MASTER THESIS 5T746 Two-Person Segmentation for Video Based Actigraph Generation in a Shared Bed Environment

Sleep studies assess the quality of sleep and body response to sleeping disorders. Sleep studies have gained prominence as empirical studies have shown the relation between untreated sleep disorders to e.g. high blood pressure, stroke, heart diseases and other severe medical conditions. Polysomnography (PSG) [1] and wrist actigraphy are widely used in sleep studies. These methods are neither comfortable nor a convenient way to monitor sleep because the user is confined by sensors and wires. There is a very promising market for contactless sleep monitoring, not only for sleep clinics, but also for home applications. With the development of video processing and analysis, video based sleep analysis promises to be a suitable alternative to wrist actigraphy for home monitoring. Compared with the commonly used wrist actigraphy, the video based method is contactless and provides a possibility of measuring the movements of the entire body. A shared bed scenario is a common case in home environment and distinguishing person of interest's (POI) movements from that of bed partner's (BP) movements is difficult as they sleep in close proximity. It becomes even more challenging when the camera is placed on the bed side table as the viewing angle would show BP to be just behind POI. This thesis presents the development and implementation of a novel sleep monitoring application that is capable of monitoring the movements from the person of interest (POI) in a shared bed scenario using a camera and an IR emitter mounted on the bed side table. The application takes video data as input and generates a video actigraph that has movement signals that relate to POI movements. The POI area in the image is segmented by using an AdaBoost classifier that uses several discriminative features like brightness, focus, gradients and motion vectors. The video actigraph is then generated by frame differencing in the POI area. The application has been ported onto an embedded platform called BeagleBoard and has the potential to run in real time. The application has been tested over diverse data sets and has been found to be robust to movements originating from under the blanket, various sleeping positions, different illumination conditions, beds and blankets. The developed sleep monitoring application can successfully replace wrist actigraphy in monitoring POI's movements. The application can detect movements from POI up to an accuracy of 92.5% compared to 70% accuracy using video actigraph generated without POI area segmentation and 75% …