【前言】
人脸疲劳检测：一种通过分析人脸特征来判断一个人是否处于疲劳状态的技术。其原理主要基于计算机视觉和机器学习方法。当人疲劳时，面部会出现一些特征变化，如眼睛闭合程度增加、眨眼频率变慢、打哈欠、头部姿态改变等。
例如，通过检测眼睛的状态来判断疲劳程度是一个关键部分。正常情况下，人的眨眼频率相对稳定，而当疲劳时，眨眼频率会降低，并且每次眨眼时眼睛闭合的时间可能会延长。同时，头部可能会不自觉地下垂或者摇晃，这些特征都可以作为疲劳检测的依据。
米尔MYC-LR3576采用8核CPU+搭载6 TOPS的NPU加速器，3D GPU，能够非常轻松的实现这个功能，下面就如何实现这一功能分享如下：
【硬件】
1、米尔MYC-LR3576开发板
2、USB摄像头
【软件】
v4l2
openCV
dlib库：dlib 是一个现代化的 C++ 工具包，它包含了许多用于机器学习、图像处理、数值计算等多种任务的算法和工具。它的设计目标是提供高性能、易于使用的库，并且在开源社区中被广泛应用。
【实现步骤】
1、安装python-opencv
2、按装dlib库
3、安装v4l2库
【代码实现】
1、引入cv2、dlib以及线程等：
import cv2
import dlib
import numpy as np
import time
from concurrent.futures import ThreadPoolExecutor
import threading
复制代码detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
复制代码def eye_aspect_ratio(eye):
A = np.linalg.norm(np.array(eye[1]) - np.array(eye[5]))
B = np.linalg.norm(np.array(eye[2]) - np.array(eye[4]))
C = np.linalg.norm(np.array(eye[0]) - np.array(eye[3]))
ear = (A + B) / (2.0 * C)
return ear
复制代码def get_head_pose(shape):
# 定义面部特征点的三维坐标
object_points = np.array([
(0.0, 0.0, 0.0), # 鼻尖
(0.0, -330.0, -65.0), # 下巴
(-225.0, 170.0, -135.0), # 左眼左眼角
(225.0, 170.0, -135.0), # 右眼右眼角
(-150.0, -150.0, -125.0), # 左嘴角
(150.0, -150.0, -125.0) # 右嘴角
], dtype=np.float32)
image_pts = np.float32([shape[i] for i in [30, 8, 36, 45, 48, 54]])
size = frame.shape
focal_length = size[1]
center = (size[1] // 2, size[0] // 2)
camera_matrix = np.array(
[[focal_length, 0, center[0]],
[0, focal_length, center[1]],
[0, 0, 1]], dtype="double"
)
dist_coeffs = np.zeros((4, 1))
(success, rotation_vector, translation_vector) = cv2.solvePnP(
object_points, image_pts, camera_matrix, dist_coeffs, flags=cv2.SOLVEPNP_ITERATIVE
)
rmat, _ = cv2.Rodrigues(rotation_vector)
angles, _, _, _, _, _ = cv2.RQDecomp3x3(rmat)
return angles
复制代码EYE_AR_THRESH = 0.3
EYE_AR_CONSEC_FRAMES = 48
复制代码我们先使用v4l2-ctl --list-devices来例出接在开发板上的列表信息：
USB Camera: USB Camera (usb-xhci-hcd.0.auto-1.2):
/dev/video60
/dev/video61
/dev/media7
复制代码在代码中填入60为摄像头的编号：
cap = cv2.VideoCapture(60)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 480) # 降低分辨率
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 320)
复制代码# 多线程处理函数
def process_frame(frame):
global COUNTER, TOTAL
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
faces = detector(gray, 0) # 第二个参数为0，表示不使用upsampling
for face in faces:
landmarks = predictor(gray, face)
shape = [(landmarks.part(i).x, landmarks.part(i).y) for i in range(68)]
left_eye = shape[36:42]
right_eye = shape[42:48]
left_ear = eye_aspect_ratio(left_eye)
right_ear = eye_aspect_ratio(right_eye)
ear = (left_ear + right_ear) / 2.0
if ear < EYE_AR_THRESH:
with lock:
COUNTER += 1
else:
with lock:
if COUNTER >= EYE_AR_CONSEC_FRAMES:
TOTAL += 1
COUNTER = 0
# 绘制68个特征点
for n in range(0, 68):
x, y = shape[n]
cv2.circle(frame, (x, y), 2, (0, 255, 0), -1)
cv2.putText(frame, f"Eye AR: {ear:.2f}", (10, 30), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
cv2.putText(frame, f"Blink Count: {TOTAL}", (10, 60), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
# 计算头部姿势
angles = get_head_pose(shape)
pitch, yaw, roll = angles
cv2.putText(frame, f"Pitch: {pitch:.2f}", (10, 120), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
cv2.putText(frame, f"Yaw: {yaw:.2f}", (10, 150), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
cv2.putText(frame, f"Roll: {roll:.2f}", (10, 180), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
# 判断疲劳状态
if COUNTER >= EYE_AR_CONSEC_FRAMES or abs(pitch) > 30 or abs(yaw) > 30 or abs(roll) > 30:
cv2.putText(frame, "Fatigue Detected!", (10, 210), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
return frame
复制代码with ThreadPoolExecutor(max_workers=2) as executor:
future_to_frame = {}
while True:
ret, frame = cap.read()
if not ret:
break
# 提交当前帧到线程池
future = executor.submit(process_frame, frame.copy())
future_to_frame[future] = frame
# 获取已完成的任务结果
for future in list(future_to_frame.keys()):
if future.done():
processed_frame = future.result()
cv2.imshow("Frame", processed_frame)
del future_to_frame[future]
break
# 计算帧数
fps_counter += 1
elapsed_time = time.time() - start_time
if elapsed_time > 1.0:
fps = fps_counter / elapsed_time
fps_counter = 0
start_time = time.time()
