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
| """
Self-Distillation Transformer (SDT) + 三项改进
基础: 文本 + 音频 + 视觉 → intra-modal Transformer → inter-modal Transformer
→ sigmoid 门 → softmax 门 → 分类
改进1: 视觉 = 外观(ViT) + 几何(FACS)
改进2: softmax 门 → 类别自适应门
改进3: + Valence-Arousal 先验
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, Tuple, List
import numpy as np
class GeometryEnhancedVisualEncoder(nn.Module):
"""
几何增强视觉编码器
论文核心改进1: 外观特征 + 面部几何描述符
外观: ViT 提取全局面部表征
几何: FACS Action Unit 描述面部结构变化
"""
def __init__(self, appearance_dim: int = 768,
geometry_dim: int = 51,
hidden_dim: int = 256):
super().__init__()
self.appearance_proj = nn.Linear(appearance_dim, hidden_dim)
self.geometry_encoder = nn.Sequential(
nn.Linear(geometry_dim, 128),
nn.ReLU(),
nn.Linear(128, hidden_dim),
nn.ReLU()
)
self.fusion = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.GELU()
)
def forward(self, appearance_feat: torch.Tensor,
geometry_feat: torch.Tensor) -> torch.Tensor:
"""
Args:
appearance_feat: ViT 特征, shape=(B, T, 768)
geometry_feat: FACS AU, shape=(B, T, 51)
Returns:
visual_repr: 增强视觉表征, shape=(B, T, hidden_dim)
"""
app = self.appearance_proj(appearance_feat)
geo = self.geometry_encoder(geometry_feat)
return self.fusion(torch.cat([app, geo], dim=-1))
class ClassWiseAdaptiveFusion(nn.Module):
"""
类别自适应模态融合
论文核心改进2: 不同情绪类别下，文本/音频/视觉权重不同
例如:
- 愤怒: 音频权重高（语气强烈）
- 悲伤: 视觉权重高（面部表情）
- 中性: 文本权重高
"""
def __init__(self, n_modalities: int = 3, n_classes: int = 7,
hidden_dim: int = 256):
super().__init__()
self.class_modality_weights = nn.Parameter(
torch.randn(n_classes, n_modalities) * 0.1 + 1.0
)
def forward(self, modalities: Dict[str, torch.Tensor],
class_logits: torch.Tensor) -> torch.Tensor:
"""
Args:
modalities: {'text': (B,T,D), 'audio': (B,T,D), 'visual': (B,T,D)}
class_logits: 当前预测 logits, shape=(B, n_classes)
Returns:
fused: 融合表征, shape=(B, T, D)
"""
class_probs = F.softmax(class_logits, dim=-1)
weights = torch.einsum('bc,cm->bm',
class_probs,
self.class_modality_weights)
weights = F.softmax(weights, dim=-1)
mod_tensor = torch.stack([
modalities['text'],
modalities['audio'],
modalities['visual']
], dim=1)
weights = weights.unsqueeze(-1).unsqueeze(-1)
fused = (mod_tensor * weights).sum(dim=1)
return fused
class ValenceArousalPrior(nn.Module):
"""
Valence-Arousal 先验
论文核心改进3: 利用情感在 VA 空间的结构
情感转移 utterance (如 sad→joy) 分类困难
VA 先验: 将类别映射到 VA 空间，对转移 utterance 施加修正
"""
EMOTION_VA_CENTERS = {
'neutral': (0.0, 0.0),
'happy': (0.8, 0.5),
'sad': (-0.7, -0.4),
'angry': (0.2, 0.8),
'fearful': (-0.6, 0.7),
'disgust': (-0.5, 0.3),
'surprised': (0.6, 0.9),
}
def __init__(self, n_classes: int = 7, va_dim: int = 2):
super().__init__()
self.va_projection = nn.Linear(n_classes, va_dim)
self.shift_correction = nn.Linear(va_dim * 2, n_classes)
def forward(self, class_logits: torch.Tensor,
