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
| """
Everybody Tracking Every Body - 多视角自我中心 3D 姿态融合
核心思路：
1. 每人佩戴摄像头 → VIO SLAM 获得头部运动轨迹
2. 第一人称视角 → 推断自身身体姿态
3. 第三人称视角（看到他人）→ 推断他人姿态
4. 扩散模型融合 → 基于头运动+外向观测
关键挑战：
- 外向观测稀疏、间歇、可靠性变化大
- 需要同时处理内容和可靠性
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class EgocentricPoseEncoder(nn.Module):
"""
自我中心视角姿态编码器
从第一人称视频提取自身身体姿态估计
"""
def __init__(self, feat_dim: int = 256):
super().__init__()
self.visual_encoder = nn.Sequential(
nn.Conv2d(3, 32, 3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(32, 64, 3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(64, 128, 3, stride=2, padding=1),
nn.ReLU(),
nn.AdaptiveAvgPool2d((4, 4)),
nn.Flatten(),
nn.Linear(128 * 16, feat_dim)
)
def forward(self, ego_frame: torch.Tensor) -> torch.Tensor:
return self.visual_encoder(ego_frame)
class ExocentricPoseEncoder(nn.Module):
"""
外向视角（看到他人）姿态编码器
从第三人称视角提取他人姿态估计
包含可靠性估计
"""
def __init__(self, feat_dim: int = 256):
super().__init__()
self.visual_encoder = nn.Sequential(
nn.Conv2d(3, 32, 3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(32, 64, 3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(64, 128, 3, stride=2, padding=1),
nn.ReLU(),
nn.AdaptiveAvgPool2d((4, 4)),
nn.Flatten(),
nn.Linear(128 * 16, feat_dim)
)
self.reliability_head = nn.Sequential(
nn.Linear(feat_dim, 64),
nn.ReLU(),
nn.Linear(64, 1),
nn.Sigmoid()
)
def forward(self, exo_frame: torch.Tensor) -> tuple:
feat = self.visual_encoder(exo_frame)
reliability = self.reliability_head(feat)
return feat, reliability
class DiffusionFusionModel(nn.Module):
"""
扩散模型融合
将头部运动先验 + 第一人称 + 第三人称观测
通过扩散过程融合为最终姿态
论文核心：条件于观测内容和可靠性
"""
def __init__(self, pose_dim: int = 17 * 3, feat_dim: int = 256):
super().__init__()
self.pose_dim = pose_dim
self.ego_encoder = nn.Linear(feat_dim, 128)
self.exo_encoder = nn.Linear(feat_dim, 128)
self.head_motion_encoder = nn.Linear(6, 128)
self.denoise_net = nn.Sequential(
nn.Linear(pose_dim + 128 * 3, 512),
nn.ReLU(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, pose_dim)
)
def forward(self, noisy_pose: torch.Tensor, t: torch.Tensor,
ego_feat: torch.Tensor, exo_feat: torch.Tensor,
head_motion: torch.Tensor) -> torch.Tensor:
"""
去噪一步
Args:
noisy_pose: (B, pose_dim) 加噪的姿态
t: (B,) 时间步
ego_feat: (B, 128) 自我中心特征
exo_feat: (B, 128) 外向观测特征
head_motion: (B, 6) 头部运动
"""
ego_cond = self.ego_encoder(ego_feat)
exo_cond = self.exo_encoder(exo_feat)
head_cond = self.head_motion_encoder(head_motion)
x = torch.cat([noisy_pose, ego_cond, exo_cond, head_cond], dim=-1)
return self.denoise_net(x)
class MultiPersonPoseTracker:
"""
多人姿态追踪系统
IMS 远期应用：座舱内多乘员姿态追踪
（不是每人佩戴摄像头，而是车内多个摄像头角度）
"""
def __init__(self, num_persons: int = 4, num_cameras: int = 3):
self.num_persons = num_persons
self.num_cameras = num_cameras
self.exo_encoders = nn.ModuleList([
ExocentricPoseEncoder() for _ in range(num_cameras)
])
self.fusion = DiffusionFusionModel()
def estimate_pose(self, camera_frames: list, head_motion: torch.Tensor):
"""
从多个摄像头角度估计所有乘员姿态
Args:
camera_frames: 每个摄像头的图像
head_motion: 驾驶员头部运动（从 IMU/SLAM）
"""
all_feats = []
for i, frame in enumerate(camera_frames):
feat, reliability = self.exo_encoders[i](frame)
all_feats.append((feat, reliability))
total_feat = sum(f * r for f, r in all_feats) / len(all_feats)
noisy_pose = torch.randn(1, 51)
for t in range(10, 0, -1):
with torch.no_grad():
noise = self.fusion(
noisy_pose, torch.tensor([t]),
total_feat, total_feat, head_motion
)
noisy_pose = noisy_pose - noise * 0.1
return noisy_pose
if __name__ == "__main__":
tracker = MultiPersonPoseTracker(num_persons=4, num_cameras=3)
frames = [torch.randn(1, 3, 224, 224) for _ in range(3)]
head_motion = torch.randn(1, 6)
pose = tracker.estimate_pose(frames, head_motion)
print(f"输出姿态: {pose.shape} (17 joints × 3D)")
print(f"参数量: {sum(p.numel() for p in tracker.parameters()):,}")
|
