Meena Ganesh

Algorithm to automatically detect abnormally long periods of inactivity in a home

An algorithm has been developed to automatically construct individual models of normal activity within a home using motion sensor data. Alerts can be generated when a period of inactivity exceeds a normal length for a particular residence. Alerting frequency has been optimized on a total of 650 days of real data from four homes of seniors who live independently. Results suggest that an inexpensive system that does not require the occupant to push any buttons or wear any devices could nonetheless alert within hours if a senior is unusually inactive. Further, such algorithms may facilitate widespread deployment of smart home technology to persons with different behavior patterns and home layouts by using automatic learning in place of potentially tedious manual configuration.

Unobtrusive vital signs monitoring with range-controlled radar

Researchers at General Electric have developed a prototype demonstration system that can identify motion and measure breathing and heart rate in periods without general body motions using a range-controlled radar device, similar to motion detectors available in home security systems. With continuous monitoring of the activity and physical vital signs, warning criteria can be used to identify potential crises augmenting existing observational systems without requiring the observed individual to wear or attach any special devices. This has potential applications in correctional facilities to complement observation for inmates at risk of suicide. It has healthcare applications where attaching traditional vital sign monitoring equipment is difficult such as for neonatal or burn patients.

A spoof detection method for contactless fingerprint collection utilizing spectrum and polarization diversity

The paper presents a spoof detection technique employing multi-spectral and multi-polarization imaging for a contactless fingerprint-capture system. While multispectral imaging has been proven to enable spoof detection for contact fingerprint imagers, these imagers typically rely on frustrated total internal reflection that requires a planar fingerprint, achieved by contact. The multispectral imaging method is based primarily on the difference in the spectral absorption profile between a real finger and a fake one. This paper will describe the expansion of this capability using blue and red light with contactless imaging in conjunction with polarization. This new method uses images at various rotated linear polarizations (each image representing a different value of specular and diffuse components), which are used to create the feature vectors representing the spectral and polarization diversity. The software extracts complex wavelet transforms (CWT) and FFT features from the images and builds a supervised learning method to train Support Vector Machine (SVM) classifiers. Experimental data was collected from a diversity of human fingers and silicon based phantoms molded from the corresponding humans. Fake and actual fingerprints were collected using individuals with a large diversity in skin tone, age, and finger dimensions. Our initial results, with an accuracy rate of at least 83%, are promising and imply that using the polarization diversity can enhance the spoof detection performance.