Sensor fusion and feature-based human/animal classification for Unattended Ground Sensors

In this paper we examine novel signal processing algorithms that utilize wavelet statistics, spectral statistics and power spectral density in addition to cadence and kurtosis for robust discrimination of humans and animals in an Unattended Ground Sensor (UGS) field. The wavelet statistics are based on the average, variance and energy of the third scale residue. The spectral statistics are based on amplitude and shape features. A learning classifier approach is used for discrimination. Training data consists of scripted events with humans walking/running along known paths; as well as riders on horses and moving vehicles on a two node sensor network. Natural events are recorded when animals, such as cows, coyotes, rabbits and kangaroo rats are in the vicinity of the sensor nodes. Each node has a three axis accelerometer and a three axis geophone and one node has a low frequency geophone in addition. In our work we use the C4.5 classifier which is a tree-based classifier and is capable of modeling complex decision surfaces while simultaneously limiting the complexity of the trees through pruning schemes. The classifier is tested on test data and the performance results are very promising — results indicate that UGS-only systems are indeed feasible for border security. The development of a successful signal processing solution to better discriminate between humans and animals would be very valuable to the Department of Homeland Security and our paper will summarize these new results.