Activity logging using lightweight classification techniques in mobile devices

Automated activity recognition enables a wide variety of applications related to child and elderly care, disease diagnosis and treatment, personal health or sports training, for which it is key to seamlessly determine and log the user's motion. This work focuses on exploring the use of smartphones to perform activity recognition without interfering in the user's lifestyle. Thus, we study how to build an activity recognition system to be continuously executed in a mobile device in background mode. The system relies on device's sensing, processing and storing capabilities to estimate significant movements/postures (walking at different paces-slow, normal, rush, running, sitting, standing). In order to evaluate the combinations of sensors, features and algorithms, an activity dataset of 16 individuals has been gathered. The performance of a set of lightweight classifiers (Naive Bayes, Decision Table and Decision Tree) working on different sensor data has been fully evaluated and optimized in terms of accuracy, computational cost and memory fingerprint. Results have pointed out that a priori information on the relative position of the mobile device with respect to the user's body enhances the estimation accuracy. Results show that computational low-cost Decision Tables using the best set of features among mean and variance and considering all the sensors (acceleration, gravity, linear acceleration, magnetometer, gyroscope) may be enough to get an activity estimation accuracy of around 88 % (78 % is the accuracy of the Naive Bayes algorithm with the same characteristics used as a baseline). To demonstrate its applicability, the activity recognition system has been used to enable a mobile application to promote active lifestyles.