A Real-Time Crowd Estimation Based on RSSI Measurements Using Convolutional Neural Network

This paper presents an indoor crowd estimation system based on Wi-Fi received signal strength indicator (RSSI). The estimation is performed based on the signal absorption rate of a human body to the Radio Frequency signals using a previously trained Convolutional Neural Network (CNN) model. For the experimental setup, several transceiver topolgies have been investigated. The diversity scheme that consists of one wireless transmitting node and two receiving nodes has shown the maximum accuracy in classifying new measurements with the minimum complexity. The collected data is confirmed with the theory of radio frequency propagation and the secondary wave principle. Moreover, to enable the real-time prediction, a transmitter-independent statistical data calibration is provided which removed the need for re-training and increases the model recall rate in the cases of restarting the transmitter or changing its specifications. For this, an automation system is built to handle data collection, pre-processing and real time prediction. The presented CNN model has shown about 93% minimum accuracy for up to the maximum capacity of the current indoor space.