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
| import torch
import torch.nn as nn
import torch.nn.functional as F
class FaceFeatureExtractor(nn.Module):
"""
驾驶员面部特征提取
双流特征:
1. 全脸特征: 整体面部姿态/表情
2. 虹膜加权眼部特征: 精确注视方向
论文Section 3.1: Facial Feature Extraction
"""
def __init__(self, embed_dim: int = 256):
super().__init__()
self.face_cnn = 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, embed_dim)
)
self.eye_cnn = 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((2, 4)),
nn.Flatten(),
nn.Linear(128 * 8, embed_dim)
)
self.iris_gate = nn.Sequential(
nn.Linear(embed_dim * 2, embed_dim),
nn.Sigmoid()
)
self.fusion = nn.Linear(embed_dim * 2, embed_dim)
def forward(self, face_img: torch.Tensor,
eye_patch: torch.Tensor) -> torch.Tensor:
"""
Args:
face_img: (B, 3, 224, 224) 全脸图像
eye_patch: (B, 3, 112, 112) 眼部裁剪
Returns:
face_feat: (B, embed_dim) 面部特征向量
"""
face_feat = self.face_cnn(face_img)
eye_feat = self.eye_cnn(eye_patch)
gate = self.iris_gate(torch.cat([face_feat, eye_feat], dim=-1))
weighted_eye = eye_feat * gate
fused = self.fusion(torch.cat([face_feat, weighted_eye], dim=-1))
return fused
class TrafficObjectEncoder(nn.Module):
"""
交通场景物体编码器
使用YOLOv8检测物体 → 空间特征编码
论文Section 3.2: Traffic Object Spatial Features
"""
def __init__(self, embed_dim: int = 256, num_objects: int = 10):
super().__init__()
self.num_objects = num_objects
self.class_embed = nn.Embedding(20, embed_dim)
self.spatial_encoder = nn.Sequential(
nn.Linear(5, 64), nn.ReLU(),
nn.Linear(64, embed_dim)
)
self.appearance_cnn = 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.AdaptiveAvgPool2d((2, 2)),
nn.Flatten(),
nn.Linear(64 * 4, embed_dim)
)
self.fusion = nn.Linear(embed_dim * 3, embed_dim)
def forward(self, bboxes: torch.Tensor, classes: torch.Tensor,
crops: torch.Tensor) -> torch.Tensor:
"""
Args:
bboxes: (B, N, 5) [x, y, w, h, conf]
classes: (B, N) 物体类别ID
crops: (B, N, 3, 32, 32) 物体裁剪图像
Returns:
obj_embeddings: (B, N, embed_dim)
"""
B, N = bboxes.shape[:2]
cls_emb = self.class_embed(classes)
sp_emb = self.spatial_encoder(bboxes)
crops_flat = crops.view(B * N, 3, 32, 32)
app_emb = self.appearance_cnn(crops_flat).view(B, N, -1)
all_emb = torch.cat([cls_emb, sp_emb, app_emb], dim=-1)
return self.fusion(all_emb)
class TransGazeObject(nn.Module):
"""
TransGaze-Object: 基于Transformer的驾驶员注视物体预测
核心创新: 使用交叉注意力直接预测注视物体,
跳过传统的PoG (Point-of-Gaze)中间估计步骤
论文Section 3.3: Transformer Cross-Attention Module
Args:
embed_dim: 嵌入维度
num_heads: 注意力头数
num_objects: 最大物体数
"""
def __init__(self, embed_dim: int = 256, num_heads: int = 8,
num_objects: int = 10):
super().__init__()
self.face_extractor = FaceFeatureExtractor(embed_dim)
self.object_encoder = TrafficObjectEncoder(embed_dim, num_objects)
self.cross_attn = nn.MultiheadAttention(
embed_dim, num_heads, batch_first=True
)
self.norm1 = nn.LayerNorm(embed_dim)
self.ffn = nn.Sequential(
nn.Linear(embed_dim, embed_dim * 4),
nn.GELU(),
nn.Linear(embed_dim * 4, embed_dim)
)
self.norm2 = nn.LayerNorm(embed_dim)
self.classifier = nn.Linear(embed_dim, num_objects + 1)
def forward(self, face_img, eye_patch, bboxes, classes, crops):
"""
Args:
face_img: (B, 3, 224, 224)
eye_patch: (B, 3, 112, 112)
