Source-linked AI summary
Extended Object Tracking with Random Hypersurface Models
Marcus Baum, Uwe D. Hanebeck
TL;DR
Extended-object tracking must estimate shape as well as kinematics because measurements arise from multiple spatially distributed sources on the target. The paper introduces the Random Hypersurface Model, which uses randomly scaled shape boundaries to relate measurements to shape parameters and support Gaussian state estimation. It derives estimators for elliptic and star-convex shapes, with detailed star-convex estimation requiring sufficiently informative measurements.
Problem
Extended-object tracking needs to estimate shape in addition to kinematic state because measurement sources are spatially distributed and depend on object geometry.
Method
The Random Hypersurface Model represents measurement sources on randomly scaled versions of a shape boundary, yielding a measurement equation for Gaussian state estimators.
Results
Specific RHMs and Gaussian estimators are developed for elliptic and free-form star-convex shapes, and the shape is tracked precisely even when its orientation changes.
Takeaways & Limitations
Free-form star-convex shape estimation supports detailed shape approximation and can pave the way for applications such as shape-based classification and group splitting detection.
Takeaways & Limitations
Detailed star-convex shape estimation requires measurement noise low relative to target extent and enough measurements per time step; otherwise, an ellipse is more suitable.
Abstract
from arXiv · showhide
The Random Hypersurface Model (RHM) is introduced that allows for estimating a shape approximation of an extended object in addition to its kinematic state. An RHM represents the spatial extent by means of randomly scaled versions of the shape boundary. In doing so, the shape parameters and the measurements are related via a measurement equation that serves as the basis for a Gaussian state estimator. Specific estimators are derived for elliptic and star-convex shapes.
1. Introduction
Extended-object tracking must estimate both kinematic state and shape because one scan can resolve multiple measurement sources whose locations depend on object geometry. The RHM provides a systematic model linking these sources to shape parameters for Gaussian estimation, with ellipse and free-form star-convex implementations.
- Motivation: Multiple spatially distributed measurement sources can be resolved in one scan, so the point-object assumption is not justified.Their locations vary across scans and depend on object shape, surface properties, and target-to-sensor geometry.
- Motivation: Extended-object tracking estimates shape parameters alongside the object's kinematic parameters.The paper illustrates geometric shape approximation using an ellipse.
- Related work: Spatial-distribution likelihoods generally lack closed-form solutions, motivating frequent use of Monte Carlo approximations for Bayesian filtering.For elliptic extent represented by a covariance matrix, closed-form expressions can be derived using random matrix theory.
- Contributions: The RHM models an unknown measurement-source location on a randomly scaled shape boundary and supplies a measurement equation for Gaussian state estimation.Specific RHMs and Gaussian estimators are developed for ellipses and free-form star-convex shapes.
- Article overview: The paper evaluates the proposed shape representations in typical extended-object tracking scenarios.The article develops the probabilistic framework, RHM, Bayes filter, shape-specific implementations, and evaluations in sequence.
2. Modeling Extended Targets
The model represents an extended target through a state containing location, shape, and optional kinematic variables, while separating extent, sensor, and temporal-evolution models. Measurements are conditionally independent given the target state and are modeled as noisy Cartesian point observations.
- State representation: The extended-object state contains target location, a shape-parameter vector, and optional kinematic variables such as velocity.The shape parameters specify a two-dimensional set S(p_k).
- Shape representation: The framework is presented for two-dimensional shapes, although the concepts are also applicable to higher-dimensional shapes.The paper gives circular shape specification by its radius as an example.
- Measurement model: At each time step, the sensor provides n_k two-dimensional point measurements that are assumed mutually independent given the target state.This makes a single-measurement model sufficient.
- Measurement model: The extent model places each measurement source within the target location translated by its shape set, m_k + S(p_k).The sensor model then maps the source to a Cartesian measurement corrupted by additive Gaussian noise.
- Dynamic model: The target's shape and kinematic parameters evolve over time under linear motion models with system matrix A_k and white Gaussian system noise.Extended objects require temporal models for both shape and kinematic parameters, unlike point targets.
3. Random Hypersurface Models
The RHM models measurement sources inside star-convex extended objects as points on randomly scaled shape boundaries, yielding an implicit measurement relation for Gaussian estimation. The section derives scaling-factor distributions and a general procedure supporting elliptic and star-convex estimators.
- Random Hypersurface Model: The RHM represents a measurement source as belonging to a scaled version of the shape boundary, with the scaling factor modeled as a one-dimensional random variable.This covers interior sources while retaining a boundary-based implicit representation.
- Implicit Measurement Equation: The implicit relation for the scaled boundary, together with the extent model, forms an implicit measurement equation relating measurement sources to shape parameters.At scaling factor 1, the scaled-boundary function reduces to the original boundary function.
- Random Hypersurface Model: For star-convex shapes, scaling the boundary corresponds to a straight-line homotopy from the object center to its boundary.The star-convexity restriction ensures the scaled boundaries represent the region.
- Probability Distribution of the Scaling Factor: For uniformly distributed measurement sources in a two-dimensional star-convex region, the squared scaling factor is uniformly distributed on [0, 1].The result follows by relating the cumulative distribution to the area of the scaled region.
