Non-intrusive apnoea/hypopnoea detection system via graph-signal analysis of Microsoft Kinect captured depth video Cheng. Yang1, Yu. Mao2, Gene. Cheung2, Vladimir. Stankovic1, Kevin. Chan3,4 1Department 2National 3School
of Electronic and Electrical Engineering, University of Strathclyde, Glasgow, United Kingdom,
{cheng.yang, vladimir.stankovic}@strath.ac.uk
Institute of Informatics, Tokyo, Japan,
{mao, cheung}@nii.ac.jp
of Medicine, University of Western Sydney, 4Camden and Campbelltown Hospitals, Sydney, NSW, Australia
[email protected]
Introduction: Diagnostic Sleep Studies
Methodologies
Non-intrusive, easily deployable system for detection of apnoea /
1. Capture depth video of patient from above-head location.
hypopnoea episodes important for diagnostic sleep studies. State of the art sleep monitoring systems are either i) cheap & nonintrusive but inaccurate, or ii) accurate but expensive & intrusive. Vibration-sensing wrist-bands (Fitbit, Jawbone UP) 1. Minimally intrusive. 2. Mostly record sleep time, not the quality of sleep.
Full multi-sensing monitoring device (Philips Alice PDx) 1. Accurate in measuring vital signs: oxygen intake, airflow, etc. 2. Expensive and intrusive with multiple body straps and tubes.
2. For each depth image, map depth measurements into two best-fitting ellipses (I: chest, II: abdomen). 3. Sleep event detection based on the variance of the major and minor radius of the best fitting ellipses (video features) in a time window of samples.
patient pillow Depth video
Side view of a sleeping patient
abdomen http://cdn0.vox-cdn.com/entry_photo_images/7275017/DSC_3047-hero_verge_medium_landscape.jpg
http://static1.businessinsider.com/image/51a8b6246bb3f7df3c000009-1200/the-fitbit-has-a-better-sleep-tracker.jpg
http://www.newscenter.philips.com/pwc_nc/main/standard/resources/corporate/press/2010/siesta/Siesta_3_small.jpg
Objective Design an inexpensive, non-intrusive apnoea / hypopnoea detection system:
chest Dual-ellipse
Results
Ellipse model 20-second samples of normal breathing and hypopnoea from a patient in upright sleeping
1. Track chest/abdomen respiratory movements using depth video*
position, covered with quilt.
captured by a Microsoft Kinect Camera.
Major and minor radius of chest ellipse (a1,
2. Detect abnormal sleep events based on derived features from video.
b1) and those of abdomen ellipse (a2, b2).
*Depth image: per-pixel distance bet’n scene objects and camera—partial geometry info in scene.
Classify events into: i) normal breathing, ii)
Background
abnormal breathing (apnoea / hypopnea).
M.-C. Yu et al. “Breath and position monitoring during sleeping with a depth camera,” in International Conference on Health Informatics, Vilamoura, Portugal, February 2012. 1. A sleep monitoring system with torso movement detection using Kinect. 2. Cannot distinguish chest and abdomen movements individually.
Classification performance using two schemes based on graph-signal analysis (graph smoothness (GS) and robust graph smoothness (RGS)) with two conventional implementations of support vector machine (SVM) (linear kernel (SVM-l) and radial basis function kernel (SVM-rbf)). (50 samples for testing)
3. Not robust: by mounting a camera on the ceiling and measuring the distance to the body, this won’t work if the patient sleeps sideway. In our case, by using a dual-ellipse model, it is still possible for us to track the breathing cycle even if the patient’s is in sideway. (from M.-C. Yu et al. (2012))
System Overview Venue: Bondi Junction Private Sleep Laboratory in Sydney, Australia.
50 training samples 30 training samples Graph-based schemes achieved perfect 2-event classification for 50 and 30 training samples, and are more robust under noise in the training set than two SVM implementations; RGS is more noise-robust than GS.
Subjects: patients admitted for diagnostic sleep studies.
Conclusion
Diagnostic sleep equipment: Alice 6 LDx / Sleepware G3
A non-intrusive apnoea/hypopnoea detection system using Kinect is proposed.
Our equipment: Microsoft Kinect Camera and a Lenovo X220 laptop.
Accurate 2-event classification (normal vs. abnormal) based on the depth
Un-obstructed camera view of the patient’s upper body (upright/sideway). Video resolution 640x480 @ 30 fps. Can operate in complete darkness.
video features (variances of radius of ellipse fitting chest / abdominal).
References C. Yang et al, “Graph-based Depth Video Denoising and Event Detection for Sleep Monitoring,” accepted to IEEE International Workshop on Multimedia Signal Processing, Jakarta, Indonesia, September, 2014.