707 lines
25 KiB
Python
707 lines
25 KiB
Python
#!/usr/bin/env python3
|
|
|
|
"""
|
|
============================================================
|
|
STEP 2b — Simultane Multiview-Optimierung für Roboterpose
|
|
============================================================
|
|
|
|
Ziel:
|
|
Aus mehreren ArUco-Detektionsdateien die gemeinsame
|
|
Roboterpose (x,y,z,a,b,c,e) schätzen und jede Kamera-Pose
|
|
sowie Marker-Weltpositionen ausgeben.
|
|
|
|
Eingabe:
|
|
--robot ../robot.json
|
|
--detections render_1a_aruco_detection.json render_1b_aruco_detection.json ...
|
|
--outDir .
|
|
|
|
Ausgabe:
|
|
multiview_pose.json
|
|
|
|
Hinweis:
|
|
Dieses Skript verwendet die Markerpositionen aus robot.json
|
|
als kinematische Constraints und optimiert gleichzeitig:
|
|
- Roboterzustand (x,y,z,a,b,c,e)
|
|
- Kameraextrinsische Parameter pro Bild
|
|
|
|
"""
|
|
|
|
import argparse
|
|
import datetime
|
|
import json
|
|
import math
|
|
import os
|
|
import time
|
|
from pathlib import Path
|
|
from typing import Any, Dict, List, Tuple
|
|
|
|
import cv2
|
|
import numpy as np
|
|
from scipy.optimize import least_squares
|
|
|
|
STATE_KEYS = ["x", "y", "z", "a", "b", "c", "e"]
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# JSON helpers
|
|
# ------------------------------------------------------------------
|
|
|
|
def load_json(path: str) -> Dict[str, Any]:
|
|
with open(path, 'r', encoding='utf-8') as f:
|
|
return json.load(f)
|
|
|
|
|
|
def save_json(data: Dict[str, Any], path: Path) -> None:
|
|
with open(path, 'w', encoding='utf-8') as f:
|
|
json.dump(data, f, indent=2)
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# robot.json helpers
|
|
# ------------------------------------------------------------------
|
|
|
|
def resolve_scalar(value: Any, default: float = 0.0) -> float:
|
|
if value is None:
|
|
return default
|
|
if isinstance(value, (int, float)):
|
|
return float(value)
|
|
try:
|
|
return float(str(value).strip())
|
|
except ValueError:
|
|
return default
|
|
|
|
|
|
def resolve_vector(value: Any, default_len: int = 3) -> Tuple[float, ...]:
|
|
if value is None:
|
|
return tuple(0.0 for _ in range(default_len))
|
|
if isinstance(value, (int, float, str)):
|
|
return (resolve_scalar(value),) + tuple(0.0 for _ in range(default_len - 1))
|
|
if isinstance(value, (list, tuple)):
|
|
resolved = [resolve_scalar(v) for v in value]
|
|
if len(resolved) < default_len:
|
|
resolved.extend([0.0] * (default_len - len(resolved)))
|
|
return tuple(resolved[:default_len])
|
|
return tuple(0.0 for _ in range(default_len))
|
|
|
|
|
|
def parse_metric_scale(robot: Dict[str, Any]) -> float:
|
|
rendering_info = robot.get('renderingInfo', {}) or {}
|
|
metric = rendering_info.get('metric', 'mm')
|
|
return 0.001 if str(metric).strip().lower() == 'mm' else 1.0
|
|
|
|
|
|
def normalize_axis(axis: Any) -> np.ndarray:
|
|
vec = np.asarray(axis, dtype=np.float64)
|
|
if vec.shape != (3,):
|
|
vec = vec.reshape(-1)[:3]
|
|
norm = np.linalg.norm(vec)
|
|
return vec / max(norm, 1e-9)
|
|
|
|
|
|
