"""
Joint Probabilistic Data Association (JPDA) algorithm.
JPDA computes association probabilities between tracks and measurements
by considering all possible joint association hypotheses, then performs
a probabilistically weighted update for each track.
This is more sophisticated than GNN which makes hard assignment decisions,
as JPDA can handle measurement origin uncertainty in cluttered environments.
"""
import warnings
from typing import Any, List, NamedTuple, Optional
import numpy as np
from numba import njit
from numpy.linalg import LinAlgError
from numpy.typing import ArrayLike, NDArray
from scipy.linalg import cho_factor, cho_solve
from scipy.stats import chi2
from pytcl.assignment_algorithms.gating import mahalanobis_batch, mahalanobis_distance
from pytcl.diagnostics import diagnostics_enabled, logger
# An innovation covariance is symmetric by construction; its asymmetry is
# float roundoff (~1e-16 relative), so sqrt(eps) sits ~8 orders above that.
_SYMMETRY_REL_TOL = float(np.sqrt(np.finfo(np.float64).eps))
def _is_finite_symmetric(S: NDArray[Any]) -> bool:
# Cholesky reads one triangle only, and which one differs between
# scipy.linalg.cho_factor (upper) and np.linalg.cholesky (lower), so an
# asymmetric S would be tested as two different matrices.
if not np.all(np.isfinite(S)):
return False
return bool(
np.max(np.abs(S - S.T), initial=0.0)
<= _SYMMETRY_REL_TOL * np.max(np.abs(S), initial=0.0)
)
[docs]
class JPDAResult(NamedTuple):
"""Result of JPDA algorithm.
Attributes
----------
association_probs : ndarray[Any]
Association probability matrix of shape (n_tracks, n_measurements + 1).
association_probs[i, j] is the probability that track i is associated
with measurement j. The last column (j = n_measurements) represents
the probability that track i has no measurement.
marginal_probs : list of ndarray
List of marginal association probabilities for each track.
marginal_probs[i][j] = P(measurement j originated from track i).
likelihood_matrix : ndarray[Any]
Measurement likelihood matrix of shape (n_tracks, n_measurements).
gated : ndarray[Any]
Boolean matrix indicating which track-measurement pairs passed gating.
"""
association_probs: NDArray[np.floating]
marginal_probs: List[NDArray[np.floating]]
likelihood_matrix: NDArray[np.floating]
gated: NDArray[np.bool_]
[docs]
class JPDAUpdate(NamedTuple):
"""Result of JPDA-based track update.
Attributes
----------
states : list of ndarray
Updated state estimates for each track.
covariances : list of ndarray
Updated covariances for each track (includes spread of means).
association_probs : ndarray[Any]
Association probability matrix.
innovations : list of ndarray
Combined weighted innovations for each track.
"""
states: List[NDArray[np.floating]]
covariances: List[NDArray[np.floating]]
association_probs: NDArray[np.floating]
innovations: List[NDArray[np.floating]]
[docs]
def compute_measurement_likelihood(
innovation: NDArray[Any],
innovation_cov: NDArray[Any],
detection_prob: float = 1.0,
) -> float:
"""
Compute measurement likelihood for a track-measurement pair.
Parameters
----------
innovation : ndarray[Any]
Measurement innovation (residual), shape (m,).
innovation_cov : ndarray[Any]
Innovation covariance, shape (m, m).
detection_prob : float
Probability of detection (Pd).
Returns
-------
likelihood : float
Measurement likelihood.
