An Efficient Human–Animal Footfall Classifier for Desert Perimeter Surveillance Using Seismic Sensors

This letter proposes a hidden Markov model–Gaussian mixture Model-based tandem approach for efficient classification of human and animal footfall signals acquired from seismic sensors buried in desert terrain. Since desert seismic signal bandwidth is below 1 kHz, we propose to design triangular filters over nonlinear scale and extract log-sigmoid frequency cepstral coefficients that form the observation vector for tandem model. We conduct extensive real-life testing with desert terrain data and demonstrate an overall accuracy of 98.78% for the proposed approach. We show that the proposed approach renders very low false positive as well as false negative rate, outperforms popular machine learning and deep learning classifiers for nonlinear data, and is amenable for real-world implementation.