File size: 34,325 Bytes
5f8bdff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0e7815e
5f8bdff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
#!/usr/bin/env python3
"""
Pose Interpolation & Smoothing Module (A12)
============================================
Provides robust smoothing strategies for pose-estimation keypoint sequences,
eliminating jitter, filling detection gaps, and removing outlier spikes
to produce clean animations.

Supported strategies
--------------------
- ``moving_average``        Box-car (sliding window) average
- ``gaussian``              Gaussian-weighted convolution
- ``exponential``           Exponential moving average (EMA) – usable online
- ``median``                Median filter – excellent against isolated spikes
- ``savitzky_golay``        Savitzky-Golay filter – preserves signal shape
- ``kalman``                1-D constant-velocity Kalman filter (online)
- ``spline``                Cubic-spline interpolation through high-confidence
                            points, discarding outliers
- ``hybrid``                (Default) Outlier detection → interpolation →
                            Savitzky-Golay smoothing

Data formats
------------
The module accepts two common representations:

1. **App‑style list of dicts** (from ``app.py`` ``all_keypoints``):

   .. code:: python

       [
           {
               "poses": [{
                   "keypoints": [
                       {"x": 0.5, "y": 0.3, "score": 0.92, "name": "nose"},
                       ...
                   ]
               }],
               "frame_id": 0, ...
           },
           ...
       ]

2. **NumPy array** with shape ``(frames, joints, 3)`` where the last axis
   holds ``[x, y, confidence]``.

Basic usage
-----------
.. code:: python

    from A12.pose_interpolator import PoseInterpolator, smooth_pose_sequence

    # High-level convenience (recommended)
    smoothed = smooth_pose_sequence(all_keypoints, strategy="hybrid")

    # Low-level API
    interp = PoseInterpolator(strategy="kalman", process_noise=0.001,
                              measurement_noise=0.05)
    arr = interp.keypoints_to_array(all_keypoints)
    smoothed_arr = interp.fit_transform(arr)
"""

from __future__ import annotations

import warnings
from collections import defaultdict
from copy import deepcopy
from enum import Enum
from typing import Any, Dict, List, Optional, Sequence, Tuple, Union

import numpy as np

import pandas as pd

# ---------------------------------------------------------------------------
# Try importing scipy – it is a transitive dependency of statsmodels /
# scikit-learn (both in requirements.txt), but we guard in case it is
# not available for some strategies.
# ---------------------------------------------------------------------------
try:
    from scipy import interpolate as _scipy_interpolate, ndimage, signal as _scipy_signal

    _HAS_SCIPY = True
except ImportError:  # pragma: no cover
    _HAS_SCIPY = False


# ---------------------------------------------------------------------------
# Standard joint-name lists (for reference / validation)
# ---------------------------------------------------------------------------

COCO_KEYPOINTS: List[str] = [
    "nose",
    "left_eye",
    "right_eye",
    "left_ear",
    "right_ear",
    "left_shoulder",
    "right_shoulder",
    "left_elbow",
    "right_elbow",
    "left_wrist",
    "right_wrist",
    "left_hip",
    "right_hip",
    "left_knee",
    "right_knee",
    "left_ankle",
    "right_ankle",
]

A11_JOINTS: List[str] = [
    "head",
    "left_shoulder",
    "left_elbow",
    "right_shoulder",
    "right_elbow",
    "left_hand",
    "right_hand",
    "left_hip",
    "right_hip",
    "left_knee",
    "right_knee",
    "left_foot",
    "right_foot",
]


class SmoothingStrategy(Enum):
    """Available smoothing strategies."""