"""
m = len(innovation)
# det(S) > 0 alone does not imply positive-definiteness: an even number
# of negative eigenvalues also multiplies to a positive determinant
# (e.g. diag(-1, -1, 1) has det = 1), so a determinant-only test admits
# negative-definite and indefinite S. Cholesky fails whenever any
# leading principal minor is non-positive, which is the actual PD test.
if not _is_finite_symmetric(innovation_cov):
warnings.warn(
"compute_measurement_likelihood: innovation covariance is not "
"symmetric or contains non-finite values; likelihood set to 0.0 "
"(numerical failure, not evidence). Check R and the covariance "
"conditioning.",
RuntimeWarning,
stacklevel=2,
)
return 0.0
try:
S_cho = cho_factor(innovation_cov, check_finite=False)
except LinAlgError:
warnings.warn(
"compute_measurement_likelihood: innovation covariance is not "
"positive definite; likelihood set to 0.0 (numerical failure, "
"not evidence). Check R and the covariance conditioning.",
RuntimeWarning,
stacklevel=2,
)
return 0.0
mahal_sq = innovation @ cho_solve(S_cho, innovation, check_finite=False)
log_det_S = 2 * np.sum(np.log(np.diag(S_cho[0])))
likelihood = detection_prob * np.exp(
-0.5 * (mahal_sq + log_det_S + m * np.log(2 * np.pi))
)
return likelihood
[docs]
def compute_likelihood_matrix(
track_states: list[NDArray[Any]],
track_covariances: list[NDArray[Any]],
measurements: NDArray[Any],
H: NDArray[Any],
R: NDArray[Any],
detection_prob: float = 1.0,
gate_threshold: Optional[float] = None,
) -> tuple[NDArray[Any], NDArray[Any]]:
"""
Compute likelihood matrix for all track-measurement pairs.
Parameters
----------
track_states : list of ndarray
State estimates for each track.
track_covariances : list of ndarray
Covariances for each track.
measurements : ndarray[Any]
Measurements, shape (n_meas, m).
H : ndarray[Any]
Measurement matrix, shape (m, n).
R : ndarray[Any]
Measurement noise covariance, shape (m, m).
detection_prob : float
Probability of detection.
gate_threshold : float, optional
Mahalanobis distance threshold for gating.
Returns
-------
likelihood_matrix : ndarray[Any]
Likelihood values, shape (n_tracks, n_meas).
gated : ndarray[Any]
Boolean gating matrix, shape (n_tracks, n_meas).
Examples
--------
>>> import numpy as np
>>> # Two tracks, two measurements
>>> states = [np.array([0.0, 1.0]), np.array([5.0, 0.0])]
>>> covs = [np.eye(2) * 0.5, np.eye(2) * 0.5]
>>> measurements = np.array([[0.1], [5.2]])
>>> H = np.array([[1, 0]])
>>> R = np.array([[0.1]])
>>> L, gated = compute_likelihood_matrix(states, covs, measurements, H, R)
>>> L.shape
(2, 2)
>>> # Track 0 should have high likelihood for measurement 0
>>> L[0, 0] > L[0, 1]
True
"""
n_tracks = len(track_states)
# Coerce eagerly: the per-track fast path below does
# `measurements - z_pred` unconditionally, which raises for a bare
# Python list (numpy's binary-op fallback converts `[]` to a shape-(0,)
# array, not (0, m), so it fails to broadcast against z_pred's shape
# (m,) for m != 1). The old per-pair loop only ever did `measurements[j]`
# inside `for j in range(n_meas)`, so n_meas == 0 skipped touching
# `measurements` at all and a bare `[]` worked by accident.
measurements = np.asarray(measurements, dtype=np.float64)
n_meas = len(measurements)
likelihood_matrix = np.zeros((n_tracks, n_meas))
gated = np.zeros((n_tracks, n_meas), dtype=bool)
if n_meas == 0:
return likelihood_matrix, gated
for i in range(n_tracks):
# Predicted measurement and innovation covariance
z_pred = H @ track_states[i]
S = H @ track_covariances[i] @ H.T + R
m = S.shape[0]
