Evaluation of various feature extraction methods for landmine detection using hidden Markov models

Hidden Markov Models (HMM) have proved to be effective 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 confidence 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 flat edge, and a decreasing edge. Recently, for the same application, other features have been used with different classifiers. In particular, the Edge Histogram Descriptor (EHD) has been used within a K-nearest neighbor classifier. Another descriptor is based on Gabor features and has been used within a discrete HMM classifier. A third feature, that is closely related to the EHD, is the Bar histogram feature. This feature has been used within a Neural Networks classifier 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.