    MOVING_AVERAGE = "moving_average"
    GAUSSIAN = "gaussian"
    EXPONENTIAL = "exponential"
    MEDIAN = "median"
    SAVITZKY_GOLAY = "savitzky_golay"
    KALMAN = "kalman"
    SPLINE = "spline"
    HYBRID = "hybrid"


# ===================================================================
# Small helpers
# ===================================================================


def _validate_array(arr: np.ndarray) -> np.ndarray:
    """Ensure *arr* is a contiguous float64 array of shape (F, J, 3)."""
    arr = np.asarray(arr, dtype=np.float64)
    if arr.ndim != 3 or arr.shape[2] != 3:
        raise ValueError(
            f"Expected array of shape (frames, joints, 3), got {arr.shape}"
        )
    return arr


def _ensure_scipy(strategy_name: str) -> None:
    if not _HAS_SCIPY:
        raise ImportError(
            f"Strategy '{strategy_name}' requires scipy, which is not installed."
        )


# ===================================================================
# 1-D Kalman filter (constant velocity) – no external dependencies
# ===================================================================


class KalmanFilter1D:
    """
    Simple 1-D constant-velocity Kalman filter.

    State: [position, velocity]ᵀ
    Measurement: position

    Parameters
    ----------
    process_noise : float
        Process noise (Q) – higher values trust the model less.
    measurement_noise : float
        Measurement noise (R) – higher values trust measurements less.

    Notes
    -----
    When ``update(None)`` is called the filter performs a pure prediction
    step, allowing it to bridge gaps in the detection sequence.
    """

    def __init__(
        self, process_noise: float = 0.001, measurement_noise: float = 0.1
    ) -> None:
        # State transition (constant velocity, dt = 1)
        self._A = np.array([[1, 1], [0, 1]], dtype=np.float64)
        self._H = np.array([[1, 0]], dtype=np.float64)  # observe position only
        self._Q = np.eye(2, dtype=np.float64) * process_noise
        self._R = np.array([[measurement_noise]], dtype=np.float64)
        self._P = np.eye(2, dtype=np.float64) * 1.0
        self._x = np.zeros(2, dtype=np.float64)

    @property
    def position(self) -> float:
        """Current position estimate."""
        return float(self._x[0])

    def reset(self, position: float = 0.0) -> None:
        """Re-initialise filter with a new position and zero velocity."""
        self._x = np.array([position, 0.0], dtype=np.float64)
        self._P = np.eye(2, dtype=np.float64) * 1.0

    def update(self, measurement: Optional[float]) -> float:
        """
        Predict + (optionally) update step.

        Parameters
        ----------
        measurement : float or None
            Observed position. If ``None``, only the prediction step runs.

        Returns
        -------
        float
            Filtered position estimate.
        """
        # -- predict --
        self._x = self._A @ self._x
        self._P = self._A @ self._P @ self._A.T + self._Q

        if measurement is not None:
            # -- update --
            y = measurement - (self._H @ self._x)  # innovation
            S = self._H @ self._P @ self._H.T + self._R
            K = self._P @ self._H.T @ np.linalg.inv(S)  # Kalman gain
            self._x = self._x + (K @ y).ravel()
            self._P = (np.eye(2) - K @ self._H) @ self._P

        return self.position


# ===================================================================
# Outlier detection utilities
# ===================================================================


def detect_outliers_velocity(
    positions: np.ndarray,
    threshold: float = 3.0,
    min_confidence: float = 0.2,
    confidences: Optional[np.ndarray] = None,
) -> np.ndarray:
    """
    Flag outliers based on inter-frame velocity.

    A point is considered an outlier if its frame-to-frame displacement
    exceeds *threshold* times the median absolute deviation of all
    non-zero displacements across the sequence.

    Parameters
    ----------
    positions : (F,) ndarray
        1-D coordinate signal (may contain NaN).
    threshold : float
        MAD multiplier.
    min_confidence : float
        Points with confidence below this are already treated as
        missing – they are **not** flagged here (they will be
        interpolated later).
    confidences : (F,) ndarray or None
        Confidence values (same length as *positions*).

