🏃 Step Counter with CNN & Peak Detection

This project provides a step counter using both traditional peak detection and a Convolutional Neural Network (CNN) to improve accuracy. It processes accelerometer data from both hands, aligns it with video-based automated step counting, and learns to predict steps precisely.

🚀 Features

✅ Automated Ground Truth Extraction:

• MediaPipe-based step detection extracts step frames directly from video using pose tracking.

✅ Data Synchronization:

• Video and accelerometer data are aligned using clap detection (detected in audio waveform).
• The number of frames from the video is matched to accelerometer sampling frequency, ensuring precise step alignment.

✅ Peak Detection (Baseline):

• Traditional ENMO-based peak detection for quick step estimation.

✅ Deep Learning-Based Step Counter:

• A CNN trained on ENMO features detects steps more reliably than peak detection.

✅ Interactive Visualization:

• Acceleration data, step peaks, and CNN predictions are plotted dynamically.

📂 Data Processing Workflow

Before training the CNN, data is preprocessed automatically and checked manuelly:

Step 1: Extract Video-Based Steps (Ground Truth)

• Each video is analyzed using MediaPipe to detect ankle and foot index motion.
• Steps are extracted, stored in step_counts.csv, then scaled to match the accelerometer sampling rate (scaled_step_counts.csv).

Step 2: Synchronize Video & Accelerometer Data

• The start and end points of accelerometer data are synchronized with the video using clap detection (clap_detection_methods.py).
• Acceleration data is sliced to match the segment of the video using save_acc_metadata_sliced.py.

Step 3: Compute ENMO & Normalize

• ENMO (Euclidean Norm Minus One) is computed from X, Y, Z acceleration.
• Values are normalized for stable training.
• 

Step 4: Peak Picker Baseline

• A simple peak detection is applied peak_picker.py for quick validation before CNN training.

🔍 Peak Detection (Baseline Approach)

The first step in detecting walking patterns is using a simple peak-picking algorithm. This method detects peaks in the ENMO signal, identifying potential step events.

• is more effizent
• only checks peaks

🤖 CNN-Based Step Detection

A 1D CNN is trained on ENMO data to detect steps more accurately.

Model Architecture

• Conv1D Layers extract temporal step features.
• Pooling Layers reduce noise.
• Fully Connected Layers output a probability for each frame.
• ✅ The CNN learns step patterns and improves detections by reducing false positives and misalignments.

📊 Evaluation & Visualization

• The model is tested on unseen data.
• Steps detected by the CNN (🔴 red dots) are compared to ground truth steps (crosses).

🛠 How to Use

1 Set Up Environment

To ensure all dependencies are installed, choose one of the two methods: Option 1: Install with requirements.txt

• pip install -r requirements.txt **Option 2: Install via environment.yml (for Conda users)
• conda env create -f environment.yml
• conda activate stepcounter

2 Preprocess Data (Extract & Sync Ground Truth)

python save_acc_metadata_sliced.py -> Extracts step frames from video & syncs with accelerometer data.

3 Run Peak Detection Baseline

python peak_picker.py -> Detects steps using a simple peak detection algorithm.

4 Train CNN Model

python training.py -> Trains the CNN on synchronized ENMO features.

🎯 Next Steps

• ✅ Refine CNN training to better align detected steps exactly on peaks.
• 🔍 Optimize CNN loss function for better step probability calibration.
• 📊 Experiment with larger datasets (different walking speeds).
• 🛠 Improve peak detection pre-filtering before CNN classification.