feat(worker): 接口1 四庭七眼测量真实实现(替换 Mock)

worker 侧从 Mock 替换为真实算法:
- face_analysis 包:detector(MediaPipe 478点) / pose(solvePnP 姿态) /
  calibration(虹膜直径法) / hair_segmenter+bisenet_model(方案B 头发分割) /
  measure(方案A兜底+B/A决策+七眼+换算) / annotation(numpy渐变线+中文标注)
- app.py:/api/v1/face/measure 接真实实现,返回 annotated_image_base64
  (不落盘不拼URL,落盘由网关做);加 X-Internal-Token 鉴权、/health 就绪态、
  可配置分辨率门槛、异常兜底
- 部署:start.sh/run_worker.sh/hair-worker.service 监听 8187;worker_config 示例
- 测试 tests/:Tier1合成真值<1e-6 + Tier2缩放不变 + Tier3叠加 + 错误码集成 +
  数值回归,pytest 24 项全绿
- 文档补实测基线表 + RTX5090/torch 说明

注:worker 为 RTX 5090(sm_120),pinned torch 2.2.2(cu121) 只到 sm_90,
BiSeNet 已自动回退 CPU(方案B 正常);要用 GPU 需换 torch cu128(≥2.7)。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
xsl
2026-06-14 16:07:28 +08:00
co-authored by Claude Opus 4.8
parent 3b706ec0ef
commit 8d3b145111
25 changed files with 1806 additions and 35 deletions
+101
View File
@@ -0,0 +1,101 @@
"""头部姿态估计(cv2.solvePnP)+ 正面照校验。
用通用 3D 头模与 6 个 MediaPipe 关键点求解欧拉角(yaw/pitch/roll,单位:度),
阈值即可写成业务可读的「yaw>15° 拒绝」,并把角度返回前端做拍照引导。
详见技术方案 §9。
"""
import os
import cv2
import numpy as np
from face_analysis.face_mesh_landmarks import PNP_INDICES
# 正面照判定阈值(度),可由环境变量覆盖,便于上线后按真实数据标定(见技术方案 §11)。
# ⚠️ 标定说明:基于通用 6 点 3D 头模 + solvePnP,对明显正面但相机略带俯仰/个体
# 脸型差异的真实照片,解出的 yaw/pitch 常落在 15~25°(roll 较稳定,多在 5° 内)。
# 因此默认阈值放宽到 30°,只拦截明显侧脸(真实侧脸 yaw 通常 40°+),
# 避免误杀正常上传图。生产可通过环境变量随时收紧/放宽,无需改代码。
YAW_THRESHOLD = float(os.getenv("FRONTAL_YAW_THR", "30"))
PITCH_THRESHOLD = float(os.getenv("FRONTAL_PITCH_THR", "30"))
ROLL_THRESHOLD = float(os.getenv("FRONTAL_ROLL_THR", "30"))
# 通用 3D 头部模型(单位 mm,近似),与 PNP_INDICES 一一对应:
# 鼻尖(1) / 下巴(152) / 左眼外角(33) / 右眼外角(263) / 左嘴角(61) / 右嘴角(291)
# ⚠️ 采用「相机坐标系」约定:x 向右、y 向下、z 向场景内(远离观察者)。
# 与 MediaPipe 像素坐标(y 下)一致,且 +z 指向人脸背面,
# 这样正面照解出的旋转矩阵≈单位阵,欧拉角≈0。
# 若只翻 y 不翻 z(或都不翻),会残留 ~180° 翻转使正面图被误判。
_MODEL_POINTS = np.array([
(0.0, 0.0, 0.0), # 鼻尖
(0.0, 63.6, 12.5), # 下巴(在鼻尖下方 → y 正)
(-43.3, -32.7, 26.0), # 左眼外角(在鼻尖上方 → y 负,且凹于鼻尖 → z 正)
(43.3, -32.7, 26.0), # 右眼外角
(-28.9, 28.9, 24.1), # 左嘴角
(28.9, 28.9, 24.1), # 右嘴角
], dtype=np.float64)
def estimate_head_pose(landmarks, image_width, image_height):
"""求解头部欧拉角,返回 (yaw, pitch, roll)(度)。solvePnP 失败返回 None。"""
lm = landmarks.landmark if hasattr(landmarks, "landmark") else landmarks
image_points = np.array([
(lm[i].x * image_width, lm[i].y * image_height)
for i in PNP_INDICES
], dtype=np.float64)
focal = float(image_width) # 近似焦距
cam_matrix = np.array([[focal, 0, image_width / 2],
[0, focal, image_height / 2],
[0, 0, 1]], dtype=np.float64)
dist = np.zeros((4, 1)) # 假设无畸变
success, rvec, _tvec = cv2.solvePnP(
_MODEL_POINTS, image_points, cam_matrix, dist,
flags=cv2.SOLVEPNP_ITERATIVE,
)
if not success:
return None
rot, _ = cv2.Rodrigues(rvec)
# 在「相机坐标系」(x右 y下 z内) 下抽取 Tait-Bryan 欧拉角,物理含义对齐:
# yaw = 绕 Y(竖轴)转 → 左右扭头
# pitch = 绕 X(横轴)转 → 上下点头
# roll = 绕 Z(光轴)转 → 面内倾斜
sy = (rot[0, 0] ** 2 + rot[1, 0] ** 2) ** 0.5
yaw = float(np.degrees(np.arctan2(-rot[2, 0], sy)))
pitch = float(np.degrees(np.arctan2(rot[2, 1], rot[2, 2])))
roll = float(np.degrees(np.arctan2(rot[1, 0], rot[0, 0])))
return yaw, pitch, roll
def check_frontal_face(landmarks, image_width, image_height,
yaw_thr=YAW_THRESHOLD, pitch_thr=PITCH_THRESHOLD,
roll_thr=ROLL_THRESHOLD):
"""正面照判定:yaw/pitch/roll 绝对值均在阈值内才算正面。
solvePnP 解算失败时返回 True(不拦截,交由后续逻辑),避免误杀。
"""
pose = estimate_head_pose(landmarks, image_width, image_height)
if pose is None:
return True
yaw, pitch, roll = pose
return abs(yaw) <= yaw_thr and abs(pitch) <= pitch_thr and abs(roll) <= roll_thr
if __name__ == "__main__":
import sys
from face_analysis.detector import detector
path = sys.argv[1] if len(sys.argv) > 1 else "tests/fixtures/frontal.jpg"
img = cv2.imread(path)
if img is None:
print(f"无法读取图片: {path}")
sys.exit(1)
h, w = img.shape[:2]
lms = detector.detect(img)
if lms is None:
print("未检出人脸")
sys.exit(1)
yaw, pitch, roll = estimate_head_pose(lms, w, h)
frontal = check_frontal_face(lms, w, h)
print(f"yaw={yaw:.2f} pitch={pitch:.2f} roll={roll:.2f} frontal={frontal}")