# det(S) > 0 does not imply positive-definiteness (an even number of
# negative eigenvalues also multiplies to a positive determinant),
# so gate on Cholesky success rather than the sign of det(S). This
# row's likelihoods and gating stay at their zero/False
# initialization; other tracks' rows are untouched.
if not _is_finite_symmetric(S):
warnings.warn(
f"compute_likelihood_matrix: innovation covariance for track "
f"{i} is not symmetric or contains non-finite values; that "
"track's likelihoods and gating are set to 0.0 (numerical "
"failure, not evidence). Check R and the covariance "
"conditioning.",
RuntimeWarning,
stacklevel=2,
)
continue
try:
np.linalg.cholesky(S)
except LinAlgError:
warnings.warn(
f"compute_likelihood_matrix: innovation covariance for track "
f"{i} is not positive definite; that track's likelihoods and "
"gating are set to 0.0 (numerical failure, not evidence). "
"Check R and the covariance conditioning.",
RuntimeWarning,
stacklevel=2,
)
continue
# S depends only on the track, not the measurement: invert it once
# and reuse the inverse (and det(S)) across every measurement,
# instead of re-solving and re-computing det(S) per (track,
# measurement) pair. An earlier version of this restructure used
# scipy.linalg.cho_factor/cho_solve here; measured (Apple M3 Max,
# 2026-08-18, this benchmark) it was ~2x SLOWER than the original
# per-pair loop -- cho_solve's Python-level wrapper overhead
# (finite-value checks, array-API dispatch) dominates at the m=2
# sizes this benchmark exercises, the same LAPACK-generic-dispatch
# tax the gating.py njit kernels were built to avoid (see
# mahalanobis_distance's docstring). np.linalg.inv, called once per
# track, plus the already-exported (previously unwired)
# `mahalanobis_batch` njit kernel amortizing the quadratic form
# over all of a track's measurements in one call, measured ~14-28x
# faster than both the original loop and the cho_factor attempt.
try:
S_inv = np.linalg.inv(S)
except LinAlgError:
# S already passed the Cholesky PD check above, so np.linalg.inv
# should never raise here; this is a defensive fallback to the
# original per-pair path for the pathological case where it
# does anyway (mahalanobis_distance's own dispatch raises the
# same error only when its own route is also singular;
# compute_measurement_likelihood's own Cholesky guard still
# zeroes the likelihood if reached with a non-PD S).
for j in range(n_meas):
innovation = measurements[j] - z_pred
mahal_dist = mahalanobis_distance(innovation, S)
if gate_threshold is None or mahal_dist <= gate_threshold:
gated[i, j] = True
likelihood_matrix[i, j] = compute_measurement_likelihood(
innovation, S, detection_prob
)
continue
det_S = np.linalg.det(S)
innovations = measurements - z_pred
mahal_sq = np.empty(n_meas)
mahalanobis_batch(innovations, S_inv, mahal_sq)
if gate_threshold is None:
track_gated = np.ones(n_meas, dtype=bool)
else:
# mahalanobis_batch, like mahalanobis_distance, returns the
# *squared* Mahalanobis distance -- gate_threshold (e.g.
# chi2.ppf(...)) is on that same squared scale, so this
# comparison mirrors the old per-pair gate exactly.
track_gated = mahal_sq <= gate_threshold
gated[i, :] = track_gated
if det_S > 0 and n_meas:
# S already passed the Cholesky PD check above, so det_S > 0
# holds mathematically; this is a defensive floor against det()
# underflowing to exactly 0.0 for a technically-PD but severely
# ill-conditioned S.
norm_const = detection_prob / np.sqrt((2 * np.pi) ** m * det_S)
likelihood_matrix[i, track_gated] = norm_const * np.exp(
-0.5 * mahal_sq[track_gated]
)
return likelihood_matrix, gated
[docs]
def jpda_probabilities(
likelihood_matrix: NDArray[Any],
gated: NDArray[Any],
detection_prob: float = 1.0,
clutter_density: float = 1e-6,
) -> NDArray[Any]:
"""
Compute JPDA association probabilities.
Uses an efficient algorithm that avoids explicit enumeration of all
joint hypotheses when possible.
