HUMAN-SUSPICIOUS-ACTIVITY-FLASK

This project leverages Computer Vision and Convolutional Neural Networks (CNN) to detect suspicious activities in real-time from CCTV surveillance footage. It aims to improve security by providing an automated approach for identifying anomalies in crowded spaces like malls, airports, and railway stations.

Features

    Real-time detection and localization of suspicious activities.
    Categorization of activities into:
1.  Suspicious Activity
2.  Criminal Activity
3.  General (Safe) Activity
    Integration with YOLO (You Only Look Once) for object detection.
    Sends automated alerts upon detecting anomalies.
    Uses CNN and LSTM for accurate temporal and spatial analysis of video frames.

System Overview

Methodology: 1. Input Video Stream: Live or recorded surveillance footage. 2. Preprocessing: Noise reduction, normalization, and frame extraction. 3. Object Detection: Detects humans and activities using a YOLOv4 model. 4. Classification: Categorizes activities using a hybrid CNN-LSTM model. 5. Alert System: Generates alerts upon detecting suspicious activities.

Tools and Technologies: Python YOLOv4 (object detection) CNN (for feature extraction) LSTM (for temporal pattern recognition) Libraries: OpenCV, TensorFlow, Keras

Project Architecture

1.  Video Input: Captures video from CCTV cameras.
2.  Preprocessing: Converts video into frames, resizes them, and normalizes.
3.  Model Training:
    Trained on datasets from Kaggle, YouTube, and public surveillance videos.
    Focused on actions like fighting, theft, or unusual gatherings.
4.  Detection & Classification:
    Uses ResNet50 and YOLO for identifying suspicious behaviors.
    Combines spatial (images) and temporal (sequence of images) analysis.
5.  Output:
    Generates visual alerts on the video feed.
    Sends notifications to designated authorities.

Dataset

Sources: Kaggle Dataset: Criminal and suspicious activity data. Custom Dataset: Created using YouTube videos and public footage.

Preprocessing: Frames resized to 64x64 pixels. Normalized pixel values for better processing.

Dataset Categories: 1. Fighting 2. Walking 3. Running

Installation

Prerequisites: Python 3.8+ Virtual environment (optional but recommended)

Steps: 1. Clone the repository:

git clone https://github.com/NaiduBM/HUMAN-SUSPICIOUS-ACTIVITY-FLASK.git

2.  Navigate to the project directory:

cd suspicious-activity-detection

3.  Install dependencies:

pip install -r requirements.txt

4.  Download pretrained YOLOv4 weights and place them in the models/ directory.

Usage

Running the Application: 1. Start the surveillance system:

python main.py

2.  Select the video feed type (Live or Recorded).
3.  Monitor the console/logs for detection alerts.

Output: Detected anomalies are highlighted in the video feed. Alerts are sent to the administrator via email/SMS.

Results Accuracy: Kaggle Dataset: 95.5% with ResNet50. Real-time Video: 99.01% with ResNet50. Performance: Real-time processing with minimal computational overhead.

Future Work Enhanced storage services for archived footage. Improved real-time processing for high-definition video. Integration with advanced alert mechanisms (e.g., mobile push notifications).

License

This project is licensed under the MIT License. See the LICENSE file for details.

Let me know if youd like further customization!