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
| """
搜索驱动VLM测试框架
论文方法：遗传算法/爬山算法主动发现失效场景
"""
import numpy as np
import random
from typing import List, Tuple, Callable
class SearchBasedTester:
"""论文核心方法：搜索驱动测试
目标：最大化测试集直径（Test Set Diameter）
- 覆盖更多边缘场景
- 发现传统固定测试集遗漏的失效案例
搜索算法：
- Genetic Algorithm (GA)
- Hill Climbing (HC)
"""
def __init__(self, vlm_predictor: Callable, target_scenario: str):
"""
Args:
vlm_predictor: VLM预测函数
target_scenario: Euro NCAP场景编号（如D-02）
"""
self.predictor = vlm_predictor
self.scenario = target_scenario
self.param_ranges = {
"D-02": {
"phone_position": ["ear", "face", "chest", "lap"],
"hand_gesture": ["hold", "typing", "scrolling"],
"duration_sec": (1.0, 10.0),
"lighting": ["day", "night", "tunnel"],
"occlusion": ["none", "partial", "heavy"],
},
"D-05": {
"gaze_offset_deg": (0, 45),
"direction": ["left", "right", "up", "down"],
"duration_sec": (0.5, 10.0),
"head_pose": ["normal", "tilted", "turned"],
}
}
def generate_random_test_case(self) -> dict:
"""随机生成测试参数"""
params = {}
scenario_params = self.param_ranges[self.scenario]
for key, value in scenario_params.items():
if isinstance(value, list):
params[key] = random.choice(value)
elif isinstance(value, tuple):
params[key] = random.uniform(value[0], value[1])
return params
def evaluate_test_case(self, test_params: dict) -> Tuple[bool, float]:
"""
评估单个测试案例
Args:
test_params: 测试参数
Returns:
is_failure: 是否为失效案例（VLM预测错误）
failure_score: 失效严重程度（0-1）
"""
pred_result = self.predictor(test_params)
if self.scenario == "D-02":
ground_truth = test_params["duration_sec"] >= 3.0
is_detected = pred_result["phone_detected"] and pred_result["duration"] >= 3.0
is_failure = not is_detected if ground_truth else False
if is_failure:
failure_score = test_params["duration_sec"] / 10.0
else:
failure_score = 0.0
elif self.scenario == "D-05":
ground_truth = test_params["duration_sec"] >= 3.0
is_detected = pred_result["gaze_deviation"] and pred_result["duration"] >= 3.0
is_failure = not is_detected if ground_truth else False
failure_score = test_params["gaze_offset_deg"] / 45.0 if is_failure else 0.0
return is_failure, failure_score
def genetic_algorithm_search(self, population_size=50, generations=20) -> List[dict]:
"""
遗传算法搜索失效场景
论文方法：最大化失效发现率
Args:
population_size: 种群大小
generations: 进化代数
Returns:
failure_cases: 发现的失效案例列表
"""
population = [self.generate_random_test_case() for _ in range(population_size)]
failure_cases = []
for gen in range(generations):
fitness_scores = []
for test_case in population:
is_failure, score = self.evaluate_test_case(test_case)
fitness_scores.append(score)
if is_failure:
failure_cases.append(test_case)
total_fitness = sum(fitness_scores)
if total_fitness == 0:
selected = random.sample(population, population_size // 2)
else:
probs = [s / total_fitness for s in fitness_scores]
selected_indices = np.random.choice(
len(population),
size=population_size // 2,
p=probs
)
selected = [population[i] for i in selected_indices]
new_population = selected.copy()
while len(new_population) < population_size:
parent1, parent2 = random.sample(selected, 2)
child = self._crossover(parent1, parent2)
new_population.append(child)
for i in range(len(new_population)):
if random.random() < 0.1:
new_population[i] = self._mutate(new_population[i])
population = new_population
return failure_cases
def _crossover(self, parent1: dict, parent2: dict) -> dict:
"""交叉操作：混合两个父代参数"""
child = {}
for key in parent1.keys():
if random.random() < 0.5:
child[key] = parent1[key]
else:
child[key] = parent2[key]
return child
def _mutate(self, test_case: dict) -> dict:
"""变异操作：随机修改一个参数"""
mutated = test_case.copy()
key = random.choice(list(mutated.keys()))
scenario_params = self.param_ranges[self.scenario]
value_range = scenario_params[key]
if isinstance(value_range, list):
mutated[key] = random.choice(value_range)
elif isinstance(value_range, tuple):
mutated[key] = random.uniform(value_range[0], value_range[1])
return mutated
def calculate_test_set_diameter(self, test_cases: List[dict]) -> float:
"""
计算测试集直径（论文核心指标）
论文引用Feldt et al. 2016 ICST：
Test Set Diameter = 最大最小距离
- 量化测试多样性
- 覆盖边缘场景能力
Args:
test_cases: 测试案例集合
Returns:
diameter: 测试集直径
"""
vectors = []
for tc in test_cases:
vec = []
for key, value in tc.items():
if isinstance(value, str):
options = self.param_ranges[self.scenario][key]
vec.append(options.index(value) if value in options else 0)
else:
range_val = self.param_ranges[self.scenario][key]
if isinstance(range_val, tuple):
normalized = (value - range_val[0]) / (range_val[1] - range_val[0])
vec.append(normalized)
vectors.append(np.array(vec))
min_distances = []
for i, v1 in enumerate(vectors):
distances = []
for j, v2 in enumerate(vectors):
if i != j:
dist = np.linalg.norm(v1 - v2)
distances.append(dist)
if distances:
min_distances.append(min(distances))
diameter = max(min_distances) if min_distances else 0
return diameter
def mock_vlm_predictor(test_params: dict) -> dict:
"""模拟VLM预测结果"""
if test_params.get("lighting") == "night" and test_params.get("occlusion") == "partial":
return {
"phone_detected": False,
"gaze_deviation": False,
"duration": 0.0
}
return {
"phone_detected": True,
"gaze_deviation": True,
"duration": test_params.get("duration_sec", 3.0)
}
if __name__ == "__main__":
tester = SearchBasedTester(mock_vlm_predictor, "D-02")
print("搜索驱动VLM测试演示：")
print("-" * 70)
failures = tester.genetic_algorithm_search(population_size=30, generations=10)
print(f"发现失效案例数: {len(failures)}")
if failures:
print("\n失效案例分析：")
for i, case in enumerate(failures[:5], 1):
print(f"#{i}: 光照={case.get('lighting')} 遮挡={case.get('occlusion')} "
f"时长={case.get('duration_sec', 0):.1f}s")
diameter = tester.calculate_test_set_diameter(failures)
print(f"\n测试集直径: {diameter:.3f}")
print("直径越大 → 覆盖边缘场景越多")
|