Counting people with WiFi signals

A privacy-friendly way to count people in a room using WiFi Channel State Information instead of cameras. Phase-rotation augmentation took an LSTM from 49% to 67% accuracy.

Bachelor thesis, Amrita Vishwa Vidyapeetham2023

Diagram of the method, from WiFi signals in two rooms through ESP32 capture, augmentation and model comparison to a people count
Diagram of the method, from WiFi signals in two rooms through ESP32 capture, augmentation and model comparison to a people count.

Cameras can count people, but people do not like being filmed. WiFi signals change as bodies move through a room, and those changes can be measured.

The setup

With a team of four, I captured Channel State Information (CSI) from two ESP32 modules in two rooms, on both the 2.4 GHz and 5 GHz bands.

The models

We compared an ANN, an LSTM and a complex-valued neural network. The LSTM did best. Data was scarce, so we augmented it with jittering, rotation and interpolation, and checked that the augmented data still made sense using k-means and agglomerative clustering. Phase-rotation augmentation lifted the LSTM from 49% to 67% accuracy.

More work