Robust Estimation

Least Median of Squares is a robust regression method that minimizes the median of the squared residuals, rather than their sum. Its breakdown point is 50%, meaning it can tolerate up to half of the data points being outliers before producing arbitrarily wrong results.

Algorithm Process:

• Randomly sample subsets of points to generate candidate models.
• For each model, compute the median of the squared residuals for all data points.
• Select the model with the smallest median squared residual.

While computationally intensive, LMedS provides a very high degree of robustness. It is often used as a benchmark or in conjunction with a final M-estimator step for refined parameter estimation after outlier removal.

The Huber Filter is a robust variant of the Kalman filter that replaces the quadratic measurement update with the Huber loss function. This modification makes the filter resilient to measurement outliers while maintaining computational efficiency close to the standard Kalman filter.

How it works:

• The prediction step remains identical to the Kalman filter.
• The update step uses the Huber penalty, which applies a linear cost to large innovations (residuals), preventing a single bad measurement from disproportionately distorting the state estimate.

This filter is particularly valuable in visual-inertial odometry (VIO) and GPS/INS fusion, where sporadic sensor failures or multipath errors can introduce severe, non-Gaussian noise.

These methods dynamically adjust the estimation strategy based on inferred data quality or system mode.

Key Approaches:

• Interacting Multiple Model (IMM): Runs a bank of parallel filters (e.g., one standard, one robust) and mixes their estimates based on model probabilities, adapting to periods of high outlier presence.
• Adaptive Thresholding: Algorithms like RANSAC can use Median Absolute Deviation (MAD) to dynamically set the inlier/outlier threshold based on residual statistics.
• Fault Detection and Isolation (FDI): Monitors innovation sequences or residual norms to detect sensor faults, triggering a switch to a degraded-mode estimator that ignores the faulty sensor.

These are critical for long-duration autonomy where sensor and environmental conditions are non-stationary.

In modern factor graph formulations for SLAM and sensor fusion, robustness is incorporated directly into the optimization's cost function.

Implementation Techniques:

• Robust Kernels (Huber, Cauchy): Applied to factors (edges) in the graph. A kernel function wraps the error term, down-weighting the influence of large errors during nonlinear least squares optimization (e.g., in g2o or GTSAM libraries).
• Switchable Constraints: Introduce binary or continuous switch variables into the graph that can effectively disable a factor if its associated measurement is deemed an outlier.
• Degeneracy-Aware Factors: Use DCS or Geman-McClure kernels that are more aggressive than Huber at rejecting outliers, which is essential for robust loop closure in SLAM.

This approach is the backbone of robust, large-scale pose graph optimization.

Bayesian filtering is the foundational probabilistic framework for recursively estimating the state of a dynamic system from noisy sensor data. It operates on the principle of maintaining a belief state—a probability distribution over possible states—and updating it with new measurements using Bayes' theorem.

• Core Concept: The filter predicts the state forward in time using a process model, then corrects this prediction using a sensor model when a new observation arrives.
• Implementations: The Kalman filter, particle filter, and histogram filter are all specific realizations of this general framework, each making different assumptions about system linearity and noise distributions.
• Role in Robustness: This framework inherently models uncertainty via probability distributions, providing a principled way to handle sensor noise. Robust estimators often modify the Bayesian update step to reduce the influence of outliers.

An M-estimator is a broad class of robust estimators that generalize maximum likelihood estimation by replacing the squared-error loss function with a more robust alternative. The 'M' stands for 'maximum likelihood-type'.

• Mechanism: Instead of minimizing the sum of squared residuals (which is highly sensitive to outliers), an M-estimator minimizes the sum of a different function, ρ, of the residuals: ∑ ρ(r_i). Common robust ρ-functions include Huber and Tukey's biweight.
• Key Property: These functions increase less rapidly than the quadratic function for large residuals, thereby down-weighting the influence of outliers on the final estimate.
• Application in Fusion: Used within optimization-based sensor fusion frameworks (like graph-based SLAM) to make the cost function resilient to erroneous loop closures or bad feature correspondences.

RANSAC is an iterative, non-deterministic algorithm used to estimate the parameters of a mathematical model from a dataset containing a large proportion of outliers. It is a cornerstone of robust model fitting in computer vision and sensor data alignment.

• Algorithm Steps: 1) Randomly select a minimal sample set needed to fit the model. 2) Fit a model to this sample. 3) Count the number of inliers (data points within a error tolerance of the model). 4) Repeat for many iterations and keep the model with the largest inlier set.
• Use Cases: Essential for tasks like point cloud registration (e.g., initial alignment for ICP), camera pose estimation from noisy feature matches, and planar surface detection in lidar scans.
• Robustness: Its strength lies in not trying to refine a model from all data points at once, but in searching for a consensus among a subset of presumably correct data.

Fault Detection and Isolation is a systematic process within a sensor fusion system to identify when a sensor is providing erroneous data (detection) and to determine which specific sensor is faulty (isolation). It is a higher-level architectural approach to robustness.

• Methods: Techniques include parity space methods, observer-based methods, and statistical testing (e.g., on innovation sequences in a Kalman filter).
• Isolation: Once a fault is detected, the system must identify the culprit sensor to exclude its data from the fusion process, preventing corruption of the overall state estimate.
• Relationship to Robust Estimation: While robust estimators like M-estimators implicitly handle sporadic outliers, FDI systems are designed to handle persistent sensor failures or biases, often triggering a system reconfiguration.

The covariance matrix and Mahalanobis distance are fundamental tools for representing uncertainty and detecting outliers in multivariate estimation problems, forming the basis for many robust data association techniques.

• Covariance Matrix (Σ): In state estimation, this matrix quantifies the estimated uncertainty of the state vector and the correlations between state variables. It is propagated through the process model and updated during measurement steps.
• Mahalanobis Distance: A statistical distance defined as d_M(x, μ) = √[(x - μ)^T Σ^{-1} (x - μ)]. It measures how many standard deviations a point x is from a distribution with mean μ and covariance Σ.
• Robust Application: This distance is used for gating in tracking algorithms. Potential sensor measurements are only considered for data association if their Mahalanobis distance to a predicted track is below a threshold, effectively filtering out statistically improbable (and likely erroneous) measurements.

The Interacting Multiple Model algorithm is an adaptive estimation technique designed for systems whose behavior or dynamics can switch between different modes. It provides robustness to model uncertainty, not just measurement outliers.

• Architecture: IMM runs a bank of parallel filters (e.g., Kalman filters), each tuned to a different process model (e.g., constant velocity, coordinated turn).
• Interaction: The estimates and covariances from these filters are mixed at each time step based on the estimated probability of each model being correct.
• Robustness Value: It allows a tracking or fusion system to remain accurate when an object's motion pattern changes unexpectedly. For example, it can seamlessly handle a tracked vehicle transitioning from straight-line driving to making a turn, where a single fixed model would fail, leading to large estimation errors.