Proposed here is a series of techniques exploiting micro-Doppler ultrasonic sensors
capable of characterizing various detected mammalian targets based on their
physiological movements captured a series of robust features. Employed is a
combination of unique and conventional digital signal processing techniques arranged in
such a manner they become capable of classifying a series of walkers. These processes
for feature extraction develops a robust feature space capable of providing discrimination
of various movements generated from bipeds and quadrupeds and further subdivided into
large or small. These movements can be exploited to provide specific information of a
given signature dividing it in a series of subset signatures exploiting wavelets to generate
start/stop times. After viewing a series spectrograms of the signature we are able to see
distinct differences and utilizing kurtosis, we generate an envelope detector capable of
isolating each of the corresponding step cycles generated during a walk. The walk cycle
is defined as one complete sequence of walking/running from the foot pushing off the
ground and concluding when returning to the ground. This time information segments
the events that are readily seen in the spectrogram but obstructed in the temporal domain
into individual walk sequences. This walking sequence is then subsequently translated
into a three dimensional waterfall plot defining the expected energy value associated with
the motion at particular instance of time and frequency. The value is capable of being
repeatable for each particular class and employable to discriminate the events. Highly
reliable classification is realized exploiting a classifier trained on a candidate sample
space derived from the associated gyrations created by motion from actors of interest.
The classifier developed herein provides a capability to classify events as an adult
humans, children humans, horses, and dogs at potentially high rates based on the tested
sample space. The algorithm developed and described will provide utility to an
underused sensor modality for human intrusion detection because of the current high-rate
of generated false alarms. The active ultrasonic sensor coupled in a multi-modal sensor
suite with binary, less descriptive sensors like seismic devices realizing a greater
accuracy rate for detection of persons of interest for homeland purposes.
|