Real-Time Transfer and Evaluation of Activity Recognition Capabilities in an Opportunistic System

This paper describes and evaluates the challenging feature of an opportunistic activity recognition system to train a newly discovered sensor with the available sensing devices to recognize activities at runtime. The term ”opportunistic” means that the system does not operate with a fixed set of sensor devices, but uses and configures the currently available sensors that just happen to be available. Therefore, the paper presents a reference implementation of an opportunistic system, referred to as OPPORTUNITY Framework, and demonstrates the transfer of recognition capabilities from a fused multisensor ensemble to an untrained sensing device within the system in a real-world setup. Main contribution of the paper is the evaluation of the approach by describing an experimental setup and presenting results in terms of accuracy and recognition rate from the machine-learning perspective as well as from the framework and system perspective by comparing predicted classes from the teaching sensor set and the newly trained sensor to obtain QoS parameters.