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
| """
EyeCue: 眼动增强第一人称视频认知分心检测
论文：https://arxiv.org/abs/2605.07859
代码：https://github.com/langzhang2000/EyeCue
核心方法：眼动-场景交互建模 + GDSQ跨模态注意力
依赖：
- torch>=2.0
- transformers>=4.30
- opencv-python>=4.8
- numpy>=1.24
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import VideoMAEModel
import numpy as np
from typing import Tuple, Dict
class EyeCue(nn.Module):
"""
EyeCue: 眼动增强第一人称视频认知分心检测框架
架构：
1. 视频编码器：VideoMAE提取时空特征
2. 眼动编码器：BiLSTM建模眼动模式
3. GDSQ模块：眼动驱动的语义查询
4. 分类器：分心/专注二元判断
输入：
- 第一人称视频：(B, T, C, H, W)
- 眼动序列：(B, T_g, 6) [x, y, pupil, fixation_dur, saccade_vel, blink_rate]
输出：
- 分心概率：(B,) 范围[0, 1]
"""
def __init__(self, config: Dict):
super().__init__()
self.video_encoder = VideoEncoder(
model_name=config.get('video_model', 'videomae_base'),
num_frames=config.get('num_frames', 16)
)
video_dim = 768
self.gaze_encoder = GazeEncoder(
input_dim=config.get('gaze_dim', 6),
hidden_dim=config.get('gaze_hidden', 256),
num_layers=config.get('gaze_layers', 2)
)
gaze_dim = 512
self.gdsq = GazeDrivenSemanticQuery(
video_dim=video_dim,
gaze_dim=gaze_dim,
num_heads=config.get('num_heads', 8)
)
self.classifier = nn.Sequential(
nn.Linear(video_dim, 256),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(256, 1)
)
def forward(self,
video: torch.Tensor,
gaze_sequence: torch.Tensor,
gaze_positions: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
前向传播
Args:
video: (B, T_v, C, H, W) 第一人称视频
gaze_sequence: (B, T_g, 6) 眼动特征序列
gaze_positions: (B, T_g, 2) 注视点坐标
Returns:
predictions: (B,) 分心概率
attention_weights: (B, T_g, T_v) 注意力权重
"""
video_features = self.video_encoder(video)
gaze_features = self.gaze_encoder(gaze_sequence)
attended_features, attention_weights = self.gdsq(
video_features, gaze_features, gaze_positions
)
pooled_features = attended_features.mean(dim=1)
predictions = self.classifier(pooled_features).squeeze(-1)
return predictions, attention_weights
class VideoEncoder(nn.Module):
"""视频编码器（VideoMAE）"""
def __init__(self, model_name='MCG-NJU/videomae-base', num_frames=16):
super().__init__()
self.model = VideoMAEModel.from_pretrained(model_name)
self.num_frames = num_frames
def forward(self, video_clip: torch.Tensor) -> torch.Tensor:
"""
Args:
video_clip: (B, T, C, H, W) RGB视频片段
Returns:
features: (B, T, 768) 时空特征
"""
video_clip = video_clip.float() / 255.0
video_clip = (video_clip - 0.5) / 0.5
outputs = self.model(pixel_values=video_clip)
return outputs.last_hidden_state
class GazeEncoder(nn.Module):
"""眼动编码器（双向LSTM）"""
def __init__(self, input_dim=6, hidden_dim=256, num_layers=2):
super().__init__()
self.lstm = nn.LSTM(input_dim, hidden_dim,
num_layers=num_layers,
batch_first=True,
bidirectional=True,
dropout=0.2)
self.fc = nn.Linear(hidden_dim * 2, 512)
def forward(self, gaze_sequence: torch.Tensor) -> torch.Tensor:
"""
Args:
gaze_sequence: (B, T, 6) 眼动特征
Returns:
features: (B, T, 512)
"""
lstm_out, _ = self.lstm(gaze_sequence)
return self.fc(lstm_out)
class GazeDrivenSemanticQuery(nn.Module):
"""眼动驱动的语义查询模块"""
def __init__(self, video_dim=768, gaze_dim=512, num_heads=8):
super().__init__()
self.cross_attention = nn.MultiheadAttention(
embed_dim=video_dim,
num_heads=num_heads,
batch_first=True,
dropout=0.1
)
self.gaze_proj = nn.Linear(gaze_dim, video_dim)
self.norm = nn.LayerNorm(video_dim)
def forward(self,
video_features: torch.Tensor,
gaze_features: torch.Tensor,
gaze_positions: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
video_features: (B, T_v, D)
gaze_features: (B, T_g, D)
gaze_positions: (B, T_g, 2) 归一化坐标
Returns:
attended_features: (B, T_g, D)
attention_weights: (B, T_g, T_v)
"""
query = self.gaze_proj(gaze_features)
key = value = video_features
attended, attention_weights = self.cross_attention(
query, key, value, need_weights=True
)
output = self.norm(query + attended)
return output, attention_weights
if __name__ == "__main__":
"""
实际测试：模拟驾驶数据
"""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
config = {
'video_model': 'MCG-NJU/videomae-base',
'num_frames': 16,
'gaze_dim': 6,
'gaze_hidden': 256,
'gaze_layers': 2,
'num_heads': 8
}
model = EyeCue(config).to(device)
print(f"模型参数量: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M")
batch_size = 4
video = torch.randn(batch_size, 16, 3, 224, 224).to(device)
gaze_sequence = torch.randn(batch_size, 30, 6).to(device)
gaze_positions = torch.rand(batch_size, 30, 2).to(device)
model.eval()
with torch.no_grad():
predictions, attention = model(video, gaze_sequence, gaze_positions)
print(f"\n输入形状:")
print(f" 视频: {video.shape}")
print(f" 眼动: {gaze_sequence.shape}")
print(f" 注视点: {gaze_positions.shape}")
print(f"\n输出:")
print(f" 分心概率: {predictions.shape}, 值范围: [{predictions.min():.3f}, {predictions.max():.3f}]")
print(f" 注意力权重: {attention.shape}, 总和: {attention.sum(dim=-1).mean():.3f}")
for i in range(batch_size):
prob = torch.sigmoid(predictions[i]).item()
status = "分心" if prob > 0.5 else "专注"
print(f"样本{i+1}: 分心概率={prob:.2%}, 判断={status}")
|