HeteroSense: An Occupancy Sensing Framework for Multi-Class Classification for Activity Recognition and Trajectory Detection

Occupancy behavior in buildings is challenging to define and quantify because of their stochastic, diverse and complex nature. Indoor environment and energy usage in buildings are significantly influenced by the actions and behavior of occupants. Accurate estimation and prediction of occupancy behaviours such as occupancy trajectories with datasets from deployed sensors can reduce energy consumption and facilitate intelligent building operations. However, identifying affordable sensors for estimating occupancy in buildings is still a challenge. To detect occupancy state in real-time, we propose a smart fusion of not necessarily optimal expensive sensors but accurate enough sensors. Sensor fusion is opted to boost the performance of the framework. In this paper, we propose HeteroSense an occupancy sensing framework that uses a label-based approach for activity recognition using machine learning classification algorithms. Based on the sensor data collected from heterogeneous sensing modalities, an algorithm is designed for trajectory detection. This paper also discusses the challenges faced during the design phase for the deployment and it summarizes the potential improvements in the field of occupancy sensing for energy efficient buildings.