Identification of an effective learning approach to landmine detection

The Robot based landmine detection problem is a multiphase problem in which one element is the classification of landmines and clutter. To design an efficient and effective classification model requires considering factors such as the failure to detect a landmine, detection time and the high amount of false alarms that occur due of improper classification. In the absence of an extensive analysis on the effectiveness of such models, this project aims to analyze 5 different classifiers namely: Hidden Markov Model, Support Vector Machine, Artificial Neural Network, Gradient Boosted Decision Tree and Adaptive Boosted Decision trees. Two GPR based datasets have been used both of which are open source and contain data for foliage and dry, desert type soils respectively. To make the study comprehensive in terms of class label proportion as well, various ratios of mine to non-mine data is considered. The comparison of the models has been done using confusion matrices with its associated measures. Based on this, a selection table has been designed which allows the user to select the classifier that is most likely to give the best performance with respect to a preferred metric and available training dataset that a user prefers to use for training.