EPT Surveillance Video - Events Detection

A real-time video surveillance system that detects intrusions in user-defined protected zones using the YOLO (You Only Look Once) object detection model via Darkflow. Built with Python, OpenCV, and TensorFlow.

---

Overview

This project was developed at EPT (Ecole Polytechnique de Tunisie) as part of an academic project on event detection in video streams. The system captures a live webcam feed, allows the user to define a "Protected Zone" on screen, and triggers alerts whenever a person is detected inside that zone during a configurable time window.

Key Features

• Real-time person detection using the Tiny-YOLO-VOC pre-trained model
• Interactive protected zone selection - draw a rectangle on the video feed to define the monitored area
• Time-based monitoring - configure start/end times for zone protection via parameters.txt
• Intrusion logging - all detections are recorded with timestamps in logfile.log
• Video recording - detected events are saved to an output video file
• Configurable detection labels - specify which objects to detect via labels_det.txt

---

Demo

1. Launching the Application

Run the detection script with the YOLO model and weights:

python projet.py --model ../cfg/tiny-yolo-voc.cfg --load ../bin/tiny-yolo-voc.weights -sz 750

2. Loading the YOLO Model

The system loads the Tiny-YOLO-VOC model and builds the neural network architecture:

3. Live Video Stream

Once the model is loaded, the webcam feed starts with a real-time timestamp overlay:

4. Defining the Protected Zone

Press i to enter zone selection mode, then draw a rectangle with the mouse to define the area to monitor. Press Enter or Space to confirm, or c to select the whole frame:

5. Intrusion Detection & Alert

When a person enters the protected zone during the configured time window, the system displays an "Alert !!!" warning and logs the event:

6. Logging & Configuration Files

parameters.txt - Defines the time window for zone monitoring:

logfile.log - Records every intrusion event with date and time:

labels_det.txt - Lists the object labels to detect (e.g., "person"):

7. Final Summary

After pressing q to quit, the system displays elapsed time, FPS, and saves the recorded video:

---

Project Structure

Projet events detection/
├── README.md                  # This file
├── src/                       # Source code
│   ├── projet.py              # Main detection script
│   ├── guide.md               # Usage guide
│   ├── labels_det.txt         # Labels to detect (e.g., "person")
│   ├── parameters.txt         # Time window configuration
│   └── logfile.log            # Detection log output
├── docs/                      # Documentation
│   ├── Sujet Projet.docx      # Project brief / subject
│   ├── Les étapes du projet.docx  # Project milestones
│   ├── plan rapport.docx      # Report outline
│   ├── Video Stream Analytics Using OpenCV.docx  # OpenCV analytics guide
│   └── references/            # Research papers & bibliography
│       ├── yolo.pdf
│       ├── 2010CLF22089_-_LUVISON.pdf
│       ├── Human-detection-from-images-and-videos-*.pdf
│       ├── Combining-motion-and-appearance-cues-*.pdf
│       └── ... (additional pattern recognition papers)
├── images/                    # Demo screenshots
│   ├── 01_run_command.png
│   ├── 02_loading_model.png
│   ├── ...
│   └── 13_final_output_summary.png
└── output/                    # Recorded output videos
    └── video.avi

---

Dependencies

• Python 3
• TensorFlow 1.x
• OpenCV 3.x
• NumPy
• imutils
• Darkflow - github.com/thtrieu/darkflow

Installation

1. Clone or download this repository.

2. Install Darkflow by following the instructions at thtrieu/darkflow.

3. Download the Tiny-YOLO-VOC weights and config files and place them in the appropriate directories (bin/ and cfg/).

4. Install Python dependencies:

   pip install tensorflow==1.x numpy opencv-python imutils

Usage

1. Configure detection labels - Edit src/labels_det.txt to list the objects you want to detect (one per line). Default is person.

2. Configure time window - Edit src/parameters.txt with the start and end datetime for zone protection, using the format YYYY-MM-DD HH:MM:SS.

3. Run the program:

   cd src/
   python projet.py --model <path_to_model.cfg> --load <path_to_weights>

4. Interactive controls:

   • Press i to draw the protected zone
   • Press Enter/Space to confirm the zone selection
   • Press c to select the entire frame as the zone
   • Press q to quit

Command-Line Arguments

Argument	Short	Default	Description
--model	-m	required	Path to YOLO model config file
--load	-l	required	Path to YOLO weights file
--confidence	-conf	0.2	Minimum detection confidence
--person	-p	0.3	Person detection threshold
--seize	-sz	650	Window display size (px)
--fps	-f	5	Output video FPS
--save	-s	video.avi	Output video file path
--codec	-cod	XVID	Output video codec

---

How It Works

The system uses a Tiny-YOLO-VOC model for real-time object detection on each video frame. When the user defines a protected zone (via an interactive ROI selector), the program computes the overlap ratio (Intersection over Union) between each detected bounding box and the protected zone. If the overlap exceeds a threshold (5%) and the current time falls within the configured monitoring window, an intrusion alert is triggered, displayed on screen, and logged to logfile.log.

---

Author

Marwen Kraiem - EPT (Ecole Polytechnique de Tunisie), 2018

---

References

Research papers used in this project are available in the docs/references/ folder, covering topics such as human detection, anomaly detection, pattern recognition, and the YOLO architecture.