prev_logits: torch.Tensor = None) -> torch.Tensor:
"""
Args:
class_logits: 当前 utterance logits, shape=(B, C)
prev_logits: 上一 utterance logits (情感转移), shape=(B, C)
Returns:
corrected_logits: 修正后 logits
"""
if prev_logits is None:
return class_logits
curr_va = self.va_projection(class_logits)
prev_va = self.va_projection(prev_logits)
va_shift = torch.cat([curr_va, prev_va], dim=-1)
correction = self.shift_correction(va_shift)
return class_logits + 0.1 * correction
class EnhancedSDT(nn.Module):
"""
增强版 SDT: 三项改进集成
完整管道:
1. 文本编码 (RoBERTa)
2. 音频编码 (WavLM)
3. 视觉编码 (ViT + FACS) ← 改进1
4. Intra-modal Transformer
5. Inter-modal Transformer
6. Sigmoid 门 → 模态增强
7. 类别自适应融合 ← 改进2
8. 分类 + VA 先验修正 ← 改进3
"""
def __init__(self, n_classes: int = 7, hidden_dim: int = 256):
super().__init__()
self.n_classes = n_classes
self.text_encoder = nn.Linear(1024, hidden_dim)
self.audio_encoder = nn.Linear(1024, hidden_dim)
self.visual_encoder = GeometryEnhancedVisualEncoder(
appearance_dim=768, geometry_dim=51, hidden_dim=hidden_dim
)
self.intra_modal = nn.TransformerEncoder(
nn.TransformerEncoderLayer(hidden_dim, nhead=4, batch_first=True),
num_layers=2
)
self.cross_attn = nn.MultiheadAttention(hidden_dim, 4, batch_first=True)
self.adaptive_fusion = ClassWiseAdaptiveFusion(
n_modalities=3, n_classes=n_classes, hidden_dim=hidden_dim
)
self.classifier = nn.Linear(hidden_dim, n_classes)
self.va_prior = ValenceArousalPrior(n_classes=n_classes)
def forward(self, text_feat, audio_feat,
appearance_feat, geometry_feat,
prev_logits=None) -> Dict[str, torch.Tensor]:
"""
Args:
text_feat: (B, T, 1024)
audio_feat: (B, T, 1024)
appearance_feat: (B, T, 768) ViT
geometry_feat: (B, T, 51) FACS AU
prev_logits: (B, C) 上一轮分类
Returns:
logits, fused_repr
"""
t = self.text_encoder(text_feat)
a = self.audio_encoder(audio_feat)
v = self.visual_encoder(appearance_feat, geometry_feat)
t = self.intra_modal(t)
a = self.intra_modal(a)
v = self.intra_modal(v)
t_attended, _ = self.cross_attn(t, torch.cat([a, v], dim=1), torch.cat([a, v], dim=1))
a_attended, _ = self.cross_attn(a, torch.cat([t, v], dim=1), torch.cat([t, v], dim=1))
v_attended, _ = self.cross_attn(v, torch.cat([t, a], dim=1), torch.cat([t, a], dim=1))
pooled = torch.cat([t_attended.mean(1), a_attended.mean(1),
v_attended.mean(1)], dim=-1)
initial_logits = self.classifier(pooled[:, :256])
fused = self.adaptive_fusion({
'text': t_attended, 'audio': a_attended, 'visual': v_attended
}, initial_logits)
logits = self.classifier(fused.mean(1))
if prev_logits is not None:
logits = self.va_prior(logits, prev_logits)
return {'logits': logits, 'fused_repr': fused}
class DriverEmotionDetector:
"""
IMS 驾驶员情绪检测器
基于 ECCV 2026 论文方法
应用场景:
1. 路怒症检测 (angry + 高 arousal)
2. 疲劳情绪 (sad + 低 arousal)
3. 焦虑检测 (fearful + 中 arousal)
4. 满意/愉悦 (happy + 中 arousal)
