🎯 Video Stability Analysis using Optical Flow and 3D CNN

Extracting Motion Patterns from Video Sequences

---

📌 Overview

This project analyzes motion patterns in video data and classifies stability using a combination of classical computer vision and deep learning techniques.

The system computes motion-based stability scores using optical flow, generates labels automatically, and trains a 3D Convolutional Neural Network (3D CNN) to classify videos into stable and unstable categories.

---

❓ Problem Statement

Raw video data is unstructured and does not directly provide information about motion stability.

The objective of this project is to:

• Extract meaningful motion information from video sequences
• Quantify motion stability using feature engineering
• Classify videos into stable and unstable categories

---

🚀 Key Features

• 🎥 Video preprocessing and frame extraction
• 🌊 Optical Flow-based motion analysis
• 📊 Stability score computation
• 🏷️ Automatic dataset labeling (heuristic-based)
• 🗂️ Dataset organization (stable vs unstable)
• 🧠 3D CNN model for classification
• 📈 Model training and evaluation

---

🧠 Project Pipeline

Raw Video Data
        ↓
Optical Flow Analysis (main.py)
        ↓
Motion Features + Stability Score
        ↓
Auto Label Generation (sort_dataset.py)
        ↓
Structured Dataset
   ├── stable/
   └── unstable/
        ↓
3D CNN Training (train.py)
        ↓
Classification Accuracy

---

📁 Project Structure

video-stability-analysis-3dcnn/
│
├── src/
│   ├── main.py              # Optical flow + feature extraction
│   ├── sort_dataset.py     # Label generation + sorting
│   ├── train.py            # Model training
│   ├── data_loader.py      # Dataset loading
│   ├── model_3dcnn.py      # 3D CNN model
│   ├── preprocessing.py
│   ├── optical_flow.py
│   ├── feature_extraction.py
│   ├── stability_index.py
│   ├── visualization.py
│
├── dataset/
│   ├── stable/
│   └── unstable/
│
├── Data/
│   └── labels.xlsx
│
├── Results/
│   └── final_results.xlsx
│
├── requirements.txt
└── README.md

---

⚙️ Installation

git clone https://github.com/CodeVoyager7777/video-stability-analysis-3dcnn.git
cd video-stability-analysis-3dcnn
python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

---

▶️ Usage

1️⃣ Run Optical Flow + Feature Extraction

python src/main.py

2️⃣ Generate Labels & Organize Dataset

python src/sort_dataset.py

3️⃣ Train the Model

python src/train.py

---

🧪 Model Details

• Model: 3D Convolutional Neural Network (3D CNN)

• Input: Video clips

• Frame size: 64 × 64

• Frames per video: 8

• Classes:

  • Stable Motion
  • Unstable Motion

---

🧠 Core Idea

Instead of relying on manual labeling, this project:

Automatically generates labels using motion-based stability scores derived from optical flow.

This enables scalable dataset creation for supervised learning.

---

📊 Results

• Achieved ~79% accuracy on test dataset
• Model learns motion-based stability patterns from video sequences

---

⚠️ Limitations

• Heuristic Labeling: Labels are generated using a mean-based threshold instead of ground-truth annotations.
• Threshold Sensitivity: The threshold may not generalize well across different datasets.
• Limited Evaluation Metrics: Performance is evaluated mainly using accuracy.
• No Model Comparison: The model is not benchmarked against alternative approaches.

---

🚀 Future Scope

• Improve labeling using annotated or semi-supervised data
• Compare with alternative models (2D CNN, LSTM, classical ML)
• Add advanced evaluation metrics (precision, recall, F1-score)
• Improve model generalization and robustness

---

🛠️ Tech Stack

• Python
• OpenCV
• NumPy & Pandas
• PyTorch
• Matplotlib

---

👨‍💻 Author

Laksh Makkar GitHub: https://github.com/CodeVoyager7777