Deep-Learning-Based Occupant Counting by Ambient RF Sensing

With the proliferation of the Internet of Things and radio-frequency identification (RFID), the ambient radio signals can be leveraged for indoor occupant monitoring. In this article, we have employed passive RFID tags in the ambient for occupant counting by a deep-learning solution. The reader collects both carrier phase and received signal strength from each tag, which are inputs to a convolutional neural network. A novel background calibration is proposed to reduce phase offsets and noises in the presence of heavy multipath, which further improves model accuracy. Our results show satisfactory performance, with 0.82 probability for detecting the correct number of occupants, and 1.0 if ±1 error is permitted. The model also exhibits occupant location and posture independent learning, allowing limited and faster training data collection. To demonstrate generalized learning without strong bias to indoor setup, we have also transferred this pre-trained model to another similar-sized room, achieving 0.85 - 1.0 accuracy for different tag-receiver placements and furnishing.