Felix Corputty
Telecommunication Engineering Study Program, School of Electrical Engineering, Telkom University

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

Implementation of On-Device Convolutional Neural Network Inference for Indoor Occupancy Detection on the ESP32-S3 Ade Aditya Ramadha; Felix Corputty; Naufal Hanan Lutfianto
Jurnal Ecotipe (Electronic, Control, Telecommunication, Information, and Power Engineering) Vol 13 No 2 (2026): Jurnal Ecotipe, October 2026
Publisher : Jurusan Teknik Elektro, Universitas Bangka Belitung

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33019/jurnalecotipe.v13i2.4616

Abstract

Indoor occupancy detection is a key enabling technology for energy-efficient smart building management, as heating, ventilation, air conditioning, and lighting systems waste significant energy when conditioned spaces are unoccupied. This paper implements a compact Convolutional Neural Network (CNN) trained on the UCI Occupancy Detection Dataset and deployed entirely on the ESP32-S3 microcontroller using TensorFlow Lite Micro with INT8 post-training quantization, eliminating any cloud dependency. Models are trained under a leakage-controlled protocol on the designated training file, selected on the first test file, and evaluated on the held-out second test file; the INT8 quantizer is calibrated from an evenly spaced spread of training windows to avoid class bias. Twelve configurations spanning four filter capacities, (F₁,F₂) ∈ {(4,8),(8,16),(16,32),(32,64)}, and three temporal window lengths, W ∈ {15,30,60} minutes, are quantized, flashed, and measured on the device for tensor arena, flash footprint, inference latency, and held-out F1. All twelve configurations execute successfully: the best compact model, CNN(8,16) at  fifteen-minute window, reaches a held-out F1 of 0.959 using 4.2 KB of tensor arena and 1.24 ms per inference, with more than 320 KB of heap remaining free; the tensor arena across the full grid stays between 2.6 and 30 KB; and INT8 quantization changes F1 by at most ±0.021 relative to the floating-point baseline. These results demonstrate that on-device CNN occupancy detection is practical on  low-cost microcontroller, that a compact two-block architecture is sufficient to reach near-optimal accuracy, and that the full pipeline from leak-free training through representative INT8 calibration to on-device measurement is reproducible.