def euler_deg_to_matrix(euler_deg: Any) -> np.ndarray:
|
|
x_deg, y_deg, z_deg = resolve_vector(euler_deg, 3)
|
|
x = math.radians(x_deg)
|
|
y = math.radians(y_deg)
|
|
z = math.radians(z_deg)
|
|
|
|
cx = math.cos(x)
|
|
sx = math.sin(x)
|
|
cy = math.cos(y)
|
|
sy = math.sin(y)
|
|
cz = math.cos(z)
|
|
sz = math.sin(z)
|
|
|
|
Rx = np.array([
|
|
[1.0, 0.0, 0.0],
|
|
[0.0, cx, -sx],
|
|
[0.0, sx, cx]
|
|
], dtype=np.float64)
|
|
|
|
Ry = np.array([
|
|
[cy, 0.0, sy],
|
|
[0.0, 1.0, 0.0],
|
|
[-sy, 0.0, cy]
|
|
], dtype=np.float64)
|
|
|
|
Rz = np.array([
|
|
[cz, -sz, 0.0],
|
|
[sz, cz, 0.0],
|
|
[0.0, 0.0, 1.0]
|
|
], dtype=np.float64)
|
|
|
|
return Rz @ Ry @ Rx
|
|
|
|
|
|
def transform_from_translation_rotation(translation: Any, rotation_deg: Any) -> np.ndarray:
|
|
T = np.eye(4, dtype=np.float64)
|
|
pos = np.asarray(resolve_vector(translation, 3), dtype=np.float64)
|
|
T[:3, 3] = pos
|
|
T[:3, :3] = euler_deg_to_matrix(rotation_deg)
|
|
return T
|
|
|
|
|
|
def axis_angle_matrix(axis: Any, angle_deg: float) -> np.ndarray:
|
|
axis_vec = normalize_axis(axis)
|
|
theta = math.radians(angle_deg)
|
|
kx, ky, kz = axis_vec
|
|
c = math.cos(theta)
|
|
s = math.sin(theta)
|
|
v = 1.0 - c
|
|
R = np.array([
|
|
[kx * kx * v + c, kx * ky * v - kz * s, kx * kz * v + ky * s],
|
|
[ky * kx * v + kz * s, ky * ky * v + c, ky * kz * v - kx * s],
|
|
[kz * kx * v - ky * s, kz * ky * v + kx * s, kz * kz * v + c]
|
|
], dtype=np.float64)
|
|
T = np.eye(4, dtype=np.float64)
|
|
T[:3, :3] = R
|
|
return T
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# Kinematics and marker extraction
|
|
# ------------------------------------------------------------------
|
|
|
|
def extract_markers(robot: Dict[str, Any], scale: float) -> Dict[int, Dict[str, Any]]:
|
|
markers = {}
|
|
links = robot.get('links', {}) or {}
|
|
marker_defaults = (robot.get('renderingInfo', {}) or {}).get('markerDefaults', {}) or {}
|
|
default_size_mm = float(marker_defaults.get('size', 25.0))
|
|
|
|
for link_name, link_info in links.items():
|
|
for marker in link_info.get('markers', []) or []:
|
|
marker_id = int(marker.get('id', -1))
|
|
if marker_id < 0:
|
|
continue
|
|
|
|
pos = resolve_vector(marker.get('position', [0, 0, 0]), 3)
|
|
size_mm = float(marker.get('size', default_size_mm))
|
|
markers[marker_id] = {
|
|
'marker_id': marker_id,
|
|
'link_name': link_name,
|
|
'position_m': np.asarray([pos[0] * scale, pos[1] * scale, pos[2] * scale], dtype=np.float64),
|
|
'normal': normalize_axis(resolve_vector(marker.get('normal', [0, 0, 1]), 3)),
|
|
'spin_deg': float(marker.get('spin', 0.0)),
|
|
'size_m': size_mm * scale,
|
|
}
|
|
|
|
return markers
|
|
|
|
|
|
def marker_plane_axes(normal: np.ndarray, spin_deg: float) -> Tuple[np.ndarray, np.ndarray]:
|
|
n = normalize_axis(normal)
|
|
candidate = np.array((0.0, 0.0, 1.0), dtype=np.float64)
|
|
if abs(np.dot(n, candidate)) > 0.99:
|
|
candidate = np.array((1.0, 0.0, 0.0), dtype=np.float64)
|
|
|
|
x_dir = np.cross(candidate, n)
|
|
x_dir /= max(np.linalg.norm(x_dir), 1e-9)
|
|
y_dir = np.cross(n, x_dir)