输入: DMS 摄像头 + 麦克风
"""
def __init__(self):
self.model = EnhancedSDT(n_classes=7)
self.emotion_actions = {
'angry': {'risk': 3, 'action': '路怒警告', 'threshold': 0.7},
'sad': {'risk': 2, 'action': '疲劳情绪', 'threshold': 0.6},
'fearful': {'risk': 2, 'action': '焦虑检测', 'threshold': 0.6},
'happy': {'risk': 0, 'action': '正常', 'threshold': 0.5},
'neutral': {'risk': 0, 'action': '正常', 'threshold': 0.5},
'disgust': {'risk': 1, 'action': '不适', 'threshold': 0.6},
'surprised': {'risk': 1, 'action': '惊讶', 'threshold': 0.7},
}
def classify(self, appearance_feat, geometry_feat,
audio_feat, text_feat=None,
prev_emotion=None) -> dict:
"""
分类驾驶员情绪
Args:
appearance_feat: 面部 ViT 特征 (B, T, 768)
geometry_feat: FACS AU (B, T, 51) — 51 个 Action Unit
audio_feat: 语音特征 (B, T, 1024)
text_feat: 转录文本特征 (可选)
prev_emotion: 上一帧情绪 logits
"""
if text_feat is None:
text_feat = torch.zeros(appearance_feat.shape[0],
appearance_feat.shape[1], 1024)
output = self.model(text_feat, audio_feat,
appearance_feat, geometry_feat,
prev_emotion)
logits = output['logits']
probs = F.softmax(logits, dim=-1)
emotions = list(self.emotion_actions.keys())
risk_scores = torch.zeros_like(probs)
for i, emo in enumerate(emotions):
risk_scores[:, i] = self.emotion_actions[emo]['risk']
total_risk = (probs * risk_scores).sum(dim=-1)
return {
'emotions': {emotions[i]: probs[0, i].item()
for i in range(len(emotions))},
'dominant_emotion': emotions[probs[0].argmax().item()],
'risk_score': total_risk[0].item(),
'logits': logits
}
if __name__ == "__main__":
print("=== 增强版 SDT 测试 ===")
model = EnhancedSDT(n_classes=7, hidden_dim=256)
B, T = 2, 5
text_feat = torch.randn(B, T, 1024)
audio_feat = torch.randn(B, T, 1024)
appearance_feat = torch.randn(B, T, 768)
geometry_feat = torch.randn(B, T, 51)
output = model(text_feat, audio_feat, appearance_feat, geometry_feat)
print(f"输入: text={text_feat.shape}, audio={audio_feat.shape}")
print(f" appearance={appearance_feat.shape}, geometry={geometry_feat.shape}")
print(f"输出: logits={output['logits'].shape}")
prev_logits = torch.randn(B, 7)
output_shift = model(text_feat, audio_feat, appearance_feat,
geometry_feat, prev_logits)
print(f"VA 先验: 转移 utterance logits={output_shift['logits'].shape}")
detector = DriverEmotionDetector()
result = detector.classify(appearance_feat, geometry_feat, audio_feat)
print(f"\n=== 驾驶员情绪检测结果 ===")
print(f"主导情绪: {result['dominant_emotion']}")
print(f"风险评分: {result['risk_score']:.2f}")
for emo, prob in sorted(result['emotions'].items(),
key=lambda x: -x[1]):
print(f" {emo}: {prob:.1%}")
print(f"\n=== 论文性能报告 ===")
print(f"{'数据集':<12} {'改进1(几何)':<15} {'改进2(融合)':<15} {'改进3(VA)'}")
print(f"{'MELD':<12} {'+0.27 F1':<15} {'+0.17 F1':<15} {'+0.30 acc'}")
print(f"{'IEMOCAP':<12} {'+4.36 F1':<15} {'+0.25 F1':<15} {'+0.74 acc'}")
print(f"\n关键: 三项改进互补, 可叠加使用")
|