Video Monitoring and Analysis of Human Behavior for Diagnosis of Obstructive Sleep Apnoea

This thesis investigates the use of the computerized video monitoring in support of the diagnosis of obstructive sleep apnoea, which is characterized by repetitive obstruction of the upper airways during sleep and resulted in arterial oxyhaemoglobin desaturation, excessive arousals, unrefreshing sleep, excessive daytime sleepiness, poor health-related quality of life, hypertension and severe life-threatening complications. According to recent research findings, the best predictors of morbidity are nocturnal oxygen saturation and movements during sleep. Although pulse oximetry is a well-established technique to analyze oxygen saturation, video monitoring and interpretation is less well developed due to the technical challenges of persistent occlusion, obscuration of the body by the bedding, variation of human behavior and the large volume of video data. This work introduces a new automatic video monitoring technique for breathing behavior anomaly detection and assisting in diagnosis of obstructive sleep apnoea. The algorithm utilizes infrared video information, imposes few positional constraints on the patient, and deals with fully or partially covered bodies. A new motion detection model is presented to capture subtle and cyclical breathing signals. A novel action template is introduced to capture the dynamic spatial-temporal shape of normal breathing activities for action recognition, and adapts as the subject’s pose changes. The onlineconstructed action template is used to classify an action as a normal breathing episode, an apnoea episode or a body movement episode. Although the presented approach is designed for diagnosis of obstructive sleep apnoea, it could be utilized in other applications that require the analysis of breathing behavior or monitoring subtle and cyclical activity. This work also introduces two novel monocular video approaches (MatchPose and RTPose) for pose recognition of the covered human body. They are recommended for different purposes: RTPose provides coarse pose estimation and is computationally efficient; MatchPose produces fine pose estimation but takes 0.4 seconds to process a 320× 240 frame. If full body pose estimation is desirable, we recommend MatchPose. On the other hand, in the interests of computational speed, we recommend incorporating RTPose with motion information. The methods assume subjects lying horizontally. In addition, a low variance error boosting algorithm is developed for training head and upper leg pose templates. In evaluation, we demonstrate that the breathing monitoring algorithm achieves high accuracy using confusion matrix in recognizing abnormal breathing activities and body movements and in classification of symptomatic and non-symptomatic subjects, and that the two pose estimation algorithms are able to identify human configurations with various poses and occlusion levels, and they are not particularly sensitive to environmental settings, including illumination and camera angle.