CCTV Footage Summarization System

A modular and scalable system for automatic CCTV video summarization using motion-based frame selection, object detection, and video summarization. The system eliminates redundant footage and efficiently highlights important events.

Problem Statement

Long-duration CCTV videos are difficult to analyze manually due to:

• High redundancy

• Large storage requirements

• Limited actionable events

This project addresses these challenges by:

• Selecting only motion-relevant frames

• Applying object detection selectively

• Generating concise summary videos

System Overview

The pipeline consists of three independent modules, orchestrated by a master controller:

• Frame Selection – Motion-based keyframe extraction

• Object Detection – YOLO-based detection on selected frames

• Summarization – Frame-to-video summary generation

Each module is reusable, independent, and scalable.

Project Architecture

Folder Structure

project/
│
├── Videos/
│   └── {video-name}.mp4
│
├── results/
│   └── {video-name}/
│       ├── selected_frames/
│       ├── detected_frames/
│       ├── summaries/
│       └── logs/
│
├── frame_preprocessing.py
├── frame_selection.py
├── object_detection.py
├── summarizer.py
├── main.py
└── README.md

Processing Pipeline

• Input CCTV video
• Motion-based frame differencing
• Keyframe extraction
• Optional object detection on keyframes
• Summary video generation (raw & detected)

Technologies Used

• Python
• OpenCV
• NumPy
• Ultralytics YOLOv8
• Matplotlib (for analysis/graphs)

How to run

1. Install dependencies

  pip install opencv-python ultralytics numpy matplotlib

2. Place Input Video

  Videos/{video-name}.mp4

3. Run main.py

  python main.py

All outputs will be stored automatically inside:

  results/{video-name}/

Output Structure (Per Video)

results/{video-name}/
├── selected_frames/     # Motion-based keyframes
├── detected_frames/     # YOLO annotated frames
├── summaries/
│   ├── summary_raw.mp4
│   └── summary_detected.mp4
└── logs/

Results & Visualizations

Motion Analysis Graph

Graph showing percentage of changed pixels over time: (observation for one video)

Object Detection Results

Annotated frames with bounding boxes:

Comparision

Comparison prior to and following application of the system: (observation for one video)

Aspect	Before application	After application
Frames processed	64231	12411
Video Duration	35:42	17:14
No. of keyframes selected	N/A	12411
Percentage of frames retained	N/A	19.32%
Percentage of frames reduction	N/A	80.68%

Experimental Settings

• Frame differencing threshold: 15
• Motion percentage threshold: 0.15%
• Summary FPS: 12
• YOLO model: YOLOv8-nano
• Detected classes: person, car

Key Advantages

• Modular and reusable design
• Efficient processing of long CCTV footage
• Reduced storage and computation
• Suitable for real-time and offline analysis

Conclusion

This project demonstrates an efficient and scalable approach to CCTV video summarization by combining classical computer vision techniques with deep learning–based object detection. The modular design ensures flexibility, maintainability, and ease of extension.

Author

• Sonjyoti Rabha
• Project Type: Academic / Final Year Project
• Domain: Computer Vision, Video Analytics

References

• A. B. Deshmukh, V. N. Aher, V. More, R. S. Pol, A. Talele, and A. Ingle, "Memory Efficient Summarization of Real-Time CCTV Surveillance System"

• Muhammad, K., Hussain, T., & Baik, S. W. "Efficient CNN based summarization of surveillance videos for resource-constrained devices". Intelligent Media Laboratory, Digital Contents Research Institute, Sejong University, Seoul-143-747, Republic of Korea, August 2018.

• L. Goel, "CCTV Footage Summarization for Increasing Efficiency in Surveillance (using ML)", International Journal of Innovative Science and Research Technology

• Ultralytics, ”YOLOv8: A State-of-the-Art Object Detection Model,” 2023. [Online]. Available: https://docs.ultralytics.com/