|
|
|
|
if abs(spin_deg) > 1e-6:
|
|
theta = math.radians(spin_deg)
|
|
cos_t = math.cos(theta)
|
|
sin_t = math.sin(theta)
|
|
x_rot = x_dir * cos_t + np.cross(n, x_dir) * sin_t + n * np.dot(n, x_dir) * (1.0 - cos_t)
|
|
y_rot = y_dir * cos_t + np.cross(n, y_dir) * sin_t + n * np.dot(n, y_dir) * (1.0 - cos_t)
|
|
return x_rot, y_rot
|
|
|
|
return x_dir, y_dir
|
|
|
|
|
|
def marker_object_corners(marker: Dict[str, Any]) -> np.ndarray:
|
|
half = marker['size_m'] * 0.5
|
|
x_dir, y_dir = marker_plane_axes(marker['normal'], marker['spin_deg'])
|
|
corners = np.stack([
|
|
-x_dir * half + y_dir * half,
|
|
x_dir * half + y_dir * half,
|
|
x_dir * half - y_dir * half,
|
|
-x_dir * half - y_dir * half
|
|
], axis=0)
|
|
return marker['position_m'].reshape(1, 3) + corners
|
|
|
|
|
|
def build_link_chain(robot: Dict[str, Any]) -> List[str]:
|
|
links = robot.get('links', {}) or {}
|
|
ordered: List[str] = []
|
|
remaining = set(links.keys())
|
|
|
|
while remaining:
|
|
progress = False
|
|
for name in list(remaining):
|
|
parent = links[name].get('parent')
|
|
if not parent or parent in ordered:
|
|
ordered.append(name)
|
|
remaining.remove(name)
|
|
progress = True
|
|
if not progress:
|
|
raise RuntimeError('Cycle detected in robot link tree or missing parent link')
|
|
return ordered
|
|
|
|
|
|
def compute_link_transforms(robot: Dict[str, Any], state: Dict[str, float], scale: float) -> Dict[str, np.ndarray]:
|
|
links = robot.get('links', {}) or {}
|
|
ordered_links = build_link_chain(robot)
|
|
transforms: Dict[str, np.ndarray] = {}
|
|
|
|
for link_name in ordered_links:
|
|
link_info = links[link_name] or {}
|
|
parent_name = link_info.get('parent')
|
|
parent_transform = transforms[parent_name] if parent_name else np.eye(4, dtype=np.float64)
|
|
|
|
mount_translation = np.asarray(resolve_vector(link_info.get('mountPosition', [0, 0, 0]), 3), dtype=np.float64) * scale
|
|
mount = transform_from_translation_rotation(
|
|
mount_translation,
|
|
link_info.get('mountRotation', [0, 0, 0])
|
|
)
|
|
|
|
joint_info = link_info.get('jointToParent', {}) or {}
|
|
joint_origin = np.asarray(resolve_vector(joint_info.get('origin', [0, 0, 0]), 3), dtype=np.float64) * scale
|
|
joint = transform_from_translation_rotation(
|
|
joint_origin,
|
|
joint_info.get('rotation', [0, 0, 0])
|
|
)
|
|
|
|
motion = np.eye(4, dtype=np.float64)
|
|
joint_type = str(joint_info.get('type', 'fixed')).strip().lower()
|
|
control_var = str(joint_info.get('variable', joint_info.get('control', ''))).strip().lower()
|
|
axis = resolve_vector(joint_info.get('axis', [1, 0, 0]), 3)
|
|
|
|
if joint_type == 'linear':
|
|
motion[:3, 3] = normalize_axis(axis) * state.get(control_var, 0.0) * scale
|
|
elif joint_type == 'revolute':
|
|
motion = axis_angle_matrix(axis, state.get(control_var, 0.0))
|
|
|
|
transforms[link_name] = parent_transform @ mount @ joint @ motion
|
|
|
|
return transforms
|
|
|
|
|
|
def compute_marker_world_position(marker: Dict[str, Any], link_transforms: Dict[str, np.ndarray]) -> np.ndarray:
|
|
