Comparing chirplet-based classification with alternate feature-extraction approaches for outdoor intrusion detection using a pir sensor platform

Prior work by a subset of the authors led to the development of a Pyroelectric Infrared (PIR) sensor platform for the purposes of distinguishing in an outdoor environment, between human and animal intrusion while rejecting false alarms arising from wind-blown vegetation. The algorithm employed there, modeled the intrusion signal as a linear combination of chirplets. The extracted chirplet-based features were fed to Support Vector Machine (SVM) for the classification. This resulted in a platform that, under the tested conditions, resulted in high classification accuracy, in excess of 95%. The current paper is aimed at determining the extent to which the classification accuracy could be attributed to the use of the algorithm employed. Fifteen different algorithms for intrusion detection and classification are examined in the current paper, these operate on a database that builds on top of the earlier database. These fifteen algorithms correspond to the different possible pairings obtained by selecting from among 5 feature-vectors and 3 classifiers. The results show that the chirplet-based feature extractor to play a major role in achieving high-accuracy classification, easily beating the performance of the other feature extractors, particularly in terms of the more challenging task of separating intrusion from vegetative clutter. The two principal conclusions that can perhaps be drawn here are that (a) chirplet-decomposition can be a very effective feature-extractor in the case of a PIR signal and (b) it is important to incorporate domain knowledge where possible in designing the most effective classification algorithms.