File size: 4,875 Bytes
ae419ed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import math
from typing import Iterable

import numpy as np


NUM_JOINTS = 17

COCO_BONES: list[tuple[int, int]] = [
    (0, 1), (0, 2), (1, 3), (2, 4),
    (5, 6), (5, 7), (7, 9), (6, 8), (8, 10),
    (5, 11), (6, 12), (11, 12),
    (11, 13), (13, 15), (12, 14), (14, 16),
]

LOWER_BODY = [11, 12, 13, 14, 15, 16]
UPPER_BODY = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]


def normalize_pose(kpts: np.ndarray, eps: float = 1e-6) -> np.ndarray:
    """Normalize COCO keypoints by per-frame visible bounding box."""
    out = np.asarray(kpts, dtype=np.float32).copy()
    xy = out[..., :2]
    conf = out[..., 2]
    for t in range(out.shape[0]):
        valid = conf[t] > 0
        if not np.any(valid):
            out[t, :, :2] = 0
            continue
        pts = xy[t, valid]
        mn = pts.min(axis=0)
        mx = pts.max(axis=0)
        center = (mn + mx) / 2.0
        size = np.maximum(mx - mn, eps)
        out[t, :, 0] = (out[t, :, 0] - center[0]) / size[0]
        out[t, :, 1] = (out[t, :, 1] - center[1]) / size[1]
        out[t, ~valid, :2] = 0
    return out


def resample_or_pad(kpts: np.ndarray, clip_len: int) -> np.ndarray:
    if len(kpts) == clip_len:
        return kpts.astype(np.float32)
    if len(kpts) <= 0:
        return np.zeros((clip_len, NUM_JOINTS, 3), dtype=np.float32)
    if len(kpts) < clip_len:
        pad = np.repeat(kpts[-1:,...], clip_len - len(kpts), axis=0)
        return np.concatenate([kpts, pad], axis=0).astype(np.float32)
    idx = np.linspace(0, len(kpts) - 1, clip_len).round().astype(np.int64)
    return kpts[idx].astype(np.float32)


def make_clips(kpts: np.ndarray, clip_len: int, stride: int) -> list[np.ndarray]:
    if len(kpts) <= clip_len:
        return [resample_or_pad(kpts, clip_len)]
    clips = []
    for start in range(0, len(kpts) - clip_len + 1, stride):
        clips.append(kpts[start:start + clip_len].astype(np.float32))
    if not clips:
        clips.append(resample_or_pad(kpts, clip_len))
    return clips


def bone_features(joint: np.ndarray) -> np.ndarray:
    bone = np.zeros_like(joint, dtype=np.float32)
    for parent, child in COCO_BONES:
        bone[:, child, :2] = joint[:, child, :2] - joint[:, parent, :2]
        bone[:, child, 2] = np.minimum(joint[:, child, 2], joint[:, parent, 2])
    return bone


def temporal_diff(x: np.ndarray) -> np.ndarray:
    diff = np.zeros_like(x, dtype=np.float32)
    diff[1:] = x[1:] - x[:-1]
    return diff


def dynamics_features(joint: np.ndarray) -> np.ndarray:
    xy = joint[..., :2]
    conf = joint[..., 2:3]
    vel = temporal_diff(xy)
    acc = temporal_diff(vel)
    center = weighted_center(xy, conf)
    center_vel = temporal_diff(center)
    torso = torso_angle(xy)
    hip = xy[:, [11, 12], 1].mean(axis=1, keepdims=True)
    hip_drop = temporal_diff(hip)
    aspect = body_aspect_ratio(xy, conf)
    global_dyn = np.concatenate([center_vel, torso, hip_drop, aspect], axis=1)
    global_dyn = np.repeat(global_dyn[:, None, :], NUM_JOINTS, axis=1)
    return np.concatenate([vel, acc, global_dyn], axis=2).astype(np.float32)


def weighted_center(xy: np.ndarray, conf: np.ndarray, eps: float = 1e-6) -> np.ndarray:
    w = np.clip(conf, 0.0, 1.0)
    return (xy * w).sum(axis=1) / (w.sum(axis=1) + eps)


def torso_angle(xy: np.ndarray) -> np.ndarray:
    shoulder = xy[:, [5, 6]].mean(axis=1)
    hip = xy[:, [11, 12]].mean(axis=1)
    vec = shoulder - hip
    angle = np.arctan2(vec[:, 1], vec[:, 0]) / math.pi
    return angle[:, None].astype(np.float32)


def body_aspect_ratio(xy: np.ndarray, conf: np.ndarray, eps: float = 1e-6) -> np.ndarray:
    ratios = []
    visible = conf[..., 0] > 0
    for t in range(xy.shape[0]):
        if not np.any(visible[t]):
            ratios.append([0.0])
            continue
        pts = xy[t, visible[t]]
        wh = pts.max(axis=0) - pts.min(axis=0)
        ratios.append([float(wh[1] / (wh[0] + eps))])
    return np.asarray(ratios, dtype=np.float32)


def mask_keypoints(
    joint: np.ndarray,
    mode: str,
    amount: float = 0.0,
    rng: np.random.Generator | None = None,
) -> np.ndarray:
    rng = rng or np.random.default_rng()
    out = joint.copy()
    if mode == "clean":
        return out
    if mode.startswith("missing"):
        prob = amount
        mask = rng.random(out.shape[:2]) < prob
        out[mask] = 0
    elif mode == "lower_body":
        out[:, LOWER_BODY] = 0
    elif mode == "upper_body":
        out[:, UPPER_BODY] = 0
    elif mode == "low_conf":
        out[out[..., 2] < 0.5] = 0
    else:
        raise ValueError(f"Unknown robustness mode: {mode}")
    return out.astype(np.float32)


def infer_label_from_path(path: str) -> int:
    parts = [p.lower() for p in path.replace("\\", "/").split("/")]
    positives = {"fall", "falls", "fallen", "positive", "1"}
    return int(any(p in positives for p in parts))