Challenge 2025 — ADAMMA

On-Device MET Class Prediction from Smartphone Accelerometers

This repository contains the Android app, training code, exported models, demonstration video, and paper for our submission to Challenge 2025 (ADAMMA). We estimate MET classes (Sedentary · Light · Moderate · Vigorous) from smartphone accelerometer data (WISDM dataset) and deploy a compact ONNX model on Android.

---

📂 Repository Layout

• app/
  Kotlin Android app with a foreground service (SensorManager), feature computation, on-device inference, and live UI.

• gradle/wrapper/
  Gradle wrapper scripts and configuration files for reproducible Android builds.

• MET_Tracker_Trainer/
  Training & evaluation:

  • main.py (entry point)
  • datasets/ (windowing, preprocessing)
  • models/ (classic ML + 1D CNN)

• video_onnx_and_apk_report/
  Packaged artifacts:

  • Android APK
  • Video demonstration of the app
  • ONNX-exported CNN model cnn.onnx
  • PDF report

---

⚙️ Environment

• Python ≥ 3.9
• PyTorch, scikit-learn, NumPy, pandas
• (Optional) CUDA for faster CNN training
• Android Studio (AGP 8.5.2, Gradle 8.7), compileSdk=34, minSdk=24

Install Python deps (from inside the trainer folder):

cd MET_Tracker_Trainer
pip install -r requirements.txt

🗂️ Dataset

We use the WISDM Activity Prediction dataset (smartphone accelerometer at ~20 Hz, 6 activities).

• Official page (Fordham WISDM Lab): https://www.cis.fordham.edu/wisdm/dataset.php
• Citation: Kwapisz et al., Activity Recognition using Cell Phone Accelerometers, SensorKDD 2010.

Note: This is the classic 6-activity WISDM dataset used in many HAR papers. There is also a newer WISDM Smartphone & Smartwatch dataset (18 activities, phone+watch, 20 Hz) on the UCI ML Repository; our work here targets the classic 6-class phone dataset to match the MET mapping.

🚀 Training

Handcrafted features + classic ML

cd MET_Tracker_Trainer
python main.py   --dataset WISDM   --dataset_path /path/to/WISDM/   --feature_extraction handcrafted   --model ml   --n_splits 5

Raw windows + compact 1D CNN

cd MET_Tracker_Trainer
python main.py   --dataset WISDM   --dataset_path /path/to/WISDM/   --feature_extraction raw   --model cnn   --n_splits 5   --batch_size 32

---

📱 Android App

• Continuous sensing at ~20 Hz; on-device inference (ONNX Runtime mobile).
• Foreground service keeps predictions running; UI shows current class.

Install the APK:

1. Enable Install unknown apps on your device.
2. Transfer and install:

   video_onnx_and_apk_report/<your-apk>.apk
   

3. Launch the app and press Start to begin sensing.

---

📦 Reproducibility & Artifacts

All artifacts are bundled for convenience:

• APK (ready to install)
• Demo video of the app running
• cnn.onnx exported model. A copy is also in app/src/main/assets/. The app must have this file in that directory to build successfully.
• Report (PDF) — the paper corresponding to this repo

All four files live in:

video_onnx_and_apk_report/

---

📊 Results (summary)

• Models evaluated: Logistic Regression, kNN, RBF-SVM, Random Forest, Gaussian NB, MLP, and a compact 1D CNN.
• The 1D CNN on raw windows offers the best accuracy/latency trade-off for on-device deployment.
• Classic ML baselines work well on the 43-D handcrafted features.

(See the PDF report for full tables and methodology.)