    Returns
    -------
    (F,) bool ndarray
        ``True`` where the point is an outlier.
    """
    positions = np.asarray(positions, dtype=np.float64)
    n = len(positions)

    outlier_mask = np.zeros(n, dtype=bool)

    # Low-confidence points are handled by the interpolation stage.
    if confidences is not None:
        confidences = np.asarray(confidences, dtype=np.float64)
        low_conf = confidences < min_confidence
    else:
        low_conf = np.isnan(positions)

    # Compute finite differences
    diffs = np.abs(np.diff(positions))
    valid_diffs = diffs[np.isfinite(diffs)]
    if len(valid_diffs) == 0:
        return outlier_mask

    mad = np.median(np.abs(valid_diffs - np.median(valid_diffs)))
    if mad == 0:
        mad = np.mean(valid_diffs) + 1e-9  # fallback

    limit = threshold * mad * 1.4826  # 1.4826 converts MAD → std for normal

    for i in range(1, n):
        if low_conf[i]:
            continue
        d = abs(positions[i] - positions[i - 1])
        if np.isfinite(d) and d > limit:
            outlier_mask[i] = True

    return outlier_mask


def detect_outliers_zscore(
    positions: np.ndarray,
    threshold: float = 3.0,
    min_confidence: float = 0.2,
    confidences: Optional[np.ndarray] = None,
) -> np.ndarray:
    """
    Flag outliers whose absolute z-score exceeds *threshold*.

    Computed against the sequence mean / std (ignoring NaN).
    """
    positions = np.asarray(positions, dtype=np.float64)
    n = len(positions)

    outlier_mask = np.zeros(n, dtype=bool)

    if confidences is not None:
        confidences = np.asarray(confidences, dtype=np.float64)
        low_conf = confidences < min_confidence
    else:
        low_conf = np.isnan(positions)

    finite = np.isfinite(positions) & ~low_conf
    if not finite.any():
        return outlier_mask

    mu = np.mean(positions[finite])
    sigma = np.std(positions[finite])
    if sigma == 0:
        return outlier_mask

    z = np.abs(positions - mu) / sigma
    outlier_mask = (z > threshold) & ~low_conf & np.isfinite(positions)
    return outlier_mask


# ===================================================================
# Core interpolator class
# ===================================================================


class PoseInterpolator:
    """
    Smooth a multi-joint pose trajectory using a configurable strategy.

    Parameters
    ----------
    strategy : str or SmoothingStrategy
        One of the strategies listed in :class:`SmoothingStrategy`.
    window_size : int
        Window length (frames) for moving‑average, gaussian, median,
        savitzky_golay.  Must be odd for savitzky_golay and median.
    poly_order : int
        Polynomial order for Savitzky‑Golay (must be < window_size).
    sigma : float
        Standard deviation of the Gaussian kernel.
    alpha : float
        Smoothing factor for exponential moving average (0 < α ≤ 1).
        Larger α gives more weight to recent observations.
    process_noise : float
        Kalman-filter process noise.
    measurement_noise : float
        Kalman-filter measurement noise.
    outlier_method : str
        ``"velocity"`` or ``"zscore"`` – used by the hybrid strategy.
    outlier_threshold : float
        MAD / z-score multiplier for outlier flagging.
    min_confidence : float
        Keypoints with confidence below this are treated as missing and
        interpolated regardless of strategy.
    fill_method : str
        How to fill missing / masked positions before smoothing:
        ``"linear"``, ``"spline"``, or ``"forward"`` (last valid
        carried forward).
    """

    def __init__(
        self,
        strategy: Union[str, SmoothingStrategy] = SmoothingStrategy.HYBRID,
        window_size: int = 5,
        poly_order: int = 2,
        sigma: float = 1.0,
        alpha: float = 0.3,
        process_noise: float = 0.001,
        measurement_noise: float = 0.05,
        outlier_method: str = "velocity",
        outlier_threshold: float = 1.0,
        min_confidence: float = 0.2,
        fill_method: str = "linear",
    ) -> None:
        if isinstance(strategy, str):
            strategy = SmoothingStrategy(strategy)
        self.strategy = strategy
        self.window_size = window_size
        self.poly_order = poly_order
        self.sigma = sigma
        self.alpha = float(np.clip(alpha, 0.0, 1.0))
        self.process_noise = process_noise
        self.measurement_noise = measurement_noise
        self.outlier_method = outlier_method
        self.outlier_threshold = outlier_threshold
        self.min_confidence = min_confidence
        self.fill_method = fill_method

        # Internal state (set during fit / transform)
        self._joint_names: List[str] = []
        self._n_frames: int = 0
        self._n_joints: int = 0

    # --- public API ---------------------------------------------------

    @staticmethod
    def keypoints_to_array(
        frames_data: List[Dict[str, Any]],
        joint_names: Optional[List[str]] = None,
    ) -> np.ndarray:
        """
        Convert *app.py* style frame dicts into a ``(F, J, 3)`` array.

