我目前正在尝试使用 Python 的多处理库将 MediaPipe 姿态估计器实现为一个独立的基于事件的进程,但它挂在 MediaPipe 的
Pose.process()
函数上。
我用另一个进程(readFrames)输入帧。每当捕获一帧时,它就会被写入共享对象并告诉 MediaPipe 进程 (MediaPipeRunner) 开始处理当前图像:
def readFrames(ns, event):
#initialize the video capture object
cap = cv2.VideoCapture(0)
while cap.isOpened():
ret, frame = cap.read()
if ret:
ns.frame = frame
event.set()
cv2.imshow('Orijinal Frame', frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
cap.release()
cv2.destroyAllWindows()
return -1
else:
return
class MediaPipeRunner(mproc.Process):
def __init__(self, name, nsFrame, nsMediaPipe, eventWait, eventPublish):
super(MediaPipeRunner, self).__init__()
# Specify a name for the instance
self.name = name
# Input and output namespaces
self.nsFrame = nsFrame
self.nsMediaPipe = nsMediaPipe
# Waiter and publisher events
self.eventWait = eventWait
self.eventPublish = eventPublish
# Create a pose estimator from MediaPipe
mp_pose = mp.solutions.pose
# Specify pose estimator parameters (static)
static_image_mode = True
model_complexity = 1
enable_segmentation = True # DONT CHANGE
min_detection_confidence = 0.5
# Create a pose estimator here
self.pose = mp_pose.Pose(
static_image_mode=static_image_mode,
model_complexity=model_complexity,
enable_segmentation=enable_segmentation,
min_detection_confidence=min_detection_confidence,
smooth_landmarks=False,
)
def run(self):
while True:
eventFrame.wait()
# This part is where it gets stuck:
results = self.pose.process(cv2.cvtColor(self.nsFrame.frame, cv2.COLOR_BGR2RGB))
if not results.pose_landmarks:
continue
self.nsMediaPipe.segmentation = results.segmentation_mask
eventMP.set()
这就是我绑定进程、命名空间和事件的方式:
if __name__=="__main__":
mgr = mproc.Manager()
nsFrame = mgr.Namespace()
nsMP = mgr.Namespace()
eventFrame = mproc.Event()
eventMP = mproc.Event()
camCap = mproc.Process(name='camCap', target=readFrames, args=(nsFrame, eventFrame, ))
camCap.daemon=True
mpCap = MediaPipeRunner('mpCap', nsFrame, nsMP, eventFrame, eventMP, )
mpCap.daemon=True
camCap.start()
mpCap.start()
camCap.join()
mpCap.join()
我是否在进程上采取了错误的步骤,或者 MediaPipe 与 Python 的多处理库不兼容?
任何帮助将不胜感激,提前致谢:)
P.S.:我通过 pip 安装了 MediaPipe,版本为 0.8.9.1。
我发现了问题:当在Python中使用
process
结构时,with
函数表现正确(不知道为什么):
with mp_pose.Pose(
static_image_mode=static_image_mode,
model_complexity=model_complexity,
enable_segmentation=enable_segmentation,
min_detection_confidence=min_detection_confidence,
smooth_landmarks=False,
) as pose:
现在这部分可以工作了!
results = self.pose.process(cv2.cvtColor(self.nsFrame.frame, cv2.COLOR_BGR2RGB))
希望对您有帮助。
我也遇到了同样的问题,通过更改解决了
results = self.pose.process(cv2.cvtColor(self.nsFrame.frame, cv2.COLOR_BGR2RGB))
到
pose= self.pose
results = pose.process(cv2.cvtColor(self.nsFrame.frame, cv2.COLOR_BGR2RGB))
不太确定为什么会这样,如果有人理解这一点,我很想得到一些意见
我认为 MediaPipe 模型只能在创建它们的线程中访问,因此一种可能的解决方法是为每个线程创建一个模型。开销并不大。
model_storage = threading.local()
# ...
if not hasattr(model_storage, 'recogniser'):
model_storage.recogniser = mp.tasks.vision.GestureRecognizer.create_from_options(
mp.tasks.vision.GestureRecognizerOptions(
base_options=mp.tasks.BaseOptions(
model_asset_path=os.path.join(os.path.dirname(__file__),
'models',
'HandGesture')
)
)
)
recogniser = model_storage.recogniser