Parameters
----------
likelihood_matrix : ndarray[Any]
Likelihood values, shape (n_tracks, n_meas).
gated : ndarray[Any]
Boolean gating matrix, shape (n_tracks, n_meas).
detection_prob : float
Probability of detection (Pd).
clutter_density : float
Spatial density of clutter (lambda_c).
Returns
-------
beta : ndarray[Any]
Association probability matrix, shape (n_tracks, n_meas + 1).
beta[i, j] = P(measurement j is from track i) for j < n_meas.
beta[i, n_meas] = P(track i has no measurement).
Examples
--------
>>> import numpy as np
>>> # Likelihood matrix: 2 tracks, 2 measurements
>>> # Track 0 has high likelihood for meas 0
>>> # Track 1 has high likelihood for meas 1
>>> likelihood = np.array([[0.9, 0.1],
... [0.1, 0.8]])
>>> gated = np.array([[True, True],
... [True, True]])
>>> beta = jpda_probabilities(likelihood, gated, detection_prob=0.9)
>>> beta.shape # 2 tracks, 3 columns (2 meas + 1 miss)
(2, 3)
>>> # Track 0 most likely associated with measurement 0
>>> np.argmax(beta[0, :2])
0
"""
n_tracks, n_meas = likelihood_matrix.shape
# Initialize beta matrix (last column is for "no measurement")
beta = np.zeros((n_tracks, n_meas + 1))
if n_meas == 0:
# No measurements - all tracks have no association
beta[:, 0] = 1.0
elif n_tracks == 0:
pass
elif n_tracks <= 5 and n_meas <= 5:
# For small problems, use exact enumeration
beta = _jpda_exact(likelihood_matrix, gated, detection_prob, clutter_density)
else:
# For larger problems, use approximate method
beta = _jpda_approximate(
likelihood_matrix, gated, detection_prob, clutter_density
)
if diagnostics_enabled():
top_marginals = [
(i, int(np.argmax(beta[i])), float(np.max(beta[i])))
for i in range(n_tracks)
]
logger.bind(site="association", algo="jpda").debug(
"jpda: {} track(s), top marginal per track: {}",
n_tracks,
"; ".join(
f"track{i} -> {'miss' if j == n_meas else f'm{j}'} p={p:.3f}"
for i, j, p in top_marginals
),
)
return beta
def _jpda_exact(
likelihood_matrix: NDArray[Any],
gated: NDArray[Any],
detection_prob: float,
clutter_density: float,
) -> NDArray[Any]:
"""
Exact JPDA computation via hypothesis enumeration.
This is exponential in complexity but exact for small problems.
"""
n_tracks, n_meas = likelihood_matrix.shape
beta = np.zeros((n_tracks, n_meas + 1))
# Generate all valid joint association hypotheses
# A hypothesis is a mapping from measurements to tracks (or clutter)
# measurement j -> track assignment[j] where -1 means clutter
# We enumerate by considering all possible measurement-to-track assignments
# where each measurement can go to at most one track that gates it
def generate_hypotheses(
meas_idx: int,
current_assignment: List[int],
used_tracks: set[Any],
) -> Any:
"""Recursively generate valid hypotheses."""