link_transform = link_transforms[marker['link_name']]
|
|
local = np.ones(4, dtype=np.float64)
|
|
local[:3] = marker['position_m']
|
|
world = link_transform @ local
|
|
return world[:3]
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# Camera / observation helpers
|
|
# ------------------------------------------------------------------
|
|
|
|
def load_intrinsics(detection_json: Dict[str, Any]) -> Tuple[np.ndarray, np.ndarray]:
|
|
cam = detection_json['camera']
|
|
K = np.asarray(cam['camera_matrix'], dtype=np.float64)
|
|
D = np.asarray(cam.get('distortion_coefficients', [0, 0, 0, 0, 0]), dtype=np.float64).reshape(-1, 1)
|
|
return K, D
|
|
|
|
|
|
def collect_views_and_observations(
|
|
detection_files: List[str],
|
|
robot_markers: Dict[int, Dict[str, Any]]
|
|
) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
|
|
views: List[Dict[str, Any]] = []
|
|
observations: List[Dict[str, Any]] = []
|
|
|
|
for idx, det_path in enumerate(detection_files):
|
|
detection_json = load_json(det_path)
|
|
K, D = load_intrinsics(detection_json)
|
|
views.append({
|
|
'index': idx,
|
|
'source_file': os.path.abspath(det_path),
|
|
'camera_id': detection_json.get('camera', {}).get('camera_id', f'cam{idx+1}'),
|
|
'image_file': detection_json.get('image', {}).get('image_file'),
|
|
'K': K,
|
|
'D': D
|
|
})
|
|
|
|
for det in detection_json.get('detections', []) or []:
|
|
marker_id = int(det.get('marker_id', -1))
|
|
if marker_id < 0 or marker_id not in robot_markers:
|
|
continue
|
|
|
|
image_points = det.get('image_points_px')
|
|
if isinstance(image_points, list) and len(image_points) == 4:
|
|
image_points = np.asarray(image_points, dtype=np.float64)
|
|
else:
|
|
center = resolve_vector(det.get('center_px', [0, 0]), 2)
|
|
image_points = np.asarray([center], dtype=np.float64)
|
|
|
|
confidence = float(det.get('confidence', 1.0))
|
|
marker = robot_markers[marker_id]
|
|
observations.append({
|
|
'view_index': idx,
|
|
'marker_id': marker_id,
|
|
'marker_link_corners': marker_object_corners(marker),
|
|
'image_points_px': image_points,
|
|
'confidence': max(0.01, min(1.0, confidence))
|
|
})
|
|
|
|
if len(views) == 0:
|
|
raise RuntimeError('No valid detection views found')
|
|
|
|
if len(observations) == 0:
|
|
raise RuntimeError('No marker observations matched robot.json markers')
|
|
|
|
return views, observations
|
|
|
|
|
|
def compute_marker_world_corners(marker: Dict[str, Any], link_transforms: Dict[str, np.ndarray]) -> np.ndarray:
|
|
link_transform = link_transforms[marker['link_name']]
|
|
local = marker_object_corners(marker)
|
|
homogeneous = np.concatenate([local, np.ones((local.shape[0], 1), dtype=np.float64)], axis=1)
|
|
world = (link_transform @ homogeneous.T).T
|
|
return world[:, :3]
|
|
|
|
|
|
def initial_camera_guess(
|
|
view: Dict[str, Any],
|
|
observations: List[Dict[str, Any]],
|
|
robot_markers: Dict[int, Dict[str, Any]],
|
|
default_state: Dict[str, float],
|
|
scale: float,
|
|
robot: Dict[str, Any]
|
|
) -> Tuple[np.ndarray, np.ndarray]:
|
|
object_points = []
|
|
image_points = []
|
|
|
|