        Parameters
        ----------
        frames_data : list of dict
            Each dict must have the structure produced by
            ``extract_joint_positions_from_movenet()`` (see module
            docstring).
        joint_names : list of str, optional
            Names of joints in the desired order.  When ``None``,
            ``COCO_KEYPOINTS`` is used.

        Returns
        -------
        ndarray of shape ``(len(frames_data), len(joint_names), 3)``
            The last axis holds ``[x, y, confidence]``. Missing values
            are represented as ``NaN``.
        """
        if joint_names is None:
            joint_names = COCO_KEYPOINTS

        n_frames = len(frames_data)
        n_joints = len(joint_names)
        arr = np.full((n_frames, n_joints, 3), np.nan, dtype=np.float64)

        for f_idx, frame in enumerate(frames_data):
            poses = frame.get("poses", [])
            if not poses:
                continue
            kps = poses[0].get("keypoints", [])
            kp_map = {kp.get("name"): kp for kp in kps}
            for j_idx, name in enumerate(joint_names):
                kp = kp_map.get(name)
                if kp is None:
                    continue
                x, y, c = kp.get("x"), kp.get("y"), kp.get("score")
                arr[f_idx, j_idx, 0] = x if x is not None else np.nan
                arr[f_idx, j_idx, 1] = y if y is not None else np.nan
                arr[f_idx, j_idx, 2] = (
                    c if c is not None else np.nan
                )

        return arr

    @staticmethod
    def array_to_keypoints(
        arr: np.ndarray,
        frames_data: List[Dict[str, Any]],
        joint_names: Optional[List[str]] = None,
    ) -> List[Dict[str, Any]]:
        """
        Write a smoothed ``(F, J, 3)`` array back into the original
        app‑style frame dicts (returns a **deep copy** of
        *frames_data* with modified keypoint coordinates).

        Confidence values are preserved from the original data;
        coordinates are overwritten with the smoothed values.
        """
        if joint_names is None:
            joint_names = COCO_KEYPOINTS

        arr = _validate_array(arr)
        n_frames, n_joints, _ = arr.shape

        out: List[Dict[str, Any]] = deepcopy(frames_data)

        for f_idx in range(min(n_frames, len(out))):
            poses = out[f_idx].get("poses", [])
            if not poses:
                continue
            kps = poses[0].get("keypoints", [])
            kp_map = {kp.get("name"): kp for kp in kps}
            for j_idx, name in enumerate(joint_names):
                kp = kp_map.get(name)
                if kp is None:
                    continue
                if not np.isnan(arr[f_idx, j_idx, 0]):
                    kp["x"] = float(arr[f_idx, j_idx, 0])
                if not np.isnan(arr[f_idx, j_idx, 1]):
                    kp["y"] = float(arr[f_idx, j_idx, 1])
                # confidence deliberately kept from original

        return out

    @staticmethod
    def array_to_dataframe(
        arr: np.ndarray,
        joint_names: Optional[List[str]] = None,
        frame_numbers: Optional[Sequence[int]] = None,
    ) -> "pd.DataFrame":
        """
        Convert ``(F, J, 3)`` array to a DataFrame compatible with
        A11 visualisation tools (columns ``<joint>_x``, ``<joint>_y``,
        optionally ``<joint>_z`` and a ``FrameNo`` column).
        """

        if joint_names is None:
            joint_names = COCO_KEYPOINTS

        arr = _validate_array(arr)
        n_frames, n_joints, _ = arr.shape

        data: Dict[str, List[float]] = {}
        for j_idx, name in enumerate(joint_names):
            data[f"{name}_x"] = arr[:, j_idx, 0].tolist()
            data[f"{name}_y"] = arr[:, j_idx, 1].tolist()

        if frame_numbers is None:
            data["FrameNo"] = list(range(n_frames))
        else:
            data["FrameNo"] = list(frame_numbers)

        return pd.DataFrame(data)

    def fit_transform(self, arr: np.ndarray) -> np.ndarray:
        """
        Run the full smoothing pipeline on one array.

