Training Activity Recognition Systems Online Using Real-time Crowdsourcing

Automated activity recognition (AR) aims to provide contextaware information and support in a wide variety of applications, such as prompting systems and monitoring public spaces for danger. This type of recognition is typically easy for people, but difficult for automated systems due to the need for understanding the context and high-level reasoning behind actions. We present Legion:AR, a system that uses real-time crowdsourcing to generate reliable sets of activity labels. Legion:AR enables online training of existing learning models in deployed recognition systems. Our evaluations with 24 workers show Legion:AR is able to generate labels in real-time, even in the presence of privacy-protecting measures such as adding a veil to hide the user’s identity.