---
license: mit
task_categories:
- feature-extraction
language:
- en
tags:
- sensor
- physics
---


# Hand Detection Training Data

This folder contains sensor data collected from mobile devices for training the hand detection model.

## Overview

The dataset includes accelerometer and gyroscope readings from 2 subjects, each holding a device with both their left and right hands. This data is used to train the Random Forest classifier that achieves 94.6% accuracy in detecting which hand is holding the device.

## Directory Structure

```
hand_data/
├── accelerometer/          # Accelerometer sensor data (primary)
│   ├── s-1_left_hand.csv   # Subject 1, left hand (39,102 samples)
│   ├── s-1_right_hand.csv  # Subject 1, right hand (30,528 samples)
│   ├── s-2_left_hand.csv   # Subject 2, left hand (44,724 samples)
│   └── s-2_right_hand.csv  # Subject 2, right hand (35,408 samples)
│
└── gyrocop/                # Gyroscope data (supplementary)
    ├── s-1_left_hand.csv   # Subject 1, left hand
    └── s-1_right_hand.csv  # Subject 1, right hand
```

## Data Format

### Accelerometer Data

Each CSV file contains timestamped 3-axis accelerometer readings:

| Column    | Type      | Description                              |
|-----------|-----------|------------------------------------------|
| timestamp | datetime  | ISO 8601 format (e.g., 2025-12-27T09:13:07.598506) |
| x         | float     | X-axis acceleration (m/s²)               |
| y         | float     | Y-axis acceleration (m/s²)               |
| z         | float     | Z-axis acceleration (m/s²)               |

**Example:**
```csv
timestamp,x,y,z
2025-12-27T09:13:07.598506,0.849452,3.895515,8.087741
2025-12-27T09:13:08.083118,0.727418,4.000800,8.099705
```

### Gyroscope Data

Similar structure with angular velocity measurements (°/s).

## Dataset Statistics

### Total Samples
- **Subject 1 (Left)**: 39,102 samples
- **Subject 1 (Right)**: 30,528 samples
- **Subject 2 (Left)**: 44,724 samples
- **Subject 2 (Right)**: 35,408 samples
- **Total**: 149,762 samples

### Collection Method
- Device: Mobile phone with accelerometer sensor
- Sampling rate: ~50-100 Hz (varies)
- Duration: Multiple sessions per subject/hand
- Environment: Normal daily usage patterns

## Data Characteristics

### X-Axis (Left/Right Tilt)
- **Primary discriminator** for hand detection
- Left hand: Positive values (device tilts right)
- Right hand: Negative values (device tilts left)
- Statistical significance: p < 0.000001

### Y-Axis (Forward/Backward Tilt)
- Secondary feature
- Shows hand-specific patterns
- Less discriminative than X-axis

### Z-Axis (Vertical)
- Represents gravity component
- Generally around 9.8 m/s² when stationary
- Varies with device orientation

### Magnitude
- Calculated: √(x² + y² + z²)
- Overall movement intensity
- Helps distinguish activity levels

## Usage in Training

This data is used in [../which_hand_you_use.ipynb](https://github.com/rockerritesh/sensor/blob/main/hand/which_hand_you_use.ipynb) for:

1. **Exploratory Data Analysis (EDA)**
   - Distribution analysis
   - Statistical testing
   - Correlation analysis
   - Time series visualization

2. **Feature Engineering**
   - Calculate magnitude
   - Window-based statistics (mean, std, min, max)
   - Temporal features (deltas, trends)

3. **Model Training**
   - Single-point Random Forest (94.6% accuracy)
   - Windowed Random Forest (96%+ accuracy)
   - PCA for visualization

## File Sizes

- `s-1_left_hand.csv`: ~2.1 MB
- `s-1_right_hand.csv`: ~1.7 MB
- `s-2_left_hand.csv`: ~2.4 MB
- `s-2_right_hand.csv`: ~2.0 MB

**Total**: ~8.2 MB (accelerometer only)

## Data Quality

### Completeness
✅ No missing values
✅ Continuous timestamps
✅ Consistent format across all files

### Statistical Validation
✅ Normal distribution per axis
✅ Significant hand differences (p < 0.05)
✅ Consistent patterns across subjects

## Privacy & Ethics

- Data collected with informed consent
- No personally identifiable information
- Used solely for research purposes
- Anonymized subject identifiers (S1, S2)

## Collection Guidelines

If collecting additional data:

1. **Consistency**: Use same device/settings
2. **Duration**: Minimum 5-10 minutes per hand
3. **Activity**: Natural usage (browsing, typing, etc.)
4. **Labeling**: Clear hand identification
5. **Format**: Match existing CSV structure

## Notes

- This data is **excluded from git** (see `.gitignore`)
- Keep data locally or use Git LFS for large files
- Model files are generated from this data
- Data collection scripts in `shared/` folder

## Related Files

- **Training**: [../which_hand_you_use.ipynb](https://github.com/rockerritesh/sensor/blob/main/hand/which_hand_you_use.ipynb)
- **Models**: `hand_classifier_*.pkl` files
- **Collection**: `collect_data.py` in shared folder

---

**Last Updated**: December 2025
**Format Version**: 1.0
**Total Samples**: 149,762