link_transforms = compute_link_transforms(robot, default_state, scale)
|
|
|
|
for obs in observations:
|
|
if obs['view_index'] != view['index']:
|
|
continue
|
|
marker = robot_markers[obs['marker_id']]
|
|
object_points.append(compute_marker_world_corners(marker, link_transforms))
|
|
image_points.append(obs['image_points_px'])
|
|
|
|
if len(object_points) == 0:
|
|
return np.zeros((3, 1), dtype=np.float64), np.array([[0.0], [0.0], [1.0]], dtype=np.float64)
|
|
|
|
object_points = np.vstack(object_points)
|
|
image_points = np.vstack(image_points)
|
|
|
|
if object_points.shape[0] < 4:
|
|
return np.zeros((3, 1), dtype=np.float64), np.array([[0.0], [0.0], [1.0]], dtype=np.float64)
|
|
|
|
success, rvec, tvec = cv2.solvePnP(
|
|
object_points,
|
|
image_points,
|
|
view['K'],
|
|
view['D'],
|
|
flags=cv2.SOLVEPNP_ITERATIVE
|
|
)
|
|
|
|
if not success:
|
|
return np.zeros((3, 1), dtype=np.float64), np.array([[0.0], [0.0], [1.0]], dtype=np.float64)
|
|
|
|
return rvec, tvec
|
|
|
|
|
|
def project_points(
|
|
points_3d: np.ndarray,
|
|
rvec: np.ndarray,
|
|
tvec: np.ndarray,
|
|
K: np.ndarray,
|
|
D: np.ndarray
|
|
) -> np.ndarray:
|
|
projected, _ = cv2.projectPoints(points_3d, rvec, tvec, K, D)
|
|
return projected.reshape(-1, 2)
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# Optimization
|
|
# ------------------------------------------------------------------
|
|
|
|
def pack_parameters(robot_state: Dict[str, float], camera_params: List[Tuple[np.ndarray, np.ndarray]]) -> np.ndarray:
|
|
state_vec = np.asarray([robot_state[k] for k in STATE_KEYS], dtype=np.float64)
|
|
cams = []
|
|
for rvec, tvec in camera_params:
|
|
cams.append(rvec.reshape(3))
|
|
cams.append(tvec.reshape(3))
|
|
return np.concatenate([state_vec] + cams)
|
|
|
|
|
|
def unpack_parameters(params: np.ndarray, n_views: int) -> Tuple[Dict[str, float], List[Tuple[np.ndarray, np.ndarray]]]:
|
|
robot_state = {STATE_KEYS[i]: float(params[i]) for i in range(len(STATE_KEYS))}
|
|
camera_params = []
|
|
offset = len(STATE_KEYS)
|
|
for _ in range(n_views):
|
|
rvec = params[offset:offset + 3].reshape(3, 1)
|
|
tvec = params[offset + 3:offset + 6].reshape(3, 1)
|
|
camera_params.append((rvec, tvec))
|
|
offset += 6
|
|
return robot_state, camera_params
|
|
|
|
|
|
def residuals_for_parameters(
|
|
params: np.ndarray,
|
|
views: List[Dict[str, Any]],
|
|
observations: List[Dict[str, Any]],
|
|
robot_markers: Dict[int, Dict[str, Any]],
|
|
robot: Dict[str, Any],
|
|
scale: float,
|
|
default_state: Dict[str, float]
|
|
) -> np.ndarray:
|
|
robot_state, camera_params = unpack_parameters(params, len(views))
|
|
link_transforms = compute_link_transforms(robot, robot_state, scale)
|
|
|
|
residuals = []
|
|
for obs in observations:
|
|
marker = robot_markers[obs['marker_id']]
|
|
world_corners = compute_marker_world_corners(marker, link_transforms)
|
|
rvec, tvec = camera_params[obs['view_index']]
|
|
proj = project_points(world_corners, rvec, tvec, views[obs['view_index']]['K'], views[obs['view_index']]['D'])
|
|
diffs = proj - obs['image_points_px']
