Human identification experiments using acoustic micro-Doppler signatures

Active acoustic scene analysis is a promising approach to distributed persistent surveillance in sensor networks. We report on the design of bandpass sampling technique for an acoustic micro-Doppler sonar to reduce the data rate to as low as 85 kbps. We then explore the use of Gaussian mixture models for human identification. We compare the classification performances using different feature vectors and from different sampling schemes. We show that the use of differential cepstral vectors of context length 2 improves the classification accuracy. We also show that the classification performance of the bandpass sampling system with an 8-bit resolution is still over 90% on a database consisting of 160 gait signatures from 8 individuals.