Goal oriented recognition of composed activities for reliable and adaptable intelligence systems

The emerging availability of already deployed sensors that can be utilized for activity and context recognition raised a new paradigm. This paradigm called opportunistic sensing utilizes the available sensing infrastructure for activity and context recognition. This work focuses on utilizing this dynamically varying sensing infrastructure to recognize high-level composed activities in an adaptable way. The proposed methods use activity relations modeled in an ontology. This domain knowledge is used to dynamically configure hidden Markov models (HMM) and evidential networks. These models are popular in activity and context recognition systems due to their high recognition accuracy. A goal oriented approach is proposed to dynamically create and instantiate these models. The goal encapsulates the recognition purpose of the activity and context recognition system and is expressed in an abstracted and semantic manner. This flexible approach utilizes the opportunistic sensing principles. It directs the dynamic configuration of the activity and context recognition system during runtime. The configured recognition models are based on the recognition purpose of the system, and the configured sensing ensemble depends on the available sensing infrastructure. This enables the dynamic configuration and adaption of the activity and context recognition system during runtime to detect composed and time sequenced activities using HMMs or evidential networks in an opportunistic way.