Executive Summary : | Human activity learning is a heavily investigated research domain owing to its significant implications on a variety of use cases including smart healthcare, home automation, security and emergency, etc. Internet of Things (IoT) by utilizing the wide spectrum of heterogeneous sensors enables pervasive and continuous sensing of various environmental parameters and human actions/interactions. Among several approaches proposed in the literature for accurately learning human activities, ambient sensing promises a unique advantage of preserving user privacy by not capturing any sensitive data through acoustic and optical channels. Usually the sensors are either embedded into the environment (monitoring ambient temperature, humidity, light, proximity etc.) or worn/carried by the human in terms of wearables/smartphones (monitoring movement, posture/gesture, physiological vitals etc.). However, these low-cost sensors are highly failure-prone, non-persistent and unreliable. Human activity learning including recognition, discovery or prediction is a complex problem as the performance of the machine learning model is greatly affected by the quality and quantity of data generated from these heterogeneous sensors. Also, learning activity from data generated by a single source may not possibly capture all aspects of an event as activities could be overlapping in space and time, hierarchical and complex. Multimodal data fusion has been quite a prominent field of research in this direction that promotes improved knowledge inference. However, centralized fusion often fails to capture the joint contribution and interdependency among diverse sensory inputs, in a typical distributed and shared sensing environment, running multiple intelligent services from different backgrounds. Context-aware computing has exhibited remarkable performance benefits in the past while addressing a diverse set of problems that accounts for the influence of varying environmental dynamics. The advantages of context-aware activity recognition is multi-faceted: 1. Various contextual parameters like calendar events, time of day, weather conditions, or subject-specific traits could add value to the inference module by addressing the problem of high spatiotemporal variation often exhibited in sensory data. 2. The richness of a context could be utilized to shape up the data pipeline for optimal resource utilization. 3. Depending on the non-occurrence of an event, non-availability or erroneous reporting of data, the sampling rate of the sensors as well as granularity of the filtration can be adapted. Finally, integrating context with the sensory data could even be helpful in terms of generating the ground truth for the machine learning model, which is free from human intervention, and rather explainable. This proposal aims to develop a context-aware decentralized framework supporting the collaboration of sensors towards improved human activity learning while optimally utilizing the resources. |