Occupancy sensing and activity recognition with cameras and wireless sensors

We present a system work combining visual cameras and wireless sensors for human occupancy detection and activity recognition. We describe our testbed system, data collected from a human subject study, observations from long-term occupancy experiments, and preliminary analytical results. We apply machine learning algorithms to the human activity recognition data, and identify challenges in applying the state-of-the-art deep learning techniques to wireless sensing of human activity. We find that packet loss due to wireless interference has a significant effect on time series classification. We also find that the convolutional neural networks significantly outperforms the conventional support vector machine method, but further experiments need to be performed to investigate environment-independent classification and the overfitting issue. Finally, we discuss future research topics that can use our testbed of wireless sensors and visual cameras to automate data labeling in deep learning model training.