|
|
weight = math.sqrt(obs['confidence'])
|
|
residuals.extend((diffs * weight).reshape(-1))
|
|
|
|
for key in STATE_KEYS:
|
|
diff = robot_state[key] - default_state.get(key, 0.0)
|
|
if key in ('x', 'y', 'z', 'e'):
|
|
w = 0.001
|
|
else:
|
|
w = 0.01
|
|
residuals.append(diff * w)
|
|
|
|
return np.asarray(residuals, dtype=np.float64)
|
|
|
|
|
|
def estimate_uncertainty(result: Any, n_params: int) -> np.ndarray:
|
|
if result.jac is None:
|
|
return np.full(n_params, float('nan'), dtype=np.float64)
|
|
J = result.jac
|
|
m, n = J.shape
|
|
JTJ = J.T @ J
|
|
try:
|
|
cov = np.linalg.pinv(JTJ)
|
|
except np.linalg.LinAlgError:
|
|
cov = np.linalg.pinv(JTJ + np.eye(n) * 1e-9)
|
|
residuals = result.fun
|
|
dof = max(1, m - n)
|
|
sigma2 = float(np.sum(residuals ** 2) / dof)
|
|
cov *= sigma2
|
|
return np.sqrt(np.diag(cov))
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# Output assembly
|
|
# ------------------------------------------------------------------
|
|
|
|
def camera_position_world(rvec: np.ndarray, tvec: np.ndarray) -> np.ndarray:
|
|
R, _ = cv2.Rodrigues(rvec)
|
|
return (-R.T @ tvec).reshape(3)
|
|
|
|
|
|
def build_output(
|
|
robot_state: Dict[str, float],
|
|
state_uncertainty: np.ndarray,
|
|
views: List[Dict[str, Any]],
|
|
camera_params: List[Tuple[np.ndarray, np.ndarray]],
|
|
observations: List[Dict[str, Any]],
|
|
robot_markers: Dict[int, Dict[str, Any]],
|
|
scale: float,
|
|
robot: Dict[str, Any],
|
|
robot_json_path: str
|
|
) -> Dict[str, Any]:
|
|
link_transforms = compute_link_transforms(robot, robot_state, scale)
|
|
|
|
marker_summary: Dict[int, Dict[str, Any]] = {}
|
|
for marker_id, marker in robot_markers.items():
|
|
marker_summary[marker_id] = {
|
|
'marker_id': marker_id,
|
|
'link_name': marker['link_name'],
|
|
'position_world_m': compute_marker_world_position(marker, link_transforms).tolist(),
|
|
'size_m': marker['size_m'],
|
|
'observation_count': 0,
|
|
'mean_confidence': None,
|
|
'mean_reprojection_error_px': None,
|
|
'observations': []
|
|
}
|
|
|
|
per_marker_errors: Dict[int, List[float]] = {mid: [] for mid in marker_summary}
|
|
per_marker_confidences: Dict[int, List[float]] = {mid: [] for mid in marker_summary}
|
|
|
|
link_transforms = compute_link_transforms(robot, robot_state, scale)
|
|
for obs in observations:
|
|
marker_id = obs['marker_id']
|
|
marker = robot_markers[marker_id]
|
|
object_points_m = compute_marker_world_corners(marker, link_transforms)
|
|
rvec, tvec = camera_params[obs['view_index']]
|
|
proj = project_points(object_points_m, rvec, tvec, views[obs['view_index']]['K'], views[obs['view_index']]['D'])
|
|
diffs = proj - obs['image_points_px']
|
|
errors = np.linalg.norm(diffs, axis=1)
|
|
repro_error = float(np.mean(errors))
|
|
per_marker_errors[marker_id].extend(errors.tolist())
|
|
per_marker_confidences[marker_id].append(obs['confidence'])
|
|
marker_summary[marker_id]['observation_count'] += 1
|
|
marker_summary[marker_id]['observations'].append({
|
|
'view_index': obs['view_index'],
|
|
