Files
calibration/imu_lidar/imu_preintegration.py
T

461 lines
15 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""Frame-to-frame IMU preintegration (Phase-A rotation + Phase-C full factor).
Phase-A: ``ΔR``, 3×3 ``Σ``, ``J_bg``.
Phase-C: ``ΔR/Δv/Δp``, 9×9 ``Σ`` (with bias RW process noise), ``J_bg``/``J_ba``.
"""
from __future__ import annotations
from dataclasses import dataclass
import numpy as np
from .geometry import orthonormalize_rotation, so3_exp, so3_log, skew
@dataclass(frozen=True)
class GyroPreintegration:
"""Rotation-only preintegration on ``[t0, t1]`` (IMU clock)."""
delta_R: np.ndarray
duration_s: float
mean_gyro_norm: float
sigma_rad: float
weight: float
bias_rad_s: np.ndarray
cov: np.ndarray
J_bg: np.ndarray
@dataclass(frozen=True)
class ImuPreintegration:
"""Full IMU preintegration on ``[t0, t1]`` (IMU clock).
``delta_R`` maps vectors from IMU frame at ``t1`` into IMU frame at ``t0``.
``delta_v`` / ``delta_p`` are body-frame increments (no gravity).
Error-state order in ``cov`` / Jacobians: ``[δθ, δv, δp]`` (9).
``J_bg`` / ``J_ba`` are 9×3: first-order correction w.r.t. constant bias deltas.
"""
delta_R: np.ndarray
delta_v: np.ndarray
delta_p: np.ndarray
duration_s: float
mean_gyro_norm: float
sigma_rad: float
weight: float
gyro_bias_rad_s: np.ndarray
acc_bias_m_s2: np.ndarray
cov: np.ndarray
J_bg: np.ndarray
J_ba: np.ndarray
def _right_jacobian(phi: np.ndarray) -> np.ndarray:
"""SO(3) right Jacobian ``Jr(φ)`` with ``Exp(φ+δ)≈Exp(φ)Exp(Jr δ)``."""
phi = np.asarray(phi, dtype=float).reshape(3)
angle = float(np.linalg.norm(phi))
if angle < 1e-8:
return np.eye(3) - 0.5 * skew(phi)
axis = phi / angle
s = skew(axis)
return (
np.eye(3)
- ((1.0 - np.cos(angle)) / angle) * s
+ ((angle - np.sin(angle)) / angle) * (s @ s)
)
def _interp_vec(times_s: np.ndarray, values: np.ndarray, t: float) -> np.ndarray:
"""Linear interpolate a 3-vector series at an exact time."""
return np.array(
[float(np.interp(t, times_s, values[:, axis])) for axis in range(3)],
dtype=float,
)
def _interp_gyro(times_s: np.ndarray, gyro_rad_s: np.ndarray, t: float) -> np.ndarray:
"""Linear interpolate gyro at an exact time."""
return _interp_vec(times_s, gyro_rad_s, t)
def _pair_weight(duration_s: float, mean_gyro_norm: float, cov_trace: float) -> float:
"""Larger weight for short, excited, low-covariance intervals."""
duration_term = 1.0 / max(duration_s, 0.05)
excite_term = min(max(mean_gyro_norm, 1e-3), 1.0)
avg_var = max(cov_trace / 3.0, 1e-8)
return float(duration_term * excite_term / avg_var)
def preintegrate_gyro(
times_s: np.ndarray,
gyro_rad_s: np.ndarray,
t0: float,
t1: float,
bias_rad_s: np.ndarray | None = None,
*,
sigma_g_rad_s_sqrt_hz: float = 1.5e-3,
) -> GyroPreintegration:
"""Discrete mid-point gyro preintegration with exact endpoints.
``delta_R`` maps vectors from IMU frame at ``t1`` into IMU frame at ``t0``
via right-invariant updates ``ΔR ← ΔR Exp((ω-b) dt)``.