- General Procedure: The RHM procedure selects a shape parameterization, forms the implicit equation, and derives a state estimator from the resulting measurement equation.The paper applies this procedure to Gaussian estimators for ellipses and free-form star-convex shapes.
4. Formal Gaussian State Estimator for Extended Objects
The formal Gaussian Bayes filter updates an extended-object state through prediction and recursive measurement incorporation. Linear system models permit Kalman time updates, while independent measurements are assimilated using single-measurement likelihoods.
- Gaussian State Representation: The filter approximates probability densities for the extended-object parameter vector with Gaussians characterized by a mean and covariance matrix.The density after each incorporated measurement is indexed by the number of measurements processed.
- Time Update: The time update predicts the previous filtering density to the next time step using Kalman filter formulas for linear system models.This step determines the predicted density before measurements at the new time step are incorporated.
- Measurement Update: The measurement update incorporates the measurement set recursively according to Bayes’ rule.Each update uses a single-measurement likelihood and a normalization factor.
- Measurement Update: The order of measurements is irrelevant under the independent measurement-generation assumption, although approximation procedures can make processing order matter.The distinction applies to a particular time step.
5. Elliptic Shapes
The paper derives an implicit measurement equation for elliptic shapes under the Random Hypersurface Model and develops Gaussian-state estimation procedures for shape and kinematic parameters.
- Ellipses are modeled by a center and a positive semi-definite shape matrix, with a vectorized parameterization derived from its Cholesky decomposition.
- Equations (8) and (1) jointly formulate ellipse fitting as an implicit measurement problem with noisy data and a random scaling factor.
- The proposed estimator algebraically reformulates the implicit equation and statistically linearizes it around the measurement source rather than directly linearizing around the measurement and state.
- The unknown measurement source is replaced by a point estimate obtained from the previous mean ellipse and the closest point on its conic to the measurement.
- The resulting measurement equation can be processed with Gaussian estimators such as the UKF or analytic moment calculation, while measurement-order differences are observed to be negligible.
- Statistical linearization introduces approximation errors, making estimation quality dependent on ellipse parameterization and the particular reformulated measurement function.
6. Star-Convex Shapes
For low-noise measurements, the paper extends the RHM to star-convex shapes, using Fourier-parameterized radius functions to obtain a detailed shape approximation and an implicit measurement equation.
- The detailed shape approximation is motivated by applications including target classification, track management, and sensor management when measurement noise is low relative to target extent.
- Star-convex shapes are represented by a radius function giving the center-to-contour distance as a function of angle and shape parameters.
- The radius function is parameterized by Fourier coefficients, with low-index terms encoding coarse shape features and high-index terms encoding finer details.
- The star-convex boundary representation and the RHM scaling relation together define an implicit measurement equation.
- To avoid uncertain-angle treatment, occurrences of the radius function are evaluated at a point angle estimate, with the most likely angle available under isotropic measurement noise.
- The simplified measurement equation maps the star-convex state, measurement noise, scaling factor, and measurement to a zero pseudo-measurement.
7. Evaluation
The RHM estimators were evaluated on stationary and moving extended objects using elliptic and star-convex shape models. Shape estimates tracked moving targets precisely, but detailed star-convex recovery required sufficiently informative measurements.
- Evaluation design: The evaluation used RHM estimators for elliptic and star-convex shapes on stationary and moving extended objects.Both shape models used the UKF for measurement updates.
- Stationary extended target: Stationary-target simulations received 300 sequential measurements, with shape estimates averaged over 20 Monte-Carlo runs.Measurements were incorporated recursively, while plotted measurements from one run were included for visualization.
- Moving extended object: The moving-target scenario used an aircraft-shaped target with uniformly sampled surface measurements and measurement noise varying across observations.The target followed the trajectory shown in Fig. 9.
- Moving extended object: The estimated shapes, averaged over 20 time steps, tracked the moving extended object precisely even when its orientation changed.Results are shown for two trajectory snippets in the elliptic and star-convex cases.
- Evaluation limitations: A detailed star-convex approximation requires enough measurements with rather low noise, whereas an ellipse remains more suitable when information is limited.The evaluation used fewer measurements per time step for elliptic shapes than for star-convex shapes.
8. Conclusions and Future Work
The paper concludes that RHM provides measurement equations linking distributed measurements to shape parameters, enabling Gaussian estimation for elliptic and free-form star-convex shapes. Detailed star-convex recovery is information-demanding, motivating adaptive shape-model complexity.
- Conclusions: RHM models measurement sources on randomly scaled shape boundaries and provides a functional relationship between measurements and shape parameters.This relationship supports shape approximation alongside extended-object tracking.
- Conclusions: Measurement equations derived for elliptic and free-form star-convex shapes allow standard Gaussian state estimators to be used.The paper presents particular RHMs for both shape classes.
- Future work: Detailed star-convex shape estimation is possible only with relatively low measurement noise and enough measurements per time step.When these conditions are absent, the paper states that an ellipse is more suitable.
- Future work: The paper identifies adaptive complexity in shape descriptions as desirable and links free-form shape estimation to possible classification and group-splitting applications.These applications are presented as directions enabled by estimating free-form shapes.