'source_file': views[obs['view_index']]['source_file'],
|
|
'image_file': views[obs['view_index']]['image_file'],
|
|
'confidence': obs['confidence'],
|
|
'mean_reprojection_error_px': repro_error,
|
|
'corner_reprojection_errors_px': errors.tolist()
|
|
})
|
|
|
|
for marker_id, summary in marker_summary.items():
|
|
if summary['observation_count'] > 0:
|
|
summary['mean_confidence'] = float(np.mean(per_marker_confidences[marker_id]))
|
|
summary['mean_reprojection_error_px'] = float(np.mean(per_marker_errors[marker_id]))
|
|
|
|
camera_outputs = []
|
|
for idx, view in enumerate(views):
|
|
rvec, tvec = camera_params[idx]
|
|
cam_pos = camera_position_world(rvec, tvec)
|
|
observed_count = sum(1 for obs in observations if obs['view_index'] == idx)
|
|
camera_outputs.append({
|
|
'view_index': idx,
|
|
'source_file': view['source_file'],
|
|
'camera_id': view['camera_id'],
|
|
'camera_position_world_m': cam_pos.tolist(),
|
|
'rvec': rvec.reshape(-1).tolist(),
|
|
'tvec': tvec.reshape(-1).tolist(),
|
|
'intrinsics': {
|
|
'camera_matrix': view['K'].tolist(),
|
|
'distortion_coefficients': view['D'].reshape(-1).tolist()
|
|
},
|
|
'observation_count': observed_count
|
|
})
|
|
|
|
robot_pose_output = {
|
|
'state': {k: float(robot_state[k]) for k in STATE_KEYS},
|
|
'uncertainty': {
|
|
'x_mm': float(state_uncertainty[0]),
|
|
'y_mm': float(state_uncertainty[1]),
|
|
'z_mm': float(state_uncertainty[2]),
|
|
'a_deg': float(state_uncertainty[3]),
|
|
'b_deg': float(state_uncertainty[4]),
|
|
'c_deg': float(state_uncertainty[5]),
|
|
'e_mm': float(state_uncertainty[6])
|
|
},
|
|
'confidence': {
|
|
'x': float(math.exp(-state_uncertainty[0] / 10.0)),
|
|
'y': float(math.exp(-state_uncertainty[1] / 10.0)),
|
|
'z': float(math.exp(-state_uncertainty[2] / 10.0)),
|
|
'a': float(math.exp(-state_uncertainty[3] / 10.0)),
|
|
'b': float(math.exp(-state_uncertainty[4] / 10.0)),
|
|
'c': float(math.exp(-state_uncertainty[5] / 10.0)),
|
|
'e': float(math.exp(-state_uncertainty[6] / max(1.0, state_uncertainty[6])))
|
|
}
|
|
}
|
|
|
|
return {
|
|
'schema_version': '1.0',
|
|
'created_utc': datetime.datetime.utcnow().isoformat() + 'Z',
|
|
'source_robot_json': os.path.abspath(robot_json_path),
|
|
'source_detections': [view['source_file'] for view in views],
|
|
'robot_pose': robot_pose_output,
|
|
'camera_poses': camera_outputs,
|
|
'marker_positions': list(marker_summary.values())
|
|
}
|
|
|
|
|
|
# ------------------------------------------------------------------
|
|
# Main
|
|
# ------------------------------------------------------------------
|
|
|
|
def main() -> None:
|
|
parser = argparse.ArgumentParser(description='Multiview optimization of robot pose and camera extrinsics')
|
|
parser.add_argument('--robot', required=True, help='Path to robot.json')
|
|
parser.add_argument('--detections', required=True, nargs='+', help='List of detection JSON files')
|
|
parser.add_argument('--outDir', required=True, help='Output directory')
|
|
parser.add_argument('--write-summary', action='store_true', help='Write summary file')
|
|
