Feasibility and Deployment Analysis of Wi-Fi-Based Human Activity Recognition on a Memory-Constrained Microcontroller

Abstract

Wi-Fi channel state information (CSI) can be used to recognize human activities without cameras or wearable sensors. It is not well known which recognition models can run on a microcontroller with little internal memory. This study runs CSI-based human activity recognition (HAR) directly on an Espressif ESP32-WROOM-32 module that has no external pseudo-static random-access memory (PSRAM). Using the UT-HAR and CSI-HAR datasets, we test classical classifiers, small convolutional neural networks (CNNs), and larger neural networks. We report accuracy, conversion to 8-bit integers, memory use, inference time, and energy. On the board, the convolutional and Transformer models need about 8–54 kB of working memory and take 9–321 ms per inference at 240 MHz. The recurrent and state-space models cannot be converted to TensorFlow Lite for Microcontrollers, and the Kolmogorov–Arnold network fails when the runtime prepares it. These failures come from the software tools, not from a lack of memory. The larger problem is accuracy on new people. On the three-user CSI-HAR dataset, testing on a person left out of training lowers accuracy by about 25–33 percentage points compared with a random split, and the 35-user CSI-Bench dataset shows a similar drop. In a 30-minute live test, the board turned 180,815 CSI packets into 2,825 windows without losing data and labeled each window 37.6 ms after it was complete. A separate test with one participant reached 46.9% accuracy within a session and 25.8% across sessions. For the models and datasets studied, recognizing new people is a bigger problem than fitting the model in memory.