On motion-sensor behavior analysis for human-activity recognition via smartphones

A wealth of sensors on smartphones has greatly facilitated people's life, which may also provide great potential for accurate human activity recognition. This paper presents an empirical study of analyzing the behavioral characteristics of smartphone inertial sensors for human activity recognition. The rationale behind is that different human activities would cause different levels of posture and motion change of smartphone. In this work, an Android application was run as a background job to monitor data of motion sensors. Sensory data from motion sensors (mainly including accelerometer and gyroscope data) were analyzed to extracted time-, frequency-, and wavelet-domain features for accurate and fine-grained characterization of human activities. Classification technique were applied to build both personalized model and generalized model for discriminating five daily human activities: going downstairs, going upstairs, walking, running, and jumping. Analyses conducted on 18 subjects showed that these human activities can be accurately recognized from smartphone-sensor behavior, with recognition rates expressed by the area under the ROC curve ranging from 84.97% to 90.65%. We also discuss a number of avenues for additional research to advance the state of the art in this area.