Impacts of feature selection on classification of individual activity recognitions for prediction of crowd disasters

Abstract

We examined the possibility of feature selection using statistical based time frequency domain (SBTFD) extracted features for human activity recognition. This is to reduce the dimensionality of features on monitoring devices to improve accuracy and minimize false negative alarm for crowd disasters. We analysed 54 SBTFD features obtained from 22,350 instances comprising of a climb down, climb up, peak shake while standing, standing, still, and walking; as classes V1, V2, to V8. Also, the benchmark dataset of 274,214 instances from nine users for accelerometer signals. Both datasets were subjected to minimum redundancy maximum relevance with information gain (MRMR-IG), correlation and chi-square techniques to select the relevant SBTFD features. We applied ten-fold cross validation using WEKA with four classifiers to classify individual behaviour classes V1 to V8. We achieved 97.8% accuracy and false negative rate of 9.5% to save human lives from crowd disasters with seven features of MRMR-IG using RF.