Also returns:
- ``cov``: 3×3 covariance of the right tangent noise on ``ΔR``
- ``J_bg``: ``ΔR(b+δb) ≈ ΔR Exp(J_bg δb)``
"""
times_s = np.asarray(times_s, dtype=float).reshape(-1)
gyro_rad_s = np.asarray(gyro_rad_s, dtype=float).reshape(-1, 3)
bias = np.zeros(3) if bias_rad_s is None else np.asarray(bias_rad_s, dtype=float).reshape(3)
duration = float(max(t1 - t0, 0.0))
empty = GyroPreintegration(
delta_R=np.eye(3),
duration_s=0.0,
mean_gyro_norm=0.0,
sigma_rad=1e3,
weight=1e-6,
bias_rad_s=bias.copy(),
cov=np.eye(3) * 1e6,
J_bg=np.zeros((3, 3)),
)
if times_s.size < 2 or duration <= 0:
return empty
t0 = float(np.clip(t0, times_s[0], times_s[-1]))
t1 = float(np.clip(t1, times_s[0], times_s[-1]))
duration = float(max(t1 - t0, 0.0))
if duration <= 0:
return empty
left = int(np.searchsorted(times_s, t0, side="left") - 1)
right = int(np.searchsorted(times_s, t1, side="right"))
left = max(left, 0)
right = min(right, times_s.size - 1)
if right <= left:
return empty
delta_r = np.eye(3)
j_bg = np.zeros((3, 3))
cov = np.zeros((3, 3))
sigma2 = float(sigma_g_rad_s_sqrt_hz) ** 2
gyro_norms: list[float] = []
for index in range(left, right):
t_a = float(times_s[index])
t_b = float(times_s[index + 1])
if t_b <= t0 or t_a >= t1:
continue
seg0 = max(t_a, t0)
seg1 = min(t_b, t1)
dt = seg1 - seg0
if dt <= 0:
continue
# Exact local endpoint interpolation. ``seg0`` and ``seg1`` are inside
# this adjacent sample interval, so scanning the full series with
# np.interp here would turn pair construction into quadratic work.
sample_dt = max(t_b - t_a, 1e-12)
u0 = (seg0 - t_a) / sample_dt
u1 = (seg1 - t_a) / sample_dt
g_a = (1.0 - u0) * gyro_rad_s[index] + u0 * gyro_rad_s[index + 1]
g_b = (1.0 - u1) * gyro_rad_s[index] + u1 * gyro_rad_s[index + 1]
omega = 0.5 * (g_a + g_b) - bias
gyro_norms.append(float(np.linalg.norm(omega)))
theta = omega * dt
jr = _right_jacobian(theta)
a_mat = so3_exp(-theta)
j_bg = a_mat @ j_bg - jr * dt
cov = a_mat @ cov @ a_mat.T + jr @ (sigma2 * dt * np.eye(3)) @ jr.T
delta_r = delta_r @ so3_exp(theta)
delta_r = orthonormalize_rotation(delta_r)
mean_gyro_norm = float(np.mean(gyro_norms)) if gyro_norms else 0.0
cov = 0.5 * (cov + cov.T)
cov = cov + np.eye(3) * 1e-12
if mean_gyro_norm < 0.02:
cov = cov * 4.0
cov_trace = float(np.trace(cov))
sigma_rad = float(np.sqrt(max(cov_trace / 3.0, 1e-12)))
weight = _pair_weight(duration, mean_gyro_norm, cov_trace)
return GyroPreintegration(
delta_R=delta_r,
duration_s=duration,
mean_gyro_norm=mean_gyro_norm,
sigma_rad=sigma_rad,
weight=weight,
bias_rad_s=bias.copy(),
cov=cov,
J_bg=np.asarray(j_bg, dtype=float),
)
def preintegrate_imu(
times_s: np.ndarray,
gyro_rad_s: np.ndarray,
acc_m_s2: np.ndarray,
t0: float,
t1: float,
gyro_bias_rad_s: np.ndarray | None = None,
acc_bias_m_s2: np.ndarray | None = None,
*,
sigma_g_rad_s_sqrt_hz: float = 1.5e-3,
sigma_a_m_s2_sqrt_hz: float = 2.0e-2,
sigma_bg_rw_rad_s_sqrt_hz: float = 1.0e-5,
sigma_ba_rw_m_s2_sqrt_hz: float = 1.0e-3,
) -> ImuPreintegration:
"""Mid-point IMU preintegration with exact endpoints and bias-RW noise.
