SmartGate: TinyML Occupancy Classifcation using 24 GHz FMCW Radar Gate Vectors on ESP32

Abstract

This paper introduces SmartGate, a full-stack edge AI system for classification of indoor room occupancy using labeled energy feature vectors extracted from a low-cost HLKLD2410C 24 GHz FMCW radar sensor module. Since common techniques for occupancy measurement are RGB cameras, passive IR motion detectors, or high-cost radar platforms that can expose micro-level IQ data, the system leverages the gate energy outputs supplied through the sensor UART. A dataset of 1,344 labeled feature samples from 3 classes (empty, still person sitting, moving person in room) was collected and trained with the Edge Impulse AutoML platform using a small two-layer fully-connected network. On-chip int8 quantization using the EON Compiler produces a deployable library for the ESP32 microcontroller (Xtensa LX6 240 MHz). System tests on a held-out test set attain 98.47% accuracy, 2 ms inference latency at peak 1.4 KB RAM use; demonstrating feasibility of simple radar gate-level features towards cloud-free, real-time occupancy analytics at the edge.