🧠 CMI 2025 - Human Motion Classification from IMU Sensors

📘 Overview

This repository contains the full workflow developed for the CMI 2025 Human Motion Classification competition. The goal of this challenge is to accurately identify human motion types (e.g., walking, sitting, running, standing) using data collected from IMU (Inertial Measurement Unit) sensors such as accelerometers, gyroscopes, and orientation sensors.

This project demonstrates end-to-end data processing, feature extraction, model training, and validation pipelines with a strong emphasis on robustness, interpretability, and efficiency.

---

📊 Dataset

The dataset consists of multivariate time-series signals captured from IMU sensors. Each sequence corresponds to one labeled human activity.

Example sensor streams:

• acc_x, acc_y, acc_z → Accelerometer axes (m/s²)
• gyro_x, gyro_y, gyro_z → Gyroscope angular velocity (rad/s)
• rot_x, rot_y, rot_z, rot_w → Orientation quaternion
• time → Timestamp or frame index

Target variable: behavior — the motion category (18 classes)

The data is typically split into train.csv, test.csv, and a metadata file such as subjects.csv or sequence_info.csv.

---

⚙️ Approach

1️⃣ Data Preprocessing

• Sensor alignment and missing-value imputation
• Outlier clipping and normalization (per sensor group)
• Sequence-level standardization and window segmentation

2️⃣ Feature Engineering

• Statistical features (mean, std, skew, kurtosis, IQR)
• Frequency-domain analysis via FFT and wavelets
• Quaternion-based angular distance and velocity features
• Inter-sensor fusion (e.g., relative acceleration, gravity-compensated motion)
• Aggregation by sequence_id or subject_id

3️⃣ Modeling

A combination of traditional and deep learning models was explored:

• Baseline: RandomForest, XGBoost, LightGBM
• Deep Models: LSTM / GRU, 1D CNN, and InceptionTime-style architectures
• Hybrid: Temporal Convolutional Networks (TCN) + Attention pooling

Training employed stratified group k-fold cross-validation to prevent data leakage between subjects.

4️⃣ Evaluation

• Primary metric: Macro F1-score
• Validation: 5-fold cross-validation
• Visualization: confusion matrices, per-class accuracy, t-SNE embeddings

---

📈 Results Summary

Model	Cross-Validation F1	Public LB	Notes
XGBoost baseline	0.82	0.80	fast, interpretable
TCN + Attention	0.91	0.90	best generalization
Ensemble (XGB + TCN)	0.92	0.91	final submission

---

🧩 Tech Stack

• Python 3.10+
• NumPy, Pandas, Polars
• Scikit-learn, XGBoost, LightGBM
• PyTorch, PyTorch Lightning
• Matplotlib, Seaborn, Plotly

---

🚀 How to Run

Clone the repository and open the notebook:

git clone https://github.com/massoudsh/FinDetect.git
cd FinDetect/CMI_2025_IMU_Motion_Classification
jupyter notebook Kaggle_CMI_2025.ipynb

Or run directly from terminal (if using VSCode / JupyterLab):

jupyter lab

Dependencies can be installed with:

pip install -r requirements.txt

---

🧪 Next Steps

• Add real-time inference pipeline (IMU sensor streaming)
• Deploy trained model as a lightweight edge classifier (ONNX / CoreML)
• Integrate interpretability (SHAP for temporal data)
• Compare transformer-based temporal encoders (e.g., TimesNet, TST)

---

👤 Author

Massoud Shemirani
FinTech | AI Developer | ML Researcher
📍 GitHub • Kaggle • LinkedIn

---

© 2025 — Developed as part of the Kaggle CMI 2025 competition.