I. Introduction
Aging population shows a dramatic increasing during the past decade. The aging population explosion has become a social problem in both United States and many other countries. Independent lifestyles are highly desired by elderly people. However, independent lifestyles of older adults often come with high risks. To assist elderly persons' independent living, many smart home technologies have been developed to track and monitor activities of elderly persons at home with various sensors. Specifically, the video sensor is able to analyze human actions in visual scenes, which has significant advantages in providing rich contextual information of human activities. However, the application of such approach may usually be hindered by its dynamic usage environment in which the scenario varies over time [1]. The visual surveillance system, which is able to automatically detect anomalies at various situations and warn operators of dangerous activities, is extremely helpful to eldercare in providing real time vid-eo-based activity monitoring and functional assessment. Conventional surveillance systems consider using rule-based method to detect predefined dangerous activities [2], [4]. These approaches perform well in certain tasks such as fall detection [5], sleeping posture classification. Other methods consider action recognition and anomaly detection as supervised learning problems in which anomalies are treated as outlier to previous trained classifier of normal action patterns [3]. In this paper, we propose to develop a visual sensing system for nighttime action monitoring of elderly people at home(Fig.1). We capture images using Infrared cameras. We develop spatiotemporal filters for noise removal in IR images during very low-lighting conditions. We develop supervised learning methods to recognize their action patterns. The supervised learning algorithm allow the action model to be adapt to dynamic usage environments, thus achieve more robust and accurate results. Overview of our proposed system.