COPO: A Novel Position-Adaptive Method for Smartphone-Based Human Activity Recognition

摘要

In recent years, smartphone-based human activity recognition has become a promising research field of mobile computing, and is widely applied in inertial positioning, fall detection, and personalized recommendation. In practical scenario, smartphone can be placed at several body positions, such as trouser pocket, jacket pocket and so on. Since data is collected from the accelerometer embedded in smart-phone, different body locations cannot generate consistent data for the same activity. As a result, the samples at a new position usually obtains low recognition rate from the classifier trained by the original data collected from other positions. In this paper, we propose a COntinuity-based POsition-adaptive recognition method, abbreviated COPO, for dealing with this problem. Considering the continuous results with high probability of correct recognition, we select them as the retraining data in COPO for updating the initial classifier. To prove the effectiveness of retraining data selecting method theoretically, we use Hidden Markov Model (HMM) to calculate the probability that the continuous recognition results Eire correctly recognized. Finally, a number of experiments are designed to verify our COPO, including data collection, performance comparison, and parameter analysis. The results show that the recognition rate of COPO is 2.62 % higher than other common methods.