Indoor scene and human activity analysis with wireless binary sensor networks.

摘要

Indoor scene analysis aims to extract the scene information and perceive the situations. The goal of human activity study is to recognize human subjects' behavior patterns. Indoor scene analysis and human activity recognition can be used to enhance the performance of various applications ranging from healthcare to surveillance and energy efficient building. Our research is focusing on human sensing, behavioral biometrics, and situation awareness in indoor environment. The goal of my research is to build low-cost wireless sensing systems, design compressive sampling structures, and develop lightweight scene analysis and human activity recognition algorithms to form a human-centric intelligent sensing framework for the applications in smart environment. This work presents a framework for indoor scene analysis and human activity recognition based on pyroelectric infrared sensor and fiber-optic sensor.;The main accomplishments of this thesis include the following aspects:;(1) Wireless binary sensing infrastructure establishment. We have built two low-cost, low data throughput, wireless sensing infrastructures based on pyroelectric infrared sensor and fiber-optic sensor for indoor scene analysis and human activity recognition. In these systems, binary sensing technique has been employed to further reduce the data load and computation complexity.;(2) Geometric sampling structure exploration. Sampling structure plays a key role in efficient information acquisition, data load reduction, and intrinsic feature determination. We have explored efficient sampling structures for both pyroelectric infrared and fiber-optic sensing systems by employing visibility modulation and space encoding schemes, respectively.;(3) Indoor scene modelling and representation. Different from the conventional object-based methods which focus on individual characteristics, we have built a statistical, low-dimensional feature based scene representation model. Such model can not only discover the number of people, but also facilitate localization and identication.;(4) Ground truth feature selection. We have created both informative and non-informative hierarchical inference models to seek the ground-truth scene bases. Meanwhile, various approaches including maximum a posteriori (MAP), expectation-maximization (EM), geometry embedding, variational Bayesian (VB), have been investigated to enhance the convergence as well as the robustness of the models.;(5) Data driven and reasoning approaches integration. Compared with the instructed sensing modality and conventional human activity recognition approaches in wireless sensing systems, we have developed a new framework which combines data driven and reasoning methods to achieve learning and inference. More specifically, the sensing context, situation context, and environment context are utilized to facilitate information professing.