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
| import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple
class MDERNet(nn.Module):
"""
MDERNet: 多模态驾驶员情感识别网络
=================================
论文核心架构
分支1: 面部表情分支
- 输入: 面部视频帧序列 [B, T, 3, H, W]
- 编码: 3D-CNN → 时序特征
- 输出: 面部情感特征 [B, D]
分支2: 驾驶行为分支
- 输入: 驾驶行为序列 [B, T, 4] (刹车, 油门, 方向盘, 速度)
- 编码: 1D-CNN + LSTM → 时序特征
- 输出: 驾驶行为特征 [B, D]
融合层: 中层级特征融合
- 方式: 双线性融合 + 交叉注意力
- 输出: 融合特征 [B, D]
分类头: 情感分类
- 输出: 7类情感 logits [B, 7]
Args:
face_dim: 面部特征维度
behavior_dim: 驾驶行为特征维度
fusion_dim: 融合后维度
num_emotions: 情感类别数
"""
def __init__(
self,
face_dim: int = 512,
behavior_dim: int = 128,
fusion_dim: int = 256,
num_emotions: int = 7,
num_behavior_features: int = 4,
sequence_length: int = 30
):
super().__init__()
self.face_branch = FaceExpressionBranch(
output_dim=face_dim,
sequence_length=sequence_length
)
self.behavior_branch = DrivingBehaviorBranch(
input_dim=num_behavior_features,
output_dim=behavior_dim,
sequence_length=sequence_length
)
self.fusion = MidLevelFusion(
face_dim=face_dim,
behavior_dim=behavior_dim,
output_dim=fusion_dim
)
self.classifier = nn.Sequential(
nn.Linear(fusion_dim, fusion_dim),
nn.LayerNorm(fusion_dim),
nn.GELU(),
nn.Dropout(0.2),
nn.Linear(fusion_dim, num_emotions)
)
def forward(
self,
face_frames: torch.Tensor,
behavior_data: torch.Tensor
) -> dict:
"""
前向传播
Args:
face_frames: [B, T, 3, H, W] 面部视频
behavior_data: [B, T, 4] 驾驶行为序列
Returns:
emotions: [B, 7] 情感logits
face_feat: [B, face_dim] 面部特征
behavior_feat: [B, behavior_dim] 行为特征
fused_feat: [B, fusion_dim] 融合特征
"""
face_feat = self.face_branch(face_frames)
behavior_feat = self.behavior_branch(behavior_data)
fused_feat = self.fusion(face_feat, behavior_feat)
emotions = self.classifier(fused_feat)
return {
"emotions": emotions,
"face_feat": face_feat,
"behavior_feat": behavior_feat,
"fused_feat": fused_feat
}
class FaceExpressionBranch(nn.Module):
"""面部表情分支: 3D-CNN"""
def __init__(self, output_dim: int = 512, sequence_length: int = 30):
super().__init__()
self.conv1 = nn.Conv3d(3, 32, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1))
self.bn1 = nn.BatchNorm3d(32)
self.conv2 = nn.Conv3d(32, 64, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1))
self.bn2 = nn.BatchNorm3d(64)
self.conv3 = nn.Conv3d(64, 128, (3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1))
self.bn3 = nn.BatchNorm3d(128)
self.pool = nn.AdaptiveAvgPool3d((1, 1, 1))
self.fc = nn.Linear(128, output_dim)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Args:
x: [B, T, 3, H, W] → [B, 3, T, H, W]
"""
x = x.permute(0, 2, 1, 3, 4)
x = F.relu(self.bn1(self.conv1(x)))
x = F.relu(self.bn2(self.conv2(x)))
x = F.relu(self.bn3(self.conv3(x)))
x = self.pool(x).flatten(1)
return self.fc(x)
class DrivingBehaviorBranch(nn.Module):
"""驾驶行为分支: 1D-CNN + LSTM"""
def __init__(self, input_dim: int = 4, output_dim: int = 128, sequence_length: int = 30):
super().__init__()
self.conv1 = nn.Conv1d(input_dim, 32, 3, padding=1)
self.conv2 = nn.Conv1d(32, 64, 3, padding=1)
self.lstm = nn.LSTM(64, output_dim, batch_first=True, num_layers=2)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Args:
x: [B, T, 4] 驾驶行为序列
"""
x = x.permute(0, 2, 1)
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = x.permute(0, 2, 1)
out, (h, c) = self.lstm(x)
return out[:, -1, :]
class MidLevelFusion(nn.Module):
"""中层级特征融合"""
def __init__(self, face_dim: int, behavior_dim: int, output_dim: int):
super().__init__()
self.face_proj = nn.Linear(face_dim, output_dim)
self.behavior_proj = nn.Linear(behavior_dim, output_dim)
self.cross_attn_f2b = nn.MultiheadAttention(output_dim, 4, batch_first=True)
self.cross_attn_b2f = nn.MultiheadAttention(output_dim, 4, batch_first=True)
self.gate = nn.Sequential(
nn.Linear(output_dim * 2, output_dim),
nn.Sigmoid()
)
self.output = nn.Sequential(
nn.Linear(output_dim, output_dim),
nn.LayerNorm(output_dim),
nn.GELU()
)
def forward(self, face_feat: torch.Tensor, behavior_feat: torch.Tensor) -> torch.Tensor:
f = self.face_proj(face_feat).unsqueeze(1)
b = self.behavior_proj(behavior_feat).unsqueeze(1)
f_enhanced, _ = self.cross_attn_f2b(f, b, b)
b_enhanced, _ = self.cross_attn_b2f(b, f, f)
f_out = f_enhanced.squeeze(1)
b_out = b_enhanced.squeeze(1)
gate = self.gate(torch.cat([f_out, b_out], dim=-1))
fused = gate * f_out + (1 - gate) * b_out
return self.output(fused)
if __name__ == "__main__":
model = MDERNet(
face_dim=512,
behavior_dim=128,
fusion_dim=256,
num_emotions=7
)
face = torch.randn(4, 30, 3, 64, 64)
behavior = torch.randn(4, 30, 4)
out = model(face, behavior)
total_params = sum(p.numel() for p in model.parameters())
print(f"MDERNet 双分支情感识别")
print(f"输入: face={face.shape}, behavior={behavior.shape}")
print(f"输出: {out['emotions'].shape}")
print(f"总参数: {total_params:,} ({total_params/1e6:.2f}M)")
print(f"面部特征: {out['face_feat'].shape}")
print(f"行为特征: {out['behavior_feat'].shape}")
print(f"融合特征: {out['fused_feat'].shape}")
|