Discrete updates (right-invariant)::
ΔR ← ΔR Exp((ω-bg) dt)
Δv ← Δv + ΔR (a-ba) dt
Δp ← Δp + Δv_old dt + 0.5 ΔR (a-ba) dt²
Propagates a 15-DoF error state ``[δθ, δv, δp, δbg, δba]`` then returns the
top-left 9×9 covariance (bias RW already folded in) and 9×3 Jacobians.
"""
times_s = np.asarray(times_s, dtype=float).reshape(-1)
gyro_rad_s = np.asarray(gyro_rad_s, dtype=float).reshape(-1, 3)
acc_m_s2 = np.asarray(acc_m_s2, dtype=float).reshape(-1, 3)
bg = np.zeros(3) if gyro_bias_rad_s is None else np.asarray(gyro_bias_rad_s, dtype=float).reshape(3)
ba = np.zeros(3) if acc_bias_m_s2 is None else np.asarray(acc_bias_m_s2, dtype=float).reshape(3)
empty = ImuPreintegration(
delta_R=np.eye(3),
delta_v=np.zeros(3),
delta_p=np.zeros(3),
duration_s=0.0,
mean_gyro_norm=0.0,
sigma_rad=1e3,
weight=1e-6,
gyro_bias_rad_s=bg.copy(),
acc_bias_m_s2=ba.copy(),
cov=np.eye(9) * 1e6,
J_bg=np.zeros((9, 3)),
J_ba=np.zeros((9, 3)),
)
if times_s.size < 2 or acc_m_s2.shape != gyro_rad_s.shape:
return empty
t0 = float(np.clip(t0, times_s[0], times_s[-1]))
t1 = float(np.clip(t1, times_s[0], times_s[-1]))
duration = float(max(t1 - t0, 0.0))
if duration <= 0:
return empty
left = int(np.searchsorted(times_s, t0, side="left") - 1)
right = int(np.searchsorted(times_s, t1, side="right"))
left = max(left, 0)
right = min(right, times_s.size - 1)
if right <= left:
return empty
delta_r = np.eye(3)
delta_v = np.zeros(3)
delta_p = np.zeros(3)
# Jacobians of [δθ, δv, δp] w.r.t. constant bias (accumulated analytically).
j_bg = np.zeros((9, 3))
j_ba = np.zeros((9, 3))
# 15×15 covariance: [θ, v, p, bg, ba]
cov15 = np.zeros((15, 15))
sg2 = float(sigma_g_rad_s_sqrt_hz) ** 2
sa2 = float(sigma_a_m_s2_sqrt_hz) ** 2
sbg2 = float(sigma_bg_rw_rad_s_sqrt_hz) ** 2
sba2 = float(sigma_ba_rw_m_s2_sqrt_hz) ** 2
gyro_norms: list[float] = []
for index in range(left, right):
t_a = float(times_s[index])
t_b = float(times_s[index + 1])
if t_b <= t0 or t_a >= t1:
continue
seg0 = max(t_a, t0)
seg1 = min(t_b, t1)
dt = seg1 - seg0
if dt <= 0:
continue
sample_dt = max(t_b - t_a, 1e-12)
u0 = (seg0 - t_a) / sample_dt
u1 = (seg1 - t_a) / sample_dt
g_a = (1.0 - u0) * gyro_rad_s[index] + u0 * gyro_rad_s[index + 1]
g_b = (1.0 - u1) * gyro_rad_s[index] + u1 * gyro_rad_s[index + 1]
a_a = (1.0 - u0) * acc_m_s2[index] + u0 * acc_m_s2[index + 1]
a_b = (1.0 - u1) * acc_m_s2[index] + u1 * acc_m_s2[index + 1]
omega = 0.5 * (g_a + g_b) - bg
acc = 0.5 * (a_a + a_b) - ba
gyro_norms.append(float(np.linalg.norm(omega)))
theta = omega * dt
jr = _right_jacobian(theta)
r_dt = so3_exp(theta)