parser.add_argument('--max-iter', type=int, default=500, help='Maximum optimizer iterations')
|
|
args = parser.parse_args()
|
|
|
|
os.makedirs(args.outDir, exist_ok=True)
|
|
|
|
robot_json_path = os.path.abspath(args.robot)
|
|
robot = load_json(robot_json_path)
|
|
scale = parse_metric_scale(robot)
|
|
|
|
default_state = {
|
|
k: float(robot.get('defaultPosition', {}).get(k, 0.0) or 0.0)
|
|
for k in STATE_KEYS
|
|
}
|
|
|
|
robot_markers = extract_markers(robot, scale)
|
|
views, observations = collect_views_and_observations(args.detections, robot_markers)
|
|
|
|
camera_guesses = []
|
|
for view in views:
|
|
rvec, tvec = initial_camera_guess(view, observations, robot_markers, default_state, scale, robot)
|
|
camera_guesses.append((rvec, tvec))
|
|
|
|
x0 = pack_parameters(default_state, camera_guesses)
|
|
|
|
progress = {
|
|
'iter': 0,
|
|
'last_cost': None,
|
|
'last_print': time.time(),
|
|
'prev_x': x0.copy()
|
|
}
|
|
|
|
def progress_callback(xk: np.ndarray) -> None:
|
|
progress['iter'] += 1
|
|
now = time.time()
|
|
if progress['iter'] == 1 or now - progress['last_print'] >= 1.0:
|
|
res = residuals_for_parameters(xk, views, observations, robot_markers, robot, scale, default_state)
|
|
cost = 0.5 * float(np.dot(res, res))
|
|
delta_cost = None
|
|
convergence = ''
|
|
if progress['last_cost'] is not None:
|
|
delta_cost = cost - progress['last_cost']
|
|
if abs(delta_cost) < 1e-3:
|
|
convergence = ' stable'
|
|
elif delta_cost < 0:
|
|
convergence = ' improving'
|
|
else:
|
|
convergence = ' worsening'
|
|
step_norm = float(np.linalg.norm(xk - progress['prev_x']))
|
|
print(
|
|
f'[Multiview] iter={progress["iter"]:4d} cost={cost:.4f}'
|
|
+ (f' delta={delta_cost:.4g}' if delta_cost is not None else '')
|
|
+ f' step={step_norm:.4g}'
|
|
+ convergence
|
|
)
|
|
progress['last_cost'] = cost
|
|
progress['last_print'] = now
|
|
progress['prev_x'] = xk.copy()
|
|
|
|
result = least_squares(
|
|
residuals_for_parameters,
|
|
x0,
|
|
args=(views, observations, robot_markers, robot, scale, default_state),
|
|
jac='2-point',
|
|
method='trf',
|
|
loss='soft_l1',
|
|
f_scale=1.0,
|
|
max_nfev=args.max_iter,
|
|
callback=progress_callback
|
|
)
|
|
|
|
robot_state, camera_params = unpack_parameters(result.x, len(views))
|
|
uncertainties = estimate_uncertainty(result, len(result.x))
|
|
|
|
output = build_output(robot_state, uncertainties[:len(STATE_KEYS)], views, camera_params, observations, robot_markers, scale, robot, robot_json_path)
|
|
|
|
out_path = Path(args.outDir) / 'multiview_pose.json'
|
|
save_json(output, out_path)
|
|
|
|
print(f'Saved: {out_path}')
|
|
if args.write_summary:
|
|
summary_path = Path(args.outDir) / 'multiview_pose_summary.json'
|
|
summary = {
|
|
'final_cost': float(result.cost),
|
|
'status': int(result.status),
|
|
'message': result.message,
|
|
'robot_state': output['robot_pose'],
|
|
'camera_count': len(views),
|
|
'marker_count': len(robot_markers)
|
|
}
|
|
save_json(summary, summary_path)
|
|
print(f'Saved: {summary_path}')
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|