In a laboratory test setting, we have collected annotated sensor traces from eight test subjects (two separate data collection runs each) to recognize four activities: sitting, walking, running, and riding a stationary bike. Using data hold one out training, we were able to achieve 99.5% accuracy and 97% accuracy cross-person (where none of the training data was from the test subject).
iLearn is a set of tools for Apple's iPhone that allows us to collect annotated sensor traces from the accelerometer as well as compute accelerometer features and perform real-time activity classification. The software is built on the standard Apple SDK and includes models for inferring some common activity sets (e.g., exercise). This classification can then be used by additional applications on the iPhone or sent to web servers via the iPhone’s highly available internet connection. iLearn also provides a set of desktop tools for using annotated sensor traces to build brand new activity inference models with open source Weka software.