if meas_idx == n_meas:
yield current_assignment.copy()
return
# Option 1: measurement is clutter
current_assignment.append(-1)
yield from generate_hypotheses(meas_idx + 1, current_assignment, used_tracks)
current_assignment.pop()
# Option 2: measurement is from a track (if gated and track not used)
for track_idx in range(n_tracks):
if gated[track_idx, meas_idx] and track_idx not in used_tracks:
current_assignment.append(track_idx)
used_tracks.add(track_idx)
yield from generate_hypotheses(
meas_idx + 1, current_assignment, used_tracks
)
used_tracks.remove(track_idx)
current_assignment.pop()
# Compute hypothesis probabilities
hypothesis_probs = []
hypothesis_assignments = []
for assignment in generate_hypotheses(0, [], set[Any]()):
# Compute probability of this hypothesis
prob = 1.0
detected_tracks = set[Any]()
for j, track_idx in enumerate(assignment):
if track_idx == -1:
# Measurement j is clutter
prob *= clutter_density
else:
# Measurement j is from track track_idx
prob *= likelihood_matrix[track_idx, j]
detected_tracks.add(track_idx)
# Account for non-detected tracks
for i in range(n_tracks):
if i in detected_tracks:
prob *= detection_prob
else:
prob *= 1.0 - detection_prob
hypothesis_probs.append(prob)
hypothesis_assignments.append(assignment)
# Normalize
total_prob = sum(hypothesis_probs)
if total_prob > 0:
hypothesis_probs = [p / total_prob for p in hypothesis_probs]
# Compute marginal association probabilities
for h_idx, (assignment, prob) in enumerate(
zip(hypothesis_assignments, hypothesis_probs)
):
detected_tracks = set[Any]()
for j, track_idx in enumerate(assignment):
if track_idx >= 0:
beta[track_idx, j] += prob
detected_tracks.add(track_idx)
# Tracks with no measurement
for i in range(n_tracks):
if i not in detected_tracks:
beta[i, n_meas] += prob
return beta
@njit(cache=True)
def _jpda_approximate_core(
likelihood_matrix: np.ndarray[Any, Any],
gated: np.ndarray[Any, Any],
detection_prob: float,
clutter_density: float,
) -> np.ndarray[Any, Any]:
"""JIT-compiled core of approximate JPDA computation."""
n_tracks = likelihood_matrix.shape[0]
n_meas = likelihood_matrix.shape[1]
beta = np.zeros((n_tracks, n_meas + 1), dtype=np.float64)
# Likelihood ratio for each measurement given each track
L = np.zeros((n_tracks, n_meas), dtype=np.float64)
for i in range(n_tracks):
for j in range(n_meas):
if gated[i, j] and clutter_density > 0:
L[i, j] = likelihood_matrix[i, j] / clutter_density
elif gated[i, j]:
L[i, j] = likelihood_matrix[i, j] * 1e10
# Compute delta factors (accounts for other tracks)
delta = np.ones((n_tracks, n_meas), dtype=np.float64)
for j in range(n_meas):
sum_L = 0.0
for i in range(n_tracks):
sum_L += L[i, j]
for i in range(n_tracks):
if sum_L > 0:
delta[i, j] = 1.0 / (1.0 + sum_L - L[i, j])
else:
delta[i, j] = 1.0
# Compute association probabilities
for i in range(n_tracks):
denom = 1.0 - detection_prob
for j in range(n_meas):
if gated[i, j]:
beta[i, j] = detection_prob * L[i, j] * delta[i, j]
denom += beta[i, j]
if denom > 0:
for j in range(n_meas):
beta[i, j] /= denom
beta[i, n_meas] = (1.0 - detection_prob) / denom
else:
beta[i, n_meas] = 1.0
return beta
def _jpda_approximate(
likelihood_matrix: NDArray[Any],
gated: NDArray[Any],
detection_prob: float,
clutter_density: float,
) -> NDArray[Any]:
"""
Approximate JPDA using parametric approach.
Uses the approach from [1] which is O(n_tracks * n_meas^2).
References
----------
- Fitzgerald, R.J., "Development of Practical PDA Logic for
Multitarget Tracking by Microprocessor", American Control
Conference, 1986.
"""
return _jpda_approximate_core(
likelihood_matrix.astype(np.float64),
gated.astype(np.bool_),
detection_prob,
clutter_density,
)
[docs]
def jpda_update(
track_states: List[ArrayLike],
track_covariances: List[ArrayLike],
measurements: ArrayLike,
H: ArrayLike,
R: ArrayLike,
detection_prob: float = 0.9,
clutter_density: float = 1e-6,
gate_probability: float = 0.99,
) -> JPDAUpdate:
"""
Perform JPDA-based track update.