        Parameters
        ----------
        arr : (F, J, 3) ndarray
            Raw coordinates + confidence.

        Returns
        -------
        (F, J, 3) ndarray
            Smoothed coordinates.  The confidence channel is passed
            through unchanged (it is used internally for filtering).
        """
        arr = _validate_array(arr)
        self._n_frames, self._n_joints, _ = arr.shape

        # Always mask low-confidence points before any processing
        arr = self._mask_low_confidence(arr)

        if self.strategy == SmoothingStrategy.KALMAN:
            smoothed = self._apply_kalman(arr)
        elif self.strategy == SmoothingStrategy.EXPONENTIAL:
            smoothed = self._apply_ema(arr)
        elif self.strategy == SmoothingStrategy.SPLINE:
            smoothed = self._apply_spline(arr)
        elif self.strategy == SmoothingStrategy.HYBRID:
            smoothed = self._apply_hybrid(arr)
        else:
            # scipy-based windowed filters
            smoothed = self._apply_windowed(arr)

        return smoothed

    def fit_transform_frames(
        self,
        frames_data: List[Dict[str, Any]],
        joint_names: Optional[List[str]] = None,
    ) -> List[Dict[str, Any]]:
        """
        High-level convenience: accept app‑style frame dicts, return
        smoothed frame dicts.

        Parameters
        ----------
        frames_data : list of dict
            Raw per-frame keypoint dicts.
        joint_names : list of str, optional
            Ordered joint names.

        Returns
        -------
        list of dict
            Deep-copied frame dicts with smoothed coordinates.
        """
        if joint_names is None:
            joint_names = COCO_KEYPOINTS
        self._joint_names = list(joint_names)

        arr = self.keypoints_to_array(frames_data, joint_names)
        smoothed = self.fit_transform(arr)
        return self.array_to_keypoints(smoothed, frames_data, joint_names)

    # --- internal steps ------------------------------------------------

    def _mask_low_confidence(self, arr: np.ndarray) -> np.ndarray:
        """Set coordinates to NaN where confidence < *min_confidence*."""
        arr = arr.copy()
        conf = arr[:, :, 2]
        low = conf < self.min_confidence
        arr[low, 0] = np.nan
        arr[low, 1] = np.nan
        return arr

    def _fill_missing(self, signal_1d: np.ndarray) -> np.ndarray:
        """
        Fill NaN values in a 1-D signal.

        Returns a copy with NaNs replaced according to *fill_method*.
        """
        s = np.asarray(signal_1d, dtype=np.float64).copy()
        n = len(s)
        valid = np.isfinite(s)

        if valid.all():
            return s

        if self.fill_method == "forward":
            # Forward fill
            last = np.nan
            for i in range(n):
                if np.isfinite(s[i]):
                    last = s[i]
                elif not np.isnan(last):
                    s[i] = last
            # Backward fill for leading NaN
            first = np.nan
            for i in range(n - 1, -1, -1):
                if np.isfinite(s[i]):
                    first = s[i]
                elif not np.isnan(first):
                    s[i] = first
            return s

        if self.fill_method == "spline":
            _ensure_scipy("spline fill")
            idx = np.arange(n)
            if valid.sum() < 3:
                # Not enough points for cubic spline → fall back to linear
                return self._fill_linear(s, idx, valid)
            try:
                spl = _scipy_interpolate.UnivariateSpline(
                    idx[valid], s[valid], s=0, ext="const"
                )
                s[~valid] = spl(idx[~valid])
            except Exception:
                s = self._fill_linear(s, idx, valid)
            return s

