Phone Position Independent Recognition of FineWalking Modes with Phone Sensors
Rui Zhou, Xiang Lu, Shuai Lu, Yang Fu and Mingjie Tang
Human activity is a key element for context-awareness and health monitoring. However, it is difficult to infer human activities from uncertain mobile sensor data, due to sensor noise and various phone positions. Activities like standing, sitting, walking, running have been investigated widely, while research on identifying fine-grained walking modes without prior knowledge of phone positions is still in its infancy. This paper proposes a two-layer algorithm to recognize different types of walking (static, slow walking, medium walking, fast walking and running) and different ways of going upstairs and downstairs (by elevator, by escalator and by stairs) using only phone sensors without prior knowledge of phone positions. Horizontal walking, upstairs and downstairs are distinguished on the first layer using only barometer data. Different types of horizontal walking are distinguished on the second layer using the features extracted from 3-axis acceleration sequences through wavelet transform and singular value decomposition. Recognition of different ways of going upstairs and downstairs is performed with altitude sequences and 3-axis acceleration sequences. Three classification algorithms of Support Vector Machine, Bayesian Network and Decision Tree are applied separately to fulfill the classification tasks. Evaluations show that the three algorithms achieve comparable recognition results and Decision Tree achieves the best with the mean recognition accuracy of horizontal walking modes as 95% and vertical walking modes as 97.7%, independent of phone positions.
Keywords: Activity recognition, Bayesian network, decision tree, phone position, support vector machine