r_mid = delta_r # rotate body accel into i0 frame before update
# Bias Jacobians (Forster-style first-order recursion).
j_r_bg = j_bg[0:3]
j_v_bg = j_bg[3:6]
j_p_bg = j_bg[6:9]
j_r_ba = j_ba[0:3]
j_v_ba = j_ba[3:6]
j_p_ba = j_ba[6:9]
acc_skew = skew(acc)
j_p_bg_new = j_p_bg + j_v_bg * dt - 0.5 * r_mid @ acc_skew @ j_r_bg * (dt**2)
j_v_bg_new = j_v_bg - r_mid @ acc_skew @ j_r_bg * dt
j_r_bg_new = r_dt.T @ j_r_bg - jr * dt
j_p_ba_new = j_p_ba + j_v_ba * dt - 0.5 * r_mid * (dt**2)
j_v_ba_new = j_v_ba - r_mid * dt
j_r_ba_new = r_dt.T @ j_r_ba
j_bg = np.vstack([j_r_bg_new, j_v_bg_new, j_p_bg_new])
j_ba = np.vstack([j_r_ba_new, j_v_ba_new, j_p_ba_new])
# Nominal state update (use pre-update Δv in position).
delta_p = delta_p + delta_v * dt + 0.5 * r_mid @ acc * (dt**2)
delta_v = delta_v + r_mid @ acc * dt
delta_r = orthonormalize_rotation(delta_r @ r_dt)
# Linearized error-state transition (15×15).
f = np.eye(15)
a_mat = so3_exp(-theta)
f[0:3, 0:3] = a_mat
f[0:3, 9:12] = -jr * dt
f[3:6, 0:3] = -r_mid @ acc_skew * dt
f[3:6, 12:15] = -r_mid * dt
f[6:9, 0:3] = -0.5 * r_mid @ acc_skew * (dt**2)
f[6:9, 3:6] = np.eye(3) * dt
f[6:9, 12:15] = -0.5 * r_mid * (dt**2)
# Noise: continuous densities σ²; Var(∫n dt)=σ² dt. Columns: n_g, n_a, n_bg, n_ba.
g_mat = np.zeros((15, 12))
g_mat[0:3, 0:3] = jr
g_mat[3:6, 3:6] = r_mid
g_mat[6:9, 3:6] = 0.5 * r_mid * dt
g_mat[9:12, 6:9] = np.eye(3)
g_mat[12:15, 9:12] = np.eye(3)
q = np.zeros((12, 12))
q[0:3, 0:3] = sg2 * dt * np.eye(3)
q[3:6, 3:6] = sa2 * dt * np.eye(3)
q[6:9, 6:9] = sbg2 * dt * np.eye(3)
q[9:12, 9:12] = sba2 * dt * np.eye(3)
cov15 = f @ cov15 @ f.T + g_mat @ q @ g_mat.T
delta_r = orthonormalize_rotation(delta_r)
mean_gyro_norm = float(np.mean(gyro_norms)) if gyro_norms else 0.0
cov9 = cov15[0:9, 0:9]
cov9 = 0.5 * (cov9 + cov9.T) + np.eye(9) * 1e-12
if mean_gyro_norm < 0.02:
cov9 = cov9.copy()
cov9[0:3, 0:3] = cov9[0:3, 0:3] * 4.0
cov_trace = float(np.trace(cov9[0:3, 0:3]))
sigma_rad = float(np.sqrt(max(cov_trace / 3.0, 1e-12)))
weight = _pair_weight(duration, mean_gyro_norm, cov_trace)
return ImuPreintegration(
delta_R=delta_r,
delta_v=np.asarray(delta_v, dtype=float),
delta_p=np.asarray(delta_p, dtype=float),
duration_s=duration,
mean_gyro_norm=mean_gyro_norm,
sigma_rad=sigma_rad,
weight=weight,
gyro_bias_rad_s=bg.copy(),
acc_bias_m_s2=ba.copy(),
cov=np.asarray(cov9, dtype=float),
J_bg=np.asarray(j_bg, dtype=float),
J_ba=np.asarray(j_ba, dtype=float),
)
def apply_bias_correction_imu(
preint: ImuPreintegration,
delta_gyro_bias: np.ndarray | None = None,
delta_acc_bias: np.ndarray | None = None,
) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
"""First-order bias correction of ``ΔR/Δv/Δp``.
