人體姿態(tài)評(píng)估中有兩種常見的顯示方式，分別是火柴人效果與BodyPix效果。其中火柴人效果本質(zhì)就是基于關(guān)鍵點(diǎn)的深度學(xué)習(xí)模型推理以后的顯示效果；Bodypix本質(zhì)就就是語義分割模型，基于Deeplabv3+ 與ResNet34構(gòu)建的一個(gè)人體語義分割模型。

火柴人

主要是基于關(guān)鍵點(diǎn)的人體姿態(tài)評(píng)估顯示方式，基于YOLOv8等人體姿態(tài)評(píng)估的關(guān)鍵點(diǎn)模型均可以實(shí)現(xiàn)，演示代碼運(yùn)行效果如下：

BodyPix

主要是基于語義分割的人體姿態(tài)評(píng)估顯示方式，支持對(duì)人體分割為10或者15個(gè)標(biāo)簽，演示代碼運(yùn)行效果如下：

實(shí)現(xiàn)代碼

我把相關(guān)的代碼使用PyQT5封裝為一個(gè)線程，相關(guān)的代碼如下：

classInferenceThread(QtCore.QThread):
fire_stats_signal=QtCore.pyqtSignal(dict)

def__init__(self,settings):
super(InferenceThread,self).__init__()
self.settings=settings
self.detector=None
ifself.settings.model_type==0:
self.detector=YOLOv8ORTPose(settings)
self.detector.show_skeleton=True
ifself.settings.model_type==1:
self.detector=YOLOv8ORTPose(settings)
self.detector.show_skeleton=True
ifself.settings.model_type==2:
weight_file_path="D:/projects/body_part_best.onnx"
self.detector=BodyPixSegment(weight_file_path)
self.input_image=settings.input_image

defrun(self):
ifself.detectorisNone:
return
ifself.input_image.endswith(".mp4"):
cap=cv.VideoCapture(self.input_image)
whileTrue:
ret,frame=cap.read()
ifretisTrue:
dst=self.detector.infer_image(frame)
self.fire_stats_signal.emit({"result":frame,"mask":dst})
else:
break
else:
frame=cv.imread(self.input_image)
dst=self.detector.infer_image(frame)
self.fire_stats_signal.emit({"result":frame,"mask":dst})
self.fire_stats_signal.emit({"done":"done"})
return

defon_yolov8_infer(self):
image_file=self.image_file_edit.text()
label_file=self.label_file_path.text()
model_file=self.weight_file_path.text()
iflen(image_file)==0orlen(label_file)==0orlen(model_file)==0:
QtWidgets.QMessageBox.warning(self,"警告","參數(shù)文件未選擇...")
return
settings=DLInferSettings()
settings.weight_file_path=self.weight_file_path.text()
settings.label_map_file_path=self.label_file_path.text()
settings.score_threshold=self.conf_spinbox.value()
settings.input_image=image_file
settings.model_type=0
ifself.rbtn0.isChecked():
settings.model_type=0
ifself.rbtn1.isChecked():
settings.model_type=1
ifself.rbtn2.isChecked():
settings.model_type=2
settings.target_deploy=1
self.work_thread=InferenceThread(settings)
self.work_thread.fire_stats_signal.connect(self.on_update_result_image)
self.work_thread.finished.connect(self.work_thread.deleteLater)
self.work_thread.start()
self.startBtn.setStyleSheet("background-color:gray;color:white")
self.startBtn.setEnabled(False)
self.stopBtn.setStyleSheet("background-color:cyan;color:black")
self.stopBtn.setEnabled(True)

運(yùn)行結(jié)果如下：