Enabling and Advancing Human and Probabilistic Context Awareness for Smart Facilities and Elder Care
Abstract:
This project is developing methods for human context awareness in smart buildings, aging-in-place, smart authorization, and emergency response. We enable this using inference using radio frequency (RF) sensing networks, in which channel measurements are made by deployed wireless networking devices. In RF sensing networks, the network is the sensor. We are developing methods to learn human context for smart facilities and elder care, including:
- The localization of people in the environment, both those wearing RFID tags which identify them, and those who do not wear any device.
- The science of using line segment crossing-based tracking
- Activity recognition – combining RF sensing and other sensing methods to reliably identify what someone is doing, not only where they are.
- Fall detection – identifying that someone who lives alone has fallen and is not getting up.
- Breathing monitoring – estimating a stationary person's breathing rate, detecting whether or not they are breathing, and locating any breathing people in the area of the deployed network.
- We are also investigating the privacy implications of the fact that otherwise secure wireless networks can be used for the above purposes; and on the security of RF sensing networks to potential attacks.
We have made significant progress in the three year duration of this grant. We have made dramatic improvements to radio tomographic imaging (RTI), primarily using multi-channel measurements and new imaging algorithms, which allows us to locate people more accurately (within 30 cm) than possible in prior work. We developed new methods to make RTI robust to noise, and to better locate radio tagged people. We have developed multiple person tracking methods which can accurately count and track up to four people in an area. We have demonstrated, via international competition, that device-free localization can be more accurate than radio device localization, a remarkable result considering the latter's history of research and development. We have also developed theory and algorithms for tracking using line segment crossing information. Standard tracking methods assume point sensors. Many systems can detect that a person has crossed a line segment, but there are few methods to analyze or track using such information. We have developed several methods to learn context from unreliable sensor information. In particular, we have shown methods and middleware that allows an application to merge sensor information and increase its reliability in classifying what activity the person is engaged in. We have developed methods to reliably detect that a person has fallen using RF sensor network data, and we can estimate a person's breathing rate and locate a stationary person by finding the source of the breathing in an area. Finally, we have developed methods for security and privacy. An eavesdropper can deploy receivers outside of a building or secure area and learn what is going on inside, as non-metal walls are transparent to radio waves. No current wireless network is secure to this “radio window” attack, and we have demonstrated a simple transmitter power variation scheme will not make such networks secure to an eavesdropper with more than one receiver. These technologies have significant application to broadly impact cyber-physical systems. For smart and secure buildings, we need to know where people are, and camera-based systems are often seen as too invasive of privacy. Even when employees are required to wear RFID badges, we must be able to locate people who are guests or are not wearing a badge, for safety and security. For elder care, existing systems require a person to wear a device, but most people do not remember or want to wear it. Finally, sensors are often ambiguous and unreliable, and our developed methods for classification and application development accounts for the probabilistic nature of sensing and is robust to errors.
Submitted by Neal Patwari
on
Abstract:
This project is developing methods for human context awareness in smart buildings, aging-in-place, smart authorization, and emergency response. We enable this using inference using radio frequency (RF) sensing networks, in which channel measurements are made by deployed wireless networking devices. In RF sensing networks, the network is the sensor. We are developing methods to learn human context for smart facilities and elder care, including:
- The localization of people in the environment, both those wearing RFID tags which identify them, and those who do not wear any device.
- The science of using line segment crossing-based tracking
- Activity recognition – combining RF sensing and other sensing methods to reliably identify what someone is doing, not only where they are.
- Fall detection – identifying that someone who lives alone has fallen and is not getting up.
- Breathing monitoring – estimating a stationary person's breathing rate, detecting whether or not they are breathing, and locating any breathing people in the area of the deployed network.
- We are also investigating the privacy implications of the fact that otherwise secure wireless networks can be used for the above purposes; and on the security of RF sensing networks to potential attacks.
We have made significant progress in the three year duration of this grant. We have made dramatic improvements to radio tomographic imaging (RTI), primarily using multi-channel measurements and new imaging algorithms, which allows us to locate people more accurately (within 30 cm) than possible in prior work. We developed new methods to make RTI robust to noise, and to better locate radio tagged people. We have developed multiple person tracking methods which can accurately count and track up to four people in an area. We have demonstrated, via international competition, that device-free localization can be more accurate than radio device localization, a remarkable result considering the latter's history of research and development. We have also developed theory and algorithms for tracking using line segment crossing information. Standard tracking methods assume point sensors. Many systems can detect that a person has crossed a line segment, but there are few methods to analyze or track using such information. We have developed several methods to learn context from unreliable sensor information. In particular, we have shown methods and middleware that allows an application to merge sensor information and increase its reliability in classifying what activity the person is engaged in. We have developed methods to reliably detect that a person has fallen using RF sensor network data, and we can estimate a person's breathing rate and locate a stationary person by finding the source of the breathing in an area. Finally, we have developed methods for security and privacy. An eavesdropper can deploy receivers outside of a building or secure area and learn what is going on inside, as non-metal walls are transparent to radio waves. No current wireless network is secure to this “radio window” attack, and we have demonstrated a simple transmitter power variation scheme will not make such networks secure to an eavesdropper with more than one receiver. These technologies have significant application to broadly impact cyber-physical systems. For smart and secure buildings, we need to know where people are, and camera-based systems are often seen as too invasive of privacy. Even when employees are required to wear RFID badges, we must be able to locate people who are guests or are not wearing a badge, for safety and security. For elder care, existing systems require a person to wear a device, but most people do not remember or want to wear it. Finally, sensors are often ambiguous and unreliable, and our developed methods for classification and application development accounts for the probabilistic nature of sensing and is robust to errors.