| Healthcare is an increasingly important application area. In this thesis, we explore issues and propose new techniques for home and remote healthcare applications. We first propose a solution to long-term home healthcare monitoring using video, audio, passive infrared (PIR), and heart-rate (HR) sensors, where long-term sleep monitoring is used as an example application. Taking the standard Actigraphy and questionnaire as ground-truth, we show that several factors in sleep quality measurement can be learned from these sensing data. A video or PIR sensor can provide performance comparable to Actigraphy while being low-cost and non-intrusive. The addition of audio and heart-rate (HR) sensors to the system (albeit a bit intrusive for adding HR sensors) makes the system more robust. With the proposed use of multimodal sensors and data fusion, our work showed the feasibility of low-cost long-term sleep monitoring at home. In the second part, we explore the use of audio surveillance for healthcare applications. Audio sensors have advantages over video sensors for monitoring in terms of privacy and computation. Investigation on the classification approaches based on generative models (e.g., Hidden Markov Model) and discriminative models (e.g., Support Vector Machine) for general audio event classification is conducted. We further propose a binary hierarchical classifier with feature selection for audio event detection which can keep the number of testing classifiers and feature dimension low while maintaining good performance. We show that audio events important to home surveillance can be detected with the proposed scheme. In the third part, we investigate solutions for reducing data transmission load and for classifications with an incomplete set of features for remote healthcare monitoring systems. We formulate a data scaling method to determine the order of data to be cut-down from the transmission list. We propose adaptations for classifiers in the server side including pre-training classifiers with subsets of features, adaptation of classifier testing models, and prediction of missing features to testing samples with partially available features. With experiments on ECG classifications, we show that the proposed data scaling procedure can be an effective technique for saving bandwidth (thus sensor transmission power) while maintaining satisfactory performance. |
