TinyML Fall Detection Wearable for Elderly Safety

Falls are the leading cause of injury for the elderly living alone, and the critical window is the minutes after the fall — when the person may be unable to reach a phone. Camera-based monitoring invades privacy and cloud-connected pendants fail when Wi-Fi drops. This project puts the intelligence on the body: a waist-worn ESP32 device reads its IMU at 100 Hz and classifies every 2-second window with a TinyML model small enough for the microcontroller's RAM. A real fall triggers a local buzzer (so a conscious wearer can cancel a false alarm within 15 seconds) and then sends a BLE alert with the fall timestamp to the caregiver's phone. Nothing is recorded or streamed; the device only keeps an event log. Because on-device inference replaces the cloud round-trip, detection-to-alert latency stays under a second even with no network at all.

1. The MPU6050 streams 3-axis acceleration and gyroscope data to the ESP32 at 100 Hz while worn at the waist.
2. Firmware maintains a sliding 2-second window and computes the signal magnitude vector plus statistical and spectral features.
3. A rule stage checks the fall signature: weightlessness dip below 0.5 g, impact spike above 2.5 g, then abnormal stillness.
4. Windows passing the signature go to the quantized 1D-CNN, which outputs the fall-vs-ADL probability.
5. On a confirmed fall, the buzzer sounds for the 15-second cancel window; if not cancelled, the BLE alert fires to the caregiver phone.
6. Every classified event is timestamped into the on-device log, viewable on the OLED.
7. The buyer-run drop-test procedure (cushioned dummy falls plus scripted ADLs) measures the achieved sensitivity/false-alarm rate for the report.

Tech stack:

• ESP32 (dual-core Xtensa, 520 KB SRAM)
• MPU6050 6-axis IMU, 100 Hz sampling
• TensorFlow Lite Micro, INT8 quantized 1D-CNN
• SisFall dataset (Sucerquia et al., 2017)
• BLE caregiver alert + companion phone app
• 0.96" SSD1306 OLED, buzzer, 3.7 V LiPo
• Python training pipeline (NumPy, TensorFlow)

Dataset & model details

• Dataset: SisFall (Fall and Movement Dataset) — 4,505 labeled recordings (2,707 ADL / 1,798 falls), 19 ADL types + 15 fall types, 38 subjects (23 young adults 19–30, 15 elderly 60–75), waist-worn device with two accelerometers + gyroscope sampled at 200 Hz (downsampled to 100 Hz in the pipeline), released by Sucerquia, López and Vargas-Bonilla (Sensors, 2017), available via PhysioNet.
• Task: 2-second motion-window binary classification (fall vs ADL); input = 6-channel IMU time series (200 samples × 6), output = fall probability per window.
• Model: 1D-CNN (~24 KB after INT8 quantization), TensorFlow Lite Micro on ESP32; rule-based fall-signature prefilter (free-fall/impact/stillness) before the network.
• Metrics: Sensitivity (fall recall) design target 94%, specificity design target 96%, false alarms per day design target below 1 — reported as design targets in the build notes, with the actual measured figures recorded during the buyer-run drop-test procedure. No accuracy is claimed as measured until that run. Not a medical device; no clinical claims are made.

• [On-device TinyML inference] A ~24 KB quantized 1D-CNN classifies 2-second IMU windows on the ESP32 itself — no phone, no cloud, works with zero network.
• [6-axis IMU front-end] MPU6050 accelerometer + gyroscope sampled at 100 Hz with signal-magnitude-vector features computed in firmware.
• [Fall signature pipeline] Free-fall dip, impact spike and post-fall stillness checks precede the CNN, cutting false alarms from sitting or bending.
• [Cancel window + buzzer] A 15-second audible cancel window lets a conscious wearer dismiss false alarms before the caregiver alert fires.
• [BLE caregiver alert] Fall event with timestamp is pushed over BLE to the caregiver's phone app (included).
• [Event log on OLED] The 0.96-inch OLED shows live status plus the recent fall/ADL event history.
• [Training pipeline included] Python scripts for SisFall preprocessing, model training and TFLite Micro conversion ship with the project.

• Assembled wearable: ESP32 + IMU + buzzer + OLED + battery in 3D-printed waist case
• Complete firmware (IMU pipeline, signature checks, inference, BLE alerts, power management)
• Quantized TFLite Micro model + Python training/conversion pipeline
• Caregiver phone app (BLE alert receiver) source
• Wiring diagram and case assembly guide
• Project report PDF (TinyML background, SisFall methodology, drop-test procedure, results, medical disclaimer)
• PPT presentation for final review
• Viva Q&A preparation document (1D-CNN, quantization, SMV, sensitivity/specificity)

• Sensitivity 94% is a design target until the buyer-run drop test — real falls vary in mechanics and the report documents this honestly.
• This is a safety/educational prototype, NOT a medical device — it does not replace emergency response or medical supervision, stated plainly in the report.
• Slow, controlled collapses (e.g. sliding down a wall) produce weak impact signatures and are the known hard case.
• Must be worn at the waist as documented; wrist or pocket placement changes the signal and is out of scope.
• ~10-hour battery is a design target; BLE alert range is room-scale (expected ~10 m).
• The device detects falls; it cannot prevent them or assess injury.