Human activity recognition prediction for crowd disaster mitigation

Abstract

Context sensing and context acquisition have remained challenging issues in addressing the problems relating to Human Activity Recognition (HAR) for mitigation of crowd disasters. In this study, classification algorithms for higher accuracy of HAR which may be significantly low for effective stampede prediction in crowd disaster mitigation were investigated. The proposed HAR prediction model consists of mobile devices (mobile phone sensing) that can be used for monitoring a crowd scene in group movement: it employs tri-axial accelerometer sensors as well as other sensors like digital compass to capture relevant raw data from participants. In a previous study of stampede prediction, HAR accuracy of 92% was achieved by implementing J48, a Decision Tree, (DT) algorithm for context acquisition using a data mining tool. The implementation of the proposed model using K-Nearest Neighbour (KNN) algorithm with real time raw data collected with smartphones provided easily deployable context-awareness mobile Android Application Package (.apk) for effective crowd disaster mitigation and real time alert to avoid occurrence of stampede. The results gave 99.92% accuracy for activity recognition which outperforms the aforementioned study. Our results will forestall possible instances of false stampede alarm and reduce instances of unreported cases with higher accuracy if implemented in real life.