Towards adaptation in sentient spaces
Sharad Mehrotra, Ronen Vaisenberg · 2012
Recent advances in sensing technologies enabled infusing information technology into physical processes. This offers unprecedented opportunities, the impact of which will rival (if not exceed) the opportunities created by the WWW. Such sensor enabled environments can realize sentient spaces that have the potential to revolutionize almost every aspect of our society. Sentient systems observe the state of the physical world, analyze and act based on it. Sentient systems enable a rich set of application domains including smart video surveillance, situational awareness for emergency response and social interactions in instrumented office environments to name a few. In this thesis, we utilize additional information, namely the semantics of the monitored world, to improve the adaptation ability of sentient systems. We design real-time algorithms that utilized these models to address challenges such as data collection in the presence of resource constraints, sensor actuation and automatic re-calibration of analysis algorithms for correct monitoring. Our approach abstracts the application level from sensor level challenges, namely: data collection, actuation and re-calibration challenges. We propose addressing these challenges by a middleware layer that predicts the future state of the monitored environment and automatically schedules, actuates and re-calibrates sensors to maximize an application specified reward function. We evaluate our approach in a monitored building setting in which surveillance cameras are using to detect events taking place in the building. Our results suggest that sensor level challenges can be abstracted and effectively addressed by a middleware layer. We believe that our approach will serve as a fundamental building block for building the next generation sentient systems.