Returns ``(delta_R, delta_v, delta_p)``.
"""
dbg = np.zeros(3) if delta_gyro_bias is None else np.asarray(delta_gyro_bias, dtype=float).reshape(3)
dba = np.zeros(3) if delta_acc_bias is None else np.asarray(delta_acc_bias, dtype=float).reshape(3)
j_bg = np.asarray(preint.J_bg, dtype=float).reshape(9, 3)
j_ba = np.asarray(preint.J_ba, dtype=float).reshape(9, 3)
delta_r = orthonormalize_rotation(preint.delta_R @ so3_exp(j_bg[0:3] @ dbg))
delta_v = preint.delta_v + j_bg[3:6] @ dbg + j_ba[3:6] @ dba
delta_p = preint.delta_p + j_bg[6:9] @ dbg + j_ba[6:9] @ dba
return delta_r, np.asarray(delta_v, dtype=float), np.asarray(delta_p, dtype=float)
def relative_rotation_from_lidar(R_X: np.ndarray, R_B: np.ndarray) -> np.ndarray:
"""Map LiDAR relative rotation into IMU frame: ``R_X R_B R_X^T``."""
r_x = orthonormalize_rotation(R_X)
r_b = orthonormalize_rotation(R_B)
return orthonormalize_rotation(r_x @ r_b @ r_x.T)
def preintegration_rotation_residual(
delta_R: np.ndarray,
R_X: np.ndarray,
R_B: np.ndarray,
) -> np.ndarray:
"""``log( delta_R^T * R_X R_B R_X^T )`` in so(3)."""
predicted = relative_rotation_from_lidar(R_X, R_B)
return so3_log(delta_R.T @ predicted)
def apply_bias_jacobian_correction(
delta_R: np.ndarray,
J_bg: np.ndarray,
delta_bias_rad_s: np.ndarray,
) -> np.ndarray:
"""First-order update ``ΔR(b+δb) ≈ ΔR Exp(J_bg δb)``."""
db = np.asarray(delta_bias_rad_s, dtype=float).reshape(3)
j_bg = np.asarray(J_bg, dtype=float).reshape(3, 3)
return orthonormalize_rotation(delta_R @ so3_exp(j_bg @ db))
def apply_constant_bias_correction(
delta_R: np.ndarray,
duration_s: float,
delta_bias_rad_s: np.ndarray,
) -> np.ndarray:
"""Legacy first-order correction when ``J_bg`` is unavailable.
``ΔR(b+δb) ≈ ΔR Exp(-δb Δt)`` (identity Jacobian approximation).
"""
db = np.asarray(delta_bias_rad_s, dtype=float).reshape(3)
return orthonormalize_rotation(delta_R @ so3_exp(-db * float(duration_s)))
def residual_whiten_matrix(cov: np.ndarray) -> np.ndarray:
"""Return ``W`` such that ``W @ e`` is approximately information-whitened.
Accepts square ``n×n`` covariances (3×3 rotation or 9×9 full IMU).
"""
matrix = np.asarray(cov, dtype=float)
if matrix.ndim != 2 or matrix.shape[0] != matrix.shape[1]:
raise ValueError("cov must be square")
n = matrix.shape[0]
matrix = 0.5 * (matrix + matrix.T) + np.eye(n) * 1e-10
try:
info = np.linalg.inv(matrix)
return np.linalg.cholesky(info).T
except np.linalg.LinAlgError:
scale = 1.0 / max(float(np.sqrt(np.trace(matrix) / n)), 1e-6)
return np.eye(n) * scale