NATOPS Gesture Recognition

A three-phase machine learning pipeline for classifying aircraft handling signals from the NATOPS (Naval Air Training and Operating Procedures Standardization) dataset using unsupervised clustering and supervised classification.

Overview

This project analyzes sensor data from hands, elbows, wrists, and thumbs to recognize six distinct aircraft handling gestures:

1. I have command
2. All clear
3. Not clear
4. Spread wings
5. Fold wings
6. Lock wings

The pipeline transforms raw data into cluster-based features, then trains multiple classifiers to recognize gesture patterns.

Dataset

Source: AALTD 2016 Challenge

Sensor Configuration: 8 body parts × 3 coordinates (x, y, z) = 24 features

• Hand tips (left/right)
• Elbows (left/right)
• Wrists (left/right)
• Thumbs (left/right)

Temporal Structure: 51 time steps per gesture sample

Pipeline Phases

Phase 1: Data Flattening

Purpose: Convert ARFF format time-series data into a tabular structure.

Output: flattened_data.csv:

• 24 sensor features (HandTipLeft_X, HandTipLeft_Y, etc.)
• class_label (1-6)
• time_step (1-51)
• sample_id
• isTest (boolean flag)

Usage:

python phase1.py

Phase 2: Clustering

Purpose: Transform raw sensor readings into cluster distribution features.

Output: natops_clustered_data.csv with:

• sid
• isTest
• class
• c_1, c_2, ..., c_k

Usage:

python phase2.py

Phase 3: Classification

Purpose: Train and evaluate supervised models on cluster-based features.

Models:

• Logistic Regression
• K-Nearest Neighbors (KNN)
• Multi-Layer Perceptron (MLP)

Outputs:

• Confusion Matrices
• Model Accuracy comparison table
• F1 Scores
• Overfitting analysis charts

Usage:

python phase3.py