Enhanced context-aware framework for individual and crowd condition prediction

Abstract

Context-aware framework is basic context-aware that utilizes contexts such as user with their individual activities, location and time, which are hidden information derived from smartphone sensors. These data are used to monitor a situation in a crowd scenario. Its application using embedded sensors has the potential to monitor tasks that are practically complicated to access. Inaccuracies observed in the individual activity recognition (IAR) due to faulty accelerometer data and data classification problem have led to its inefficiency when used for prediction. This study developed a solution to this problem by introducing a method of feature extraction and selection, which provides a higher accuracy by selecting only the relevant features and minimizing false negative rate (FNR) of IAR used for crowd condition prediction. The approach used was the enhanced context-aware framework (EHCAF) for the prediction of human movement activities during an emergency. Three new methods to ensure high accuracy and low FNR were introduced. Firstly, an improved statistical-based time-frequency domain (SBTFD) representing and extracting hidden context information from sensor signals with improved accuracy was introduced. Secondly, a feature selection method (FSM) to achieve improved accuracy with statistical-based time-frequency domain (SBTFD) and low false negative rate was used. Finally, a method for individual behaviour estimation (IBE) and crowd condition prediction in which the threshold and crowd density determination (CDD) was developed and used, achieved a low false negative rate. The approach showed that the individual behaviour estimation used the best selected features, flow velocity estimation and direction to determine the disparity value of individual abnormality behaviour in a crowd. These were used for individual and crowd density determination evaluation in terms of inflow, outflow and crowd turbulence during an emergency. Classifiers were used to confirm features ability to differentiate individual activity recognition data class. Experimenting SBTFD with decision tree (J48) classifier produced a maximum of 99:2% accuracy and 3:3% false negative rate. The individual classes were classified based on 7 best features, which produced a reduction in dimension, increased accuracy to 99:1% and had a low false negative rate (FNR) of 2:8%. In conclusion, the enhanced context-aware framework that was developed in this research proved to be a viable solution for individual and crowd condition prediction in our society.