        # Default: linear
        idx = np.arange(n)
        return self._fill_linear(s, idx, valid)

    @staticmethod
    def _fill_linear(s: np.ndarray, idx: np.ndarray, valid: np.ndarray) -> np.ndarray:
        """Linear interpolation (also handles edge extrapolation)."""
        n = len(s)
        s_filled = s.copy()
        if valid.sum() >= 2:
            s_filled[~valid] = np.interp(idx[~valid], idx[valid], s[valid])
        elif valid.sum() == 1:
            s_filled[~valid] = s[valid][0]
        else:
            s_filled[:] = 0.0
        return s_filled

    # --- strategy implementations --------------------------------------

    def _apply_windowed(self, arr: np.ndarray) -> np.ndarray:
        """scipy-based sliding-window filters."""
        _ensure_scipy(self.strategy.value)
        result = arr.copy()
        ws = self._effective_window()

        for j in range(self._n_joints):
            for c in [0, 1]:  # x, y
                sig = result[:, j, c]
                valid = np.isfinite(sig)
                if not valid.any():
                    continue

                # Fill gaps first
                sig_filled = self._fill_missing(sig)

                if self.strategy == SmoothingStrategy.MOVING_AVERAGE:
                    kernel = np.ones(ws) / ws
                    smoothed = np.convolve(sig_filled, kernel, mode="same")
                elif self.strategy == SmoothingStrategy.GAUSSIAN:
                    # Create Gaussian kernel
                    ax = np.arange(-(ws // 2), ws // 2 + 1)
                    kernel = np.exp(-0.5 * (ax / self.sigma) ** 2)
                    kernel /= kernel.sum()
                    smoothed = np.convolve(sig_filled, kernel, mode="same")
                elif self.strategy == SmoothingStrategy.MEDIAN:
                    smoothed = ndimage.median_filter(sig_filled, size=ws)
                elif self.strategy == SmoothingStrategy.SAVITZKY_GOLAY:
                    if ws >= len(sig_filled):
                        ws = len(sig_filled) if len(sig_filled) % 2 == 1 else len(sig_filled) - 1
                    if ws <= self.poly_order:
                        ws = self.poly_order + 2
                    if ws % 2 == 0:
                        ws += 1
                    try:
                        smoothed = _scipy_signal.savgol_filter(
                            sig_filled, ws, self.poly_order, mode="nearest"
                        )
                    except Exception:
                        smoothed = sig_filled
                else:
                    smoothed = sig_filled

                # Restore original NaN positions so downstream code
                # knows which points were originally missing.
                smoothed[~valid] = np.nan
                result[:, j, c] = smoothed

        return result

    def _apply_ema(self, arr: np.ndarray) -> np.ndarray:
        """Exponential moving average (online-capable)."""
        result = arr.copy()
        alpha = self.alpha
        for j in range(self._n_joints):
            for c in [0, 1]:
                sig = arr[:, j, c]
                out = np.empty_like(sig)
                ema = np.nan
                for i in range(len(sig)):
                    if np.isfinite(sig[i]):
                        if np.isnan(ema):
                            ema = sig[i]
                        else:
                            ema = alpha * sig[i] + (1 - alpha) * ema
                    out[i] = ema
                result[:, j, c] = out
        return result

    def _apply_kalman(self, arr: np.ndarray) -> np.ndarray:
        """Per-joint, per-coordinate Kalman filter (forward pass)."""
        result = arr.copy()
        kf = KalmanFilter1D(self.process_noise, self.measurement_noise)
        for j in range(self._n_joints):
            for c in [0, 1]:
                kf.reset()
                # Initialise with first valid point (if any)
                initialized = False
                for i in range(self._n_frames):
                    val = arr[i, j, c]
                    if np.isfinite(val):
                        kf.reset(float(val))
                        initialized = True
                        break
                if not initialized:
                    result[:, j, c] = np.nan
                    continue
                result[0, j, c] = kf.position
                for i in range(1, self._n_frames):
                    meas = arr[i, j, c]
                    pos = kf.update(
                        float(meas) if np.isfinite(meas) else None
                    )
                    result[i, j, c] = pos
        return result

    def _apply_spline(self, arr: np.ndarray) -> np.ndarray:
        """
        Fit a cubic smoothing spline through high-confidence points.

