to multiple correlated variables. The Mahalanobis distance is intrinsically linked to its geometry.

• Mathematical Basis: The probability density function of this distribution depends on the Mahalanobis distance of a point from the mean.
• Contours of Constant Density: Points with the same Mahalanobis distance from the mean lie on an ellipse (or ellipsoid), not a circle. This accounts for scale and correlation.
• Assumption: Many sensor fusion algorithms assume measurement and process noise are multivariate normal, making Mahalanobis distance the natural metric for statistical tests.

Data association is the problem of correctly assigning incoming sensor detections to existing tracks or landmarks. Mahalanobis distance is a key metric for evaluating assignment hypotheses.

• Challenge: In cluttered environments, multiple detections may be plausible matches for a single track.
• Nearest Neighbor Standard Filter: Often selects the detection with the smallest Mahalanobis distance to the predicted track position.
• Complex Methods: More advanced algorithms like Multiple Hypothesis Tracking (MHT) or Joint Probabilistic Data Association (JPDA) use Mahalanobis distance within probabilistic frameworks to manage ambiguity over time.