Adaptive landmine detection and localization system based on incremental one-class classification

Clearing large civilian areas of antipersonnel landmines is a very severe problem that should be solved efficiently. Although many methods have been developed for landmine detection and discrimination using ground penetrating radar data, the problem has not yet been properly solved, especially, as landmine and innocuous object classes are imbalanced. One-class classification is a competitive method for landmine detection as data are unbalanced, but it separates the target from outliers along the target class large variance directions, which results in higher error. As a solution, covariance-guided one-class support vector machine (COSVM) emphasizes low-variance projectional directions of the training data, which results in high accuracy and error minimization. However, in the case of a large-scale dataset, COSVM requires a large memory and enormous amount of training time. Moreover, it is inflexible with dynamic data. For these reasons, we investigate the effectiveness of incremental covariance-guided one-class support vector machine (ICOSVM) to build an adaptive landmine detection and localization system. In fact, the ICOSVM has the advantage of incrementally projecting the data onto low-variance directions, thereby improving detection performance. Experimental results have shown clearly the superiority and efficiency of the proposed landmine detection and localization system. (C) 2018 Society of Photo-Optical Instrumentation Engineers (SPIE).