        Low-confidence points are excluded from the fit and replaced
        by the spline evaluation.
        """
        _ensure_scipy("spline")
        result = arr.copy()
        n = self._n_frames
        idx = np.arange(n, dtype=np.float64)

        for j in range(self._n_joints):
            for c in [0, 1]:
                sig = arr[:, j, c]
                valid = np.isfinite(sig)
                if valid.sum() < 3:
                    result[:, j, c] = self._fill_missing(sig)
                    continue

                try:
                    spl = _scipy_interpolate.UnivariateSpline(
                        idx[valid], sig[valid], s=len(valid) * 0.5
                    )
                    result[:, j, c] = spl(idx)
                except Exception:
                    result[:, j, c] = self._fill_missing(sig)

        return result

    def _apply_hybrid(self, arr: np.ndarray) -> np.ndarray:
        """
        Hybrid pipeline:

        1. Detect positional outliers (velocity or z-score).
        2. Mask outliers + low-confidence points → NaN.
        3. Interpolate NaN gaps.
        4. Apply Savitzky-Golay smoothing.
        """
        result = arr.copy()

        for j in range(self._n_joints):
            for c in [0, 1]:
                sig = arr[:, j, c]
                conf = arr[:, j, 2]

                # Step 1 – outlier detection
                if self.outlier_method == "zscore":
                    outlier = detect_outliers_zscore(
                        sig, self.outlier_threshold, self.min_confidence, conf
                    )
                else:
                    outlier = detect_outliers_velocity(
                        sig, self.outlier_threshold, self.min_confidence, conf
                    )

                # Step 2 – mask
                sig_clean = sig.copy()
                sig_clean[outlier] = np.nan
                # low-confidence already masked by _mask_low_confidence

                # Step 3 – interpolate
                sig_filled = self._fill_missing(sig_clean)

                # Step 4 – Savitzky-Golay
                _ensure_scipy("savitzky_golay")
                ws = self._effective_window()
                if ws >= len(sig_filled):
                    ws = len(sig_filled) if len(sig_filled) % 2 == 1 else len(sig_filled) - 1
                if ws <= self.poly_order:
                    ws = self.poly_order + 2
                if ws % 2 == 0:
                    ws += 1
                try:
                    smoothed = _scipy_signal.savgol_filter(
                        sig_filled, ws, self.poly_order, mode="nearest"
                    )
                except Exception:
                    smoothed = sig_filled

                # Restore NaN for originally completely missing frames
                orig_missing = ~np.isfinite(sig) & ~outlier
                smoothed[orig_missing] = np.nan
                result[:, j, c] = smoothed

        return result

    def _effective_window(self) -> int:
        """Clamp window size to available frames and ensure odd."""
        ws = min(self.window_size, self._n_frames)
        if ws % 2 == 0:
            ws -= 1
        return max(ws, 3)


# ===================================================================
# High-level convenience function
# ===================================================================


def smooth_pose_sequence(
    frames_data: List[Dict[str, Any]],
    strategy: Union[str, SmoothingStrategy] = SmoothingStrategy.HYBRID,
    joint_names: Optional[List[str]] = None,
    **kwargs: Any,
) -> List[Dict[str, Any]]:
    """
    Smooth an entire pose sequence with a single call.

    Parameters
    ----------
    frames_data : list of dict
        Per-frame keypoint dicts in the format produced by
        ``extract_joint_positions_from_movenet()`` in ``app.py``.
    strategy : str or SmoothingStrategy
        Smoothing strategy to use (default: ``"hybrid"``).
    joint_names : list of str, optional
        Ordered joint names (defaults to COCO 17).
    **kwargs
        Passed through to :class:`PoseInterpolator` (window_size,
        alpha, outlier_threshold, …).

    Returns
    -------
    list of dict
        Deep copy of *frames_data* with smoothed (x, y) coordinates.

