Table 1.

Comparison of previous studies and the proposed method on sleep posture classification.

Sensor	Method	Dataset Used	Advantages	Limitations
Pressure-sensing mat	3D human pose estimation based on deep learning method [14]	Simulation dataset	A pressure-sensing mat is robust to covering.	A pressure-sensing mat has high cost and complex maintenance for home use.
Thermal camera	Human pose estimation based on deep learning method [15]	Simulation dataset	A thermal camera is robust to illuminationchanges and covering.	A thermal camera has high cost for home use.
Depth camera	Sleep posture classification based on deep learning method [16,18]	Simulation dataset	A depth camera is robust to low light intensity.	-Depth measurements of depth cameras suffer from various noise factors. -The depth camera is not prevalent at home surveillance.
Infrared camera	Sleep vs. wake states detection in young children based on motion analysis [22]	Real sleep data	-A single infrared camera is convenient and low-cost. -A general-purpose head detector is used.	This method only succeeds for 50% of nights.
	Human pose estimation based on deep learning method (OpenPose) [24,27]	Simulation dataset	The method can extract features of the skeleton effectively.	Their data [24] is obtained from mannequins in a simulated hospital room, and this method cannot perform well on real data [25].
	Sleep posture classificationbased on deep learning method [25,29]	Simulation dataset	The deep learning method can achieve good accuracy.	It focuses on classifying posture without detecting the upper-body and head region.
	Sleep posture detection and classification based on deep learning method (proposed method)	Simulation and real sleep dataset	A unified framework for simultaneously detecting and classifying upper-body pose and head pose is proposed.	Training personal data to learn CNN is required.