Acoustic Doppler Sensors for Surveillance
Acoustic Doppler measurements may be utilized to characterize walkers and to recognize them.
Background & Objective: Gait is a strong characteristic of individuals, and may be used to recognize them. Traditionally Gait has been characterized through video recordings. Subjects are detected, segmented out, and their image characteristics measured either directly or through stick models etc. These methods work best when the subject is walking at right angles to the vector to the camera. In this project we aim to develop a different acoustic Doppler based mechanism to characterize gait. The Doppler-based sensor will be cheaper than conventional video based sensor, and have complimentary capabilities.
Technical Discussion: Gait consists of a sequence of movements of the walker's limbs. During this movement, various parts of the body, such as the feet, shin, knees, thighs, arms, elbows etc. progress through a cycle of movements in which their velocity also changes cyclically. The ensemble of cyclic variations of these velocities characterizes the gait. If a high-frequency tone is incident on a walker, the frequency of the reflected signal get modulated by the velocities of the various moving body parts. The set and pattern of modulated frequencies in the reflected signal characterize the gait. We capture the reflected signal through a resonant transducer. The signal is frequency demodulated to enhance all frequency components. A series of cepstrum-like feature vectors are derived from it. We learn a Gaussian Mixture model for each subject from a short recording of training data. Walkers are recognized thereafter by regular MAP classification. Our results indicate that we are able to identify subjects with over 90% accuracy using this approach.
Outside Collaborations: Kaustubh Kalgaonkar, GA Tech, Intern.
Future Direction: Results indicate that we can not only identify subjects, but also classify them according to height, gender etc. Further research is being conducted in this direction. We have also developed a new time-series model to represent temporal patterns in the Doppler signal. These will also be investigated.
Contact: Bhiksha Raj
Technology Area: Multimedia
Modification Date: September 12, 2007