    Examples
    --------
    >>> # Quick hybrid smoothing (recommended for animations)
    >>> smoothed = smooth_pose_sequence(all_keypoints, strategy="hybrid")

    >>> # Light EMA for near-real-time use
    >>> smoothed = smooth_pose_sequence(all_keypoints, strategy="exponential",
    ...                                 alpha=0.15)

    >>> # Strong outlier removal for noisy recordings
    >>> smoothed = smooth_pose_sequence(all_keypoints, strategy="hybrid",
    ...                                 outlier_method="zscore",
    ...                                 outlier_threshold=2.5,
    ...                                 window_size=7)
    """
    interpolator = PoseInterpolator(strategy=strategy, **kwargs)
    return interpolator.fit_transform_frames(frames_data, joint_names=joint_names)


# ===================================================================
# Smoke test (runs when module is executed directly)
# ===================================================================

if __name__ == "__main__":
    # Generate a synthetic trajectory with gaps and spikes
    np.random.seed(42)
    n_frames = 100
    n_joints = 3  # nose, left_shoulder, right_shoulder
    t = np.linspace(0, 4 * np.pi, n_frames)

    # Ground truth – smooth sinusoid
    true_x = np.sin(t) * 0.3 + 0.5
    true_y = np.cos(t) * 0.2 + 0.5

    # Build array: (F, J, 3)
    raw = np.zeros((n_frames, n_joints, 3), dtype=np.float64)
    for j in range(n_joints):
        raw[:, j, 0] = true_x + np.random.randn(n_frames) * 0.02
        raw[:, j, 1] = true_y + np.random.randn(n_frames) * 0.02
        raw[:, j, 2] = 0.9  # high confidence

    # Inject outliers
    raw[20, 0, 0] += 0.4  # spike
    raw[50, 0, 1] -= 0.3
    raw[75, 0, 0] += 0.5

    # Inject gaps
    raw[40:45, 1, :] = np.nan
    raw[60:65, 1, :] = np.nan
    raw[80, 1, :] = np.nan

    # --- Test each strategy -------------------------------------------
    strategies = [
        SmoothingStrategy.HYBRID,
        SmoothingStrategy.MOVING_AVERAGE,
        SmoothingStrategy.GAUSSIAN,
        SmoothingStrategy.EXPONENTIAL,
        SmoothingStrategy.MEDIAN,
        SmoothingStrategy.SAVITZKY_GOLAY,
        SmoothingStrategy.KALMAN,
        SmoothingStrategy.SPLINE,
    ]

    print(f"{'Strategy':<22s} {'MAE (x)':>10s} {'MAE (y)':>10s}")
    print("-" * 44)

    for strat in strategies:
        interp = PoseInterpolator(strategy=strat)
        smoothed = interp.fit_transform(raw.copy())

        # Mean absolute error against ground truth (only first joint)
        mae_x = np.nanmean(np.abs(smoothed[:, 0, 0] - true_x))
        mae_y = np.nanmean(np.abs(smoothed[:, 0, 1] - true_y))
        print(f"{strat.value:<22s} {mae_x:10.6f} {mae_y:10.6f}")

    # --- Test high-level convenience ----------------------------------
    frames_data = PoseInterpolator.array_to_keypoints(
        raw,
        [
            {
                "poses": [
                    {
                        "pose_id": 0,
                        "keypoints": [
                            {
                                "x": raw[i, j, 0],
                                "y": raw[i, j, 1],
                                "score": raw[i, j, 2],
                                "name": COCO_KEYPOINTS[j],
                            }
                            for j in range(n_joints)
                        ],
                    }
                ],
                "frame_id": i,
            }
            for i in range(n_frames)
        ],
        joint_names=COCO_KEYPOINTS[:n_joints],
    )

    smoothed_frames = smooth_pose_sequence(frames_data, strategy="hybrid")
    print(f"\nHigh-level convenience: processed {len(smoothed_frames)} frames ✓")

    # Convert to DataFrame for A11 compatibility
    df = PoseInterpolator.array_to_dataframe(
        raw, joint_names=COCO_KEYPOINTS[:n_joints]
    )
    print(f"DataFrame conversion: {df.shape} ✓")
    print("\nAll tests passed.")