Parameters
----------
track_states : list of array_like
Predicted state estimates for each track.
track_covariances : list of array_like
Predicted covariances for each track.
measurements : array_like
Measurements, shape (n_meas, m).
H : array_like
Measurement matrix, shape (m, n).
R : array_like
Measurement noise covariance, shape (m, m).
detection_prob : float
Probability of detection (Pd). Default 0.9.
clutter_density : float
Spatial density of clutter. Default 1e-6.
gate_probability : float
Probability for chi-squared gate. Default 0.99.
Returns
-------
result : JPDAUpdate
Updated states and covariances with association probabilities.
Examples
--------
>>> import numpy as np
>>> # Two tracks, three measurements
>>> x1 = np.array([0., 1.]) # [position, velocity]
>>> x2 = np.array([5., -1.])
>>> P = np.eye(2) * 0.1
>>> measurements = np.array([[0.1], [5.2], [10.0]])
>>> H = np.array([[1., 0.]]) # Measure position
>>> R = np.array([[0.1]])
>>> result = jpda_update([x1, x2], [P, P], measurements, H, R)
>>> len(result.states)
2
Notes
-----
The JPDA update consists of:
1. Compute measurement likelihoods and gating
2. Compute association probabilities
3. Compute combined innovations for each track
4. Update each track with weighted innovation
5. Compute covariance with spread of means term
References
----------
- Bar-Shalom, Y. and Fortmann, T.E., "Tracking and Data Association",
Academic Press, 1988.
"""
# Convert inputs
track_states = [np.asarray(x, dtype=np.float64).flatten() for x in track_states]
track_covariances = [np.asarray(P, dtype=np.float64) for P in track_covariances]
measurements = np.asarray(measurements, dtype=np.float64)
H = np.asarray(H, dtype=np.float64)
R = np.asarray(R, dtype=np.float64)
n_tracks = len(track_states)
n_meas = len(measurements)
m = H.shape[0]
# Compute gate threshold
gate_threshold = chi2.ppf(gate_probability, df=m)
# Compute likelihood matrix and gating.
# Pass detection_prob=1.0 so the likelihoods are pure Gaussian densities;
# jpda_probabilities applies Pd itself (passing Pd here would count it twice).
likelihood_matrix, gated = compute_likelihood_matrix(
track_states,
track_covariances,
measurements,
H,
R,
1.0,
gate_threshold,
)
# Compute association probabilities
beta = jpda_probabilities(likelihood_matrix, gated, detection_prob, clutter_density)
# Update each track
updated_states = []
updated_covs = []
innovations = []
for i in range(n_tracks):
x = track_states[i]
P = track_covariances[i]
# Innovation covariance
S = H @ P @ H.T + R
# Kalman gain
K = P @ H.T @ np.linalg.inv(S)
# Predicted measurement
z_pred = H @ x
# Compute combined innovation (weighted sum of innovations)
combined_innovation = np.zeros(m)
for j in range(n_meas):
if gated[i, j]:
innovation_j = measurements[j] - z_pred
combined_innovation += beta[i, j] * innovation_j
innovations.append(combined_innovation)
# Updated state
x_upd = x + K @ combined_innovation
# Updated covariance with spread of means term
# P_upd = beta_0 * P + (1 - beta_0) * P_c + P_spread
# where P_c = P - K @ S @ K.T (covariance after update)
# and P_spread = K @ E[y*y'] @ K.T - K @ E[y] @ E[y]' @ K.T
beta_0 = beta[i, n_meas] # Probability of no measurement
# Covariance reduction from update
P_c = P - K @ S @ K.T
# Spread of innovations (Bar-Shalom):
# P_spread = K (sum_j beta_j y_j y_j^T - y_comb y_comb^T) K^T
y_outer_sum = np.zeros((m, m))
for j in range(n_meas):
if gated[i, j]:
innovation_j = measurements[j] - z_pred
y_outer_sum += beta[i, j] * np.outer(innovation_j, innovation_j)
y_outer_sum -= np.outer(combined_innovation, combined_innovation)
P_spread = K @ y_outer_sum @ K.T
# Combined covariance
P_upd = beta_0 * P + (1 - beta_0) * P_c + P_spread
# Ensure symmetry
P_upd = (P_upd + P_upd.T) / 2
updated_states.append(x_upd)
updated_covs.append(P_upd)
return JPDAUpdate(
states=updated_states,
covariances=updated_covs,
association_probs=beta,
innovations=innovations,
)
[docs]
def jpda(
track_states: List[ArrayLike],
track_covariances: List[ArrayLike],
measurements: ArrayLike,
H: ArrayLike,
R: ArrayLike,
detection_prob: float = 0.9,
clutter_density: float = 1e-6,
gate_probability: float = 0.99,
) -> JPDAResult:
"""
Compute JPDA association probabilities.
