feat: implement ground plane orchestration

py_workspace/tests/test_ground_plane.py

@@ -5,8 +5,11 @@ from aruco.ground_plane import (
     detect_floor_plane,
     compute_consensus_plane,
     compute_floor_correction,
+    refine_ground_from_depth,
     FloorPlane,
     FloorCorrection,
+    GroundPlaneConfig,
+    GroundPlaneMetrics,
 )
 
 
@@ -315,3 +318,160 @@ def test_compute_floor_correction_bounds():
 
     assert not result.valid
     assert "exceeds limit" in result.reason
+
+
+def test_refine_ground_from_depth_disabled():
+    config = GroundPlaneConfig(enabled=False)
+    extrinsics = {"cam1": np.eye(4)}
+    camera_data = {"cam1": {"depth": np.zeros((10, 10)), "K": np.eye(3)}}
+
+    new_extrinsics, metrics = refine_ground_from_depth(camera_data, extrinsics, config)
+
+    assert not metrics.success
+    assert "disabled" in metrics.message
+    assert new_extrinsics == extrinsics
+
+
+def test_refine_ground_from_depth_insufficient_cameras():
+    # Only 1 camera, need 2
+    config = GroundPlaneConfig(min_valid_cameras=2, min_inliers=10)
+
+    # Create fake depth map that produces a plane
+    # Plane at y=-1.0
+    width, height = 20, 20
+    K = np.eye(3)
+    K[0, 2] = 10
+    K[1, 2] = 10
+    K[0, 0] = 20
+    K[1, 1] = 20
+
+    # Generate points on plane y=-1.0
+    # In camera frame (assuming cam at origin looking -Z), floor is at Y=-1.0
+    # But wait, standard camera frame is Y-down.
+    # Let's assume world frame is Y-up.
+    # If cam is at origin, and looking down -Z (OpenGL) or +Z (OpenCV).
+    # Let's use identity extrinsics -> cam frame = world frame.
+    # World frame Y-up.
+    # So we want points with y=-1.0.
+    # But unproject_depth gives points in camera frame.
+    # If we want world y=-1.0, and T=I, then cam y=-1.0.
+    # But unproject uses OpenCV convention: Y-down.
+    # So y=-1.0 means 1m UP in camera frame.
+    # Let's just make points that form A plane, doesn't matter which one,
+    # as long as it's detected.
+
+    # Let's make a flat plane at Z=2.0 (fronto-parallel)
+    depth_map = np.full((height, width), 2.0, dtype=np.float32)
+
+    # Need to ensure we have enough points for RANSAC
+    # 20x20 = 400 points.
+    # Stride default is 8. 20/8 = 2. 2x2 = 4 points.
+    # RANSAC n=3. So 4 points is enough.
+    # But min_inliers=10. 4 < 10.
+    # So we need to reduce stride or increase size.
+    config.stride = 1
+
+    camera_data = {"cam1": {"depth": depth_map, "K": K}}
+    extrinsics = {"cam1": np.eye(4)}
+
+    new_extrinsics, metrics = refine_ground_from_depth(camera_data, extrinsics, config)
+
+    assert not metrics.success
+    assert "Found 1 valid planes" in metrics.message
+    assert metrics.num_cameras_valid == 1
+
+
+def test_refine_ground_from_depth_success():
+    # 2 cameras, both seeing floor at y=-1.0
+    # We want to correct it to y=0.0
+    config = GroundPlaneConfig(
+        min_valid_cameras=2,
+        min_inliers=10,
+        target_y=0.0,
+        max_translation_m=2.0,
+        ransac_dist_thresh=0.05,
+    )
+
+    width, height = 20, 20
+    K = np.eye(3)
+    K[0, 2] = 10
+    K[1, 2] = 10
+    K[0, 0] = 20
+    K[1, 1] = 20
+
+    # Create points on plane y=-1.0 in WORLD frame
+    # Cam 1 at origin. T_world_cam = I.
+    # So points in cam 1 should be at y=-1.0.
+    # OpenCV cam: Y-down. So y=-1.0 is UP.
+    # Let's just use the fact that we transform points to world frame before detection.
+    # So if we make depth map such that unprojected points + extrinsics -> plane y=-1.0.
+
+    # Let's manually mock the detection to avoid complex depth map math
+    # We can't easily mock internal functions without monkeypatching.
+    # Instead, let's construct a depth map that corresponds to a plane.
+    # Simplest: Camera looking down at floor.
+    # Cam at (0, 2, 0) looking at (0, 0, 0).
+    # World floor at y=0.
+    # Cam floor distance = 2.0.
+    # But here we want to simulate a MISALIGNED floor.
+    # Say we think floor is at y=-1.0 (in our current world frame).
+    # So we generate points at y=-1.0.
+
+    # Let's try a simpler approach:
+    # Create depth map for a plane Z=2.0 in camera frame.
+    # Set extrinsics such that this plane becomes Y=-1.0 in world frame.
+    # Plane Z=2.0 in cam frame: (x, y, 2).
+    # We want R * (x, y, 2) + t = (X, -1, Z).
+    # Let R = Rotation(-90 deg around X).
+    # R = [[1, 0, 0], [0, 0, 1], [0, -1, 0]]
+    # R * (x, y, 2) = (x, 2, -y).
+    # We want Y_world = -1.
+    # So 2 + ty = -1 => ty = -3.
+    # So if we put cam at y=-3, rotated -90X.
+    # Then Z=2 plane becomes Y=-1 plane.
+
+    # Rotation -90 deg around X
+    Rx_neg90 = np.array([[1, 0, 0], [0, 0, 1], [0, -1, 0]])
+    t = np.array([0, -3, 0])
+    T_world_cam = np.eye(4)
+    T_world_cam[:3, :3] = Rx_neg90
+    T_world_cam[:3, 3] = t
+
+    # Depth map: constant 2.0
+    depth_map = np.full((height, width), 2.0, dtype=np.float32)
+
+    # Need to ensure we have enough points for RANSAC
+    # 20x20 = 400 points.
+    # Stride default is 8. 20/8 = 2. 2x2 = 4 points.
+    # RANSAC n=3. So 4 points is enough.
+    # But min_inliers=10. 4 < 10.
+    # So we need to reduce stride or increase size.
+    config.stride = 1
+
+    camera_data = {
+        "cam1": {"depth": depth_map, "K": K},
+        "cam2": {"depth": depth_map, "K": K},
+    }
+    extrinsics = {
+        "cam1": T_world_cam,
+        "cam2": T_world_cam,
+    }
+
+    new_extrinsics, metrics = refine_ground_from_depth(camera_data, extrinsics, config)
+
+    assert metrics.success
+    assert metrics.num_cameras_valid == 2
+    assert metrics.correction_applied
+
+    # We started with floor at y=-1.0. Target is y=0.0.
+    # So we expect translation of +1.0 in Y.
+    # T_corr should have ty approx 1.0.
+    T_corr = metrics.correction_transform
+    assert abs(T_corr[1, 3] - 1.0) < 0.1  # Allow some slack for RANSAC noise
+
+    # Check new extrinsics
+    # New T = T_corr @ Old T
+    # Old T origin y = -3.
+    # New T origin y should be -3 + 1 = -2.
+    T_new = new_extrinsics["cam1"]
+    assert abs(T_new[1, 3] - (-2.0)) < 0.1