Performance evaluation of classifiers on activity recognition for disasters mitigation using smartphone sensing

• Main Authors: Fatai Idowu, Sadiq, Selamat, Ali, Ibrahim, Roliana
• Format: Article
• Language: English
• Published: Penerbit UTM Press 2015
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://eprints.utm.my/id/eprint/58784/1/FataiIdowu2015_PerformanceEvaluationOfClassifiersOnActivityRecognition.pdf; http://eprints.utm.my/id/eprint/58784/; http://dx.doi.org/10.11113/jt.v77.6320
• Institution: Universiti Teknologi Malaysia

Activity recognition (ARs) is a classification problem that cuts across many domains. The introduction of ARs accuracy which may be significantly low with decision tree algorithm and the use of smartphone sensing in previous studies has proven its relevance for effective disaster mitigation in our society. Smartphone sensing is an approach found to be useful for activity recognition to monitor people in large gatherings due to the power of embedded sensors on the handheld devices. In this paper, a multitask activity recognition architecture is proposed for proper monitoring of people in large gatherings to control disaster occurrences in crowd, flood, road and fire accidents using related activity scenario in time of danger. We implement the proposed architecture to determine the outcome of activity recognized with K-nearest neighbour (KNN) for k= 3 and 4 to compare performance to that of weka using accelerometer and digital compass (dc) sensors on the same dataset. The results of ARs accuracy of 100% and 99% in weka, 85% and 89% with KNN shows an improved performance in both tools. The performance of Multilayer Perceptron (MLP), Support Vector Machine (SVM), Naives baye (NB), Decision tree (DT), against KNN were investigated using precision, recall and f-measure in weka as well. The results show significant improvement with performance parameters on accelerometer and dc against the use of accelerometer sensor only with KNN and DT having low number of classified activity recognized on training and testing data.