Research Study On Prayer Activity Recognition Model Using Convolutional Neural Network (PARC)

Alzheimer's disease is the most common type of dementia in many people today.People with Alzheimer's suffer from forgetfulness or memory loss,which may affect their daily activities.The prayer activity is a must and obligatory activity for Muslims to perform five times a day.However,it can be difficult for those who face forgetfulness problems in prayer.This usually occurs due to a lack of concentration,overlapped and multiple concurrent thoughts.We associate this with the early stage of Alzheimer disease.Advancement in smartphone devices with built-in sensor technology-based applications has allowed us to help people in performing daily activities with full of satisfaction and confidence.The Muslim population is around 1.9 billion,24.4%of the world's population.They perform specific kinds of religious activities(Pray),Each pray(Salat)consists of a different number of repetitive units(steps)called 'Rakats'(Prescribe and specific Movements).People accidentally miss the unit(Rakat)due to inattention during prayer.This leads to incomplete prayer activity.To solve this problem,an intelligent automated smartphone activity recognition monitoring app could be built to support physical activity in Alzheimer's patients,to guide worshipper.In our work,a smartphone sensor-based deep learning model is proposed to recognize the prayer activities(Qayam,Ruku,and Sujud).In this context,we proposed the complete approach including data preprocessing,data segmentation,deep feature extraction and CNN classification.We utilized a convolutional neural network(CNN)to extract deep features,which are task-dependent and non-handcrafted.Results were analyzed on four body positions.Since our dataset was imbalanced and we observed low results on an imbalanced data set.To deal with this problem,three states of the art balanced techniques were utilized to make data set.After getting a balanced data set,we observed high improvements in the performance of our model.SMOTe technique based balanced data set achieved the best results.The results are reported in terms of balanced and imbalanced data set across the different body position.On the unbalanced data set,the average accuracy of the left-hand position,right-hand position,left-trouser pocket.and right trouser pocket are 67.6%,65%,65.6%and 66%,respectively.On the other hand,the average accuracy of the left-hand position,right-hand position,left-trouser pocket.and right trouser pocket on a balanced data set is 93%,90%,91.6% and 91% respectively.