This is a convenience function that computes association probabilities
without performing the state update.
Parameters
----------
track_states : list of array_like
Predicted state estimates for each track.
track_covariances : list of array_like
Predicted covariances for each track.
measurements : array_like
Measurements, shape (n_meas, m).
H : array_like
Measurement matrix, shape (m, n).
R : array_like
Measurement noise covariance, shape (m, m).
detection_prob : float
Probability of detection. Default 0.9.
clutter_density : float
Spatial density of clutter. Default 1e-6.
gate_probability : float
Probability for chi-squared gate. Default 0.99.
Returns
-------
result : JPDAResult
Association probabilities and related information.
Examples
--------
Compute association probabilities for 2 tracks and 3 measurements:
>>> import numpy as np
>>> # Two tracks with [x, vx] state
>>> states = [np.array([0.0, 1.0]), np.array([10.0, -0.5])]
>>> covariances = [np.eye(2) * 0.5, np.eye(2) * 0.5]
>>> # Three position measurements
>>> measurements = np.array([[0.1], [9.8], [5.0]])
>>> H = np.array([[1, 0]]) # measure position only
>>> R = np.array([[0.1]])
>>> result = jpda(states, covariances, measurements, H, R)
>>> result.association_probs.shape # (2 tracks, 4 columns: 3 meas + miss)
(2, 4)
>>> # Track 0 should have high prob for measurement 0
>>> result.association_probs[0, 0] > 0.5
True
See Also
--------
jpda_update : JPDA with state update.
"""
# Convert inputs
track_states = [np.asarray(x, dtype=np.float64).flatten() for x in track_states]
track_covariances = [np.asarray(P, dtype=np.float64) for P in track_covariances]
measurements = np.asarray(measurements, dtype=np.float64)
H = np.asarray(H, dtype=np.float64)
R = np.asarray(R, dtype=np.float64)
m = H.shape[0]
# Compute gate threshold
gate_threshold = chi2.ppf(gate_probability, df=m)
# Compute likelihood matrix and gating.
# Pass detection_prob=1.0 so the likelihoods are pure Gaussian densities;
# jpda_probabilities applies Pd itself (passing Pd here would count it twice).
likelihood_matrix, gated = compute_likelihood_matrix(
track_states,
track_covariances,
measurements,
H,
R,
1.0,
gate_threshold,
)
# Compute association probabilities
beta = jpda_probabilities(likelihood_matrix, gated, detection_prob, clutter_density)
# Compute marginal probabilities for each track
n_tracks = len(track_states)
n_meas = len(measurements)
marginal_probs = []
for i in range(n_tracks):
marginal_probs.append(beta[i, :n_meas].copy())
return JPDAResult(
association_probs=beta,
marginal_probs=marginal_probs,
likelihood_matrix=likelihood_matrix,
gated=gated,
)
__all__ = [
"JPDAResult",
"JPDAUpdate",
"compute_measurement_likelihood",
"compute_likelihood_matrix",
"jpda_probabilities",
"jpda_update",
"jpda",
]