Hidden Markov Models (HMM) have proved to be eective for detecting buried land mines using data collected by
a moving-vehicle-mounted ground penetrating radar (GPR). The general framework for a HMM-based landmine
detector consists of building a HMM model for mine signatures and a HMM model for clutter signatures. A
test alarm is assigned a condence proportional to the probability of that alarm being generated by the mine
model and inversely proportional to its probability in the clutter model. The HMM models are built based on
features extracted from GPR training signatures. These features are expected to capture the salient properties
of the 3-dimensional alarms in a compact representation. The baseline HMM framework for landmine detection
is based on gradient features. It models the time varying behavior of GPR signals, encoded using edge direction
information, to compute the likelihood that a sequence of measurements is consistent with a buried landmine.
In particular, the HMM mine models learns the hyperbolic shape associated with the signature of a buried mine
by three states that correspond to the succession of an increasing edge, a
at edge, and a decreasing edge.
Recently, for the same application, other features have been used with dierent classiers. In particular, the
Edge Histogram Descriptor (EHD) has been used within a K-nearest neighbor classier. Another descriptor is
based on Gabor features and has been used within a discrete HMM classier. A third feature, that is closely
related to the EHD, is the Bar histogram feature. This feature has been used within a Neural Networks classier
for handwritten word recognition. In this paper, we propose an evaluation of the HMM based landmine detection
framework with several feature extraction techniques. We adapt and evaluate the EHD, Gabor, Bar, and baseline
gradient feature extraction methods. We compare the performance of these features using a large and diverse
GPR data collection.