bboxes: (B, N, 5)
classes: (B, N)
crops: (B, N, 3, 32, 32)
Returns:
logits: (B, N+1) 注视物体概率（含背景）
attn_weights: (B, 1, N) 注意力权重
"""
face_feat = self.face_extractor(face_img, eye_patch)
query = face_feat.unsqueeze(1)
obj_emb = self.object_encoder(bboxes, classes, crops)
attn_out, attn_weights = self.cross_attn(query, obj_emb, obj_emb)
x = self.norm1(query + attn_out)
x = self.norm2(x + self.ffn(x))
logits = self.classifier(x.squeeze(1))
return logits, attn_weights
if __name__ == "__main__":
import numpy as np
import time
from sklearn.metrics import accuracy_score
NUM_OBJECTS = 8
BATCH_SIZE = 16
EMBED_DIM = 256
model = TransGazeObject(
embed_dim=EMBED_DIM,
num_heads=8,
num_objects=NUM_OBJECTS
)
total_params = sum(p.numel() for p in model.parameters())
print(f"TransGaze-Object参数量: {total_params:,}")
np.random.seed(42)
def generate_sample(num_obj=5):
"""生成一个驾驶场景样本"""
face = torch.rand(3, 224, 224)
eye = torch.rand(3, 112, 112)
bboxes = torch.rand(num_obj, 5) * torch.tensor([1, 1, 0.3, 0.3, 1])
classes = torch.randint(0, 10, (num_obj,))
crops = torch.rand(num_obj, 3, 32, 32)
return face, eye, bboxes, classes, crops
print("\n--- 生成测试数据 ---")
faces, eyes, bbox_list, class_list, crop_list = [], [], [], [], []
labels = []
for i in range(BATCH_SIZE):
n = np.random.randint(3, NUM_OBJECTS + 1)
face, eye, bbox, cls, crop = generate_sample(n)
faces.append(face)
eyes.append(eye)
bbox_list.append(bbox)
class_list.append(cls)
crop_list.append(crop)
labels.append(np.random.randint(0, n + 1))
max_obj = NUM_OBJECTS
bboxes_padded = torch.zeros(BATCH_SIZE, max_obj, 5)
classes_padded = torch.zeros(BATCH_SIZE, max_obj, dtype=torch.long)
crops_padded = torch.zeros(BATCH_SIZE, max_obj, 3, 32, 32)
masks = torch.zeros(BATCH_SIZE, max_obj, dtype=torch.bool)
for i, (bbox, cls, crop) in enumerate(zip(bbox_list, class_list, crop_list)):
n = len(bbox)
bboxes_padded[i, :n] = bbox
classes_padded[i, :n] = cls
crops_padded[i, :n] = crop
masks[i, :n] = True
face_batch = torch.stack(faces)
eye_batch = torch.stack(eyes)
labels = torch.tensor(labels)
print(f" Face: {face_batch.shape}")
print(f" Eye: {eye_batch.shape}")
print(f" BBox: {bboxes_padded.shape}")
print(f" Labels: {labels.shape}")
print("\n--- 前向传播 ---")
model.eval()
with torch.no_grad():
logits, attn = model(face_batch, eye_batch, bboxes_padded,
classes_padded, crops_padded)
preds = torch.argmax(logits, dim=1)
print(f" Logits: {logits.shape}")
print(f" Attention: {attn.shape}")
print(f" Predictions: {preds}")
print("\n--- 训练流程 ---")
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
model.train()
for epoch in range(5):
optimizer.zero_grad()
logits, _ = model(face_batch, eye_batch, bboxes_padded,
classes_padded, crops_padded)
loss = criterion(logits, labels)
loss.backward()
optimizer.step()
acc = accuracy_score(labels.numpy(), torch.argmax(logits, dim=1).detach().numpy())
print(f"Epoch {epoch+1}: Loss={loss.item():.4f}, Acc={acc:.2%}")
model.eval()
with torch.no_grad():
start = time.time()
for _ in range(50):
_ = model(face_batch[:1], eye_batch[:1], bboxes_padded[:1],
classes_padded[:1], crops_padded[:1])
elapsed = (time.time() - start) / 50
print(f"\n单样本推理延迟: {elapsed*1000:.1f}ms")
|
