Non-Intrusive Activity Detection and Prediction in Smart Residential Spaces
Aniruddha Patel, Chinmay Prabhudesai, Barış Akşanlı · 2018
Non-intrusive human activity detection and prediction is an important and challenging problem in smart and pervasive spaces. The advantages of such a design are the reduced dependency on the users and fewer security/privacy concerns. However, these also make it difficult to effectively and accurately understand the activities in real-time. In residential spaces, this can be even a bigger challenge due to nonuniform space boundaries and multiple people sharing the space. In this paper, we present a system, that consists of both hardware and software components, capable of detecting and predicting human activities in a smart residential environment. Our system deploys a finite state machine-based activity detection with 96% accuracy in real-time. Afterwards, we use several machine learning methods to create an effective activity prediction framework. We demonstrate that we can achieve up to 98.5% activity prediction accuracy with 4ms delay, making it a perfect real-time system example. Since our smart and pervasive space implementation does not use any intrusive sensor or data acquisition unit (such as wearables, camera, or audio sources), we reduce the dependency to the user and potential security/privacy issues.