TY - GEN
T1 - Recognizing daily living activity using embedded sensors in smartphones
T2 - 12th International Conference on Advanced Data Mining and Applications, ADMA 2016
AU - Ruan, Wenjie
AU - Chea, Leon
AU - Sheng, Quan Z.
AU - Yao, Lina
PY - 2016
Y1 - 2016
N2 - Smartphones are widely available commercial devices and using them as a basis to creates the possibility of future widespread usage and potential applications. This paper utilizes the embedded sensors in a smartphone to recognise a number of common human actions and postures. We group the range of all possible human actions into five basic action classes, namely walking, standing, sitting, crouching and lying. We also consider the postures pertaining to three of the above actions, including standing postures (backward, straight, forward and bend), sitting postures (lean, upright, slouch and rest) and lying postures (back, side and stomach). Training data was collected through a number of people performing a sequence of these actions and postures with a smartphone in their shirt pockets. We analysed and compared three classification algorithms, namely k Nearest Neighbour (kNN), Decision Tree Learning (DTL) and Linear Discriminant Analysis (LDA) in terms of classification accuracy and efficiency (training time as well as classification time). kNN performed the best overall compared to the other two and is believed to be the most appropriate classification algorithm to use for this task. The developed system is in the form of an Android app. Our system can real-time accesses the motion data from the three sensors and on-line classifies a particular action or posture using the kNN algorithm. It successfully recognizes the specified actions and postures with very high precision and recall values of generally above 96%.
AB - Smartphones are widely available commercial devices and using them as a basis to creates the possibility of future widespread usage and potential applications. This paper utilizes the embedded sensors in a smartphone to recognise a number of common human actions and postures. We group the range of all possible human actions into five basic action classes, namely walking, standing, sitting, crouching and lying. We also consider the postures pertaining to three of the above actions, including standing postures (backward, straight, forward and bend), sitting postures (lean, upright, slouch and rest) and lying postures (back, side and stomach). Training data was collected through a number of people performing a sequence of these actions and postures with a smartphone in their shirt pockets. We analysed and compared three classification algorithms, namely k Nearest Neighbour (kNN), Decision Tree Learning (DTL) and Linear Discriminant Analysis (LDA) in terms of classification accuracy and efficiency (training time as well as classification time). kNN performed the best overall compared to the other two and is believed to be the most appropriate classification algorithm to use for this task. The developed system is in the form of an Android app. Our system can real-time accesses the motion data from the three sensors and on-line classifies a particular action or posture using the kNN algorithm. It successfully recognizes the specified actions and postures with very high precision and recall values of generally above 96%.
UR - http://www.scopus.com/inward/record.url?scp=85000428307&partnerID=8YFLogxK
U2 - 10.1007/978-3-319-49586-6_17
DO - 10.1007/978-3-319-49586-6_17
M3 - Conference proceeding contribution
AN - SCOPUS:85000428307
SN - 9783319495859
T3 - Lecture Notes in Artificial Intelligence
SP - 250
EP - 265
BT - Advanced data mining and applications
A2 - Li, Jinyan
A2 - Li, Xue
A2 - Wang, Shuliang
A2 - Li, Jianxin
A2 - Sheng, Quan Z.
PB - Springer, Springer Nature
CY - Cham
Y2 - 12 December 2016 through 15 December 2016
ER -