cv2.putText(processed_frame, f"FPS: {fps:.2f}", (10, 90), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2)
if cv2.waitKey(1) & 0xFF == ord('q'):
复制代码
根据检测的结果，我们就可以来实现疲劳提醒等等的功能。
整体代码如下：
import cv2
import dlib
import numpy as np
import time
from concurrent.futures import ThreadPoolExecutor
import threading
# 初始化dlib的面部检测器和特征点预测器
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
# 修改字体大小
font_scale = 0.5 # 原来的字体大小是0.7，现在改为0.5
# 定义计算眼睛纵横比的函数
def eye_aspect_ratio(eye):
A = np.linalg.norm(np.array(eye[1]) - np.array(eye[5]))
B = np.linalg.norm(np.array(eye[2]) - np.array(eye[4]))
C = np.linalg.norm(np.array(eye[0]) - np.array(eye[3]))
ear = (A + B) / (2.0 * C)
return ear
# 定义计算头部姿势的函数
def get_head_pose(shape):
# 定义面部特征点的三维坐标
object_points = np.array([
(0.0, 0.0, 0.0), # 鼻尖
(0.0, -330.0, -65.0), # 下巴
(-225.0, 170.0, -135.0), # 左眼左眼角
(225.0, 170.0, -135.0), # 右眼右眼角
(-150.0, -150.0, -125.0), # 左嘴角
(150.0, -150.0, -125.0) # 右嘴角
], dtype=np.float32)
image_pts = np.float32([shape[i] for i in [30, 8, 36, 45, 48, 54]])
size = frame.shape
focal_length = size[1]
center = (size[1] // 2, size[0] // 2)
camera_matrix = np.array(
[[focal_length, 0, center[0]],
[0, focal_length, center[1]],
[0, 0, 1]], dtype="double"
)
dist_coeffs = np.zeros((4, 1))
(success, rotation_vector, translation_vector) = cv2.solvePnP(
object_points, image_pts, camera_matrix, dist_coeffs, flags=cv2.SOLVEPNP_ITERATIVE
)
rmat, _ = cv2.Rodrigues(rotation_vector)
angles, _, _, _, _, _ = cv2.RQDecomp3x3(rmat)
return angles
# 定义眼睛纵横比阈值和连续帧数阈值
EYE_AR_THRESH = 0.3
EYE_AR_CONSEC_FRAMES = 48
# 初始化计数器
COUNTER = 0
TOTAL = 0
# 创建锁对象
lock = threading.Lock()
# 打开摄像头
cap = cv2.VideoCapture(60)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 480) # 降低分辨率
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 320)
# 初始化帧计数器和时间戳
fps_counter = 0
start_time = time.time()
# 多线程处理函数
def process_frame(frame):
global COUNTER, TOTAL
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
faces = detector(gray, 0) # 第二个参数为0，表示不使用upsampling
for face in faces:
landmarks = predictor(gray, face)
shape = [(landmarks.part(i).x, landmarks.part(i).y) for i in range(68)]
left_eye = shape[36:42]
right_eye = shape[42:48]
left_ear = eye_aspect_ratio(left_eye)
right_ear = eye_aspect_ratio(right_eye)
ear = (left_ear + right_ear) / 2.0
if ear < EYE_AR_THRESH:
with lock:
COUNTER += 1
else:
with lock:
if COUNTER >= EYE_AR_CONSEC_FRAMES:
TOTAL += 1
COUNTER = 0
# 绘制68个特征点
for n in range(0, 68):
x, y = shape[n]
cv2.circle(frame, (x, y), 2, (0, 255, 0), -1)
cv2.putText(frame, f"Eye AR: {ear:.2f}", (10, 30), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
cv2.putText(frame, f"Blink Count: {TOTAL}", (10, 60), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
# 计算头部姿势
angles = get_head_pose(shape)
pitch, yaw, roll = angles
cv2.putText(frame, f"Pitch: {pitch:.2f}", (10, 120), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
cv2.putText(frame, f"Yaw: {yaw:.2f}", (10, 150), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
cv2.putText(frame, f"Roll: {roll:.2f}", (10, 180), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
# 判断疲劳状态
if COUNTER >= EYE_AR_CONSEC_FRAMES or abs(pitch) > 30 or abs(yaw) > 30 or abs(roll) > 30:
cv2.putText(frame, "Fatigue Detected!", (10, 210), cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), 2)
return frame
with ThreadPoolExecutor(max_workers=2) as executor:
future_to_frame = {}
while True:
ret, frame = cap.read()
if not ret:
break
# 提交当前帧到线程池
future = executor.submit(process_frame, frame.copy())
future_to_frame[future] = frame
# 获取已完成的任务结果
for future in list(future_to_frame.keys()):
if future.done():
processed_frame = future.result()
cv2.imshow("Frame", processed_frame)
del future_to_frame[future]
break
# 计算帧数
fps_counter += 1
elapsed_time = time.time() - start_time
if elapsed_time > 1.0:
fps = fps_counter / elapsed_time
fps_counter = 0
start_time = time.time()
cv2.putText(processed_frame, f"FPS: {fps:.2f}", (10, 90), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# 释放摄像头并关闭所有窗口
cap.release()
cv2.destroyAllWindows()
复制代码【米尔MYC-LR3576核心板及开发板】
这块开发板性能强大，能轻松实现对人脸的疲劳检测，通过计算结果后进入非常多的工业、人工智能等等的实用功能。
/3 
