Classifying Abnormal Activities in Exam Using Multi-class Markov Chain LDA Based on MODEC Features

Paper

Abstract

In this paper, we apply MCMCLDA (Multi-class Markov Chain Latent Dirichlet Allocation) model to classify abnormal activity of students in an examination. Abnormal activity in exams is defined as a cheating activity. We compare the usage of Harris3D interest point detector and a human joints detector, MODEC (Multimodal Decomposable Models), as the feature detector. Experiment results show that using MODEC to detect arm joints and head location as interest point gives better performance in accuracy and computational time than Harris3D when classifying cheating activity. MODEC suffers low accuracy due to its inability to differentiate elbow and wrist when the object wears clothes with indistinguishable colors. Meanwhile, Harris3D detects too many irrelevant interest point to recognize cheating activity reliably.