1.Abstract

The rapid increase in robbery, theft, and violent incidents involving weapons has become a significant global concern, leading to substantial financial losses and threats to public safety. To address this issue, this project proposes an AI-powered Weapon Detection System using the advanced YOLOv7 (You Only Look Once version 7) object detection architecture. The system is designed to automatically detect and identify potential threats in real-time video streams, enabling faster response and preventive action.

The core objective of this project is to build a custom object detection model capable of identifying specific weapon classes such as pistols and knives, along with additional objects including money, credit cards, smartphones, and purses, which are often associated with theft scenarios. Unlike standard object detection models trained on generic datasets, this system utilizes a custom dataset curated with labeled images of these target classes to improve detection accuracy in real-world situations.

The project begins with data collection and preprocessing, where diverse images of the defined classes are gathered and annotated. This is followed by data visualization to understand class distribution and improve dataset quality. A pre-trained YOLOv7 model is then fine-tuned using transfer learning techniques, significantly reducing training time while maintaining high accuracy. The model training is conducted on the Google Cloud Platform (GCP), leveraging its GPU capabilities for efficient computation.

Once trained, the model performs real-time inference on video input to detect weapons and suspicious objects. Upon detection of a weapon, the system triggers an alert using the MQTT (Message Queuing Telemetry Transport) protocol, sending instant notifications to a mobile application. Additionally, a Flask-based web application is developed to stream live video and display detection results in a user-friendly interface, integrating both frontend and backend components seamlessly.

The proposed system demonstrates the effectiveness of combining deep learning, cloud computing, and IoT communication to build a scalable and intelligent surveillance solution. By enabling early detection of weapons and suspicious activities, this project contributes to enhancing security measures in public and private environments.

2. Objectives

1. To develop a real-time weapon detection system using YOLOv7.
2. To train a custom object detection model for specific classes like pistol and knife.
3. To detect suspicious objects such as money, credit cards, smartphones, and purses.
4. To implement real-time video inference for surveillance.
5. To send instant alerts using MQTT protocol.
6. To build a web-based interface using Flask for monitoring.

3.Existing System

1. Traditional surveillance systems rely on manual monitoring by humans, which is time-consuming and prone to errors.
2. Earlier approaches for weapon detection used basic image processing and classical machine learning algorithms, which required manual feature extraction.
3. Some existing systems use deep learning models like YOLOv3, YOLOv4, SSD, and Faster R-CNN for object detection, but may lack optimization for real-time performance in all scenarios.
4. Most systems are not integrated with real-time alert mechanisms, making them less effective in critical situations.
5. Limited use of custom datasets leads to poor detection accuracy for specific objects like pistols and knives.

Limitations:

1. Requires continuous human monitoring.
2. No automatic threat identification.
3. Delayed response to suspicious activities.
4. Inefficient in large-scale environments.
5. High dependency on human attention.

4.Proposed System

1. Develop an AI-powered Weapon Detection System using the YOLOv7 model for accurate and real-time object detection.
2. Train the model using a custom dataset containing images of pistols, knives, money, credit cards, smartphones, and purses.
3. Implement real-time video processing to detect weapons instantly from live camera feeds.
4. Integrate MQTT protocol to send immediate alerts to users when a weapon is detected.
5. Use a Flask web application to stream live video and display detection results.
6. Deploy the model on Google Cloud Platform (GCP) for efficient training using GPU resources.

5. Implementation Procedure

Module 1: Dataset Collection

1. Collect images of pistols, knives, and other objects.
2. Label images using annotation tools.

Module 2: Data Visualization

1. Analyze dataset distribution.
2. Verify annotations and image quality.

Module 3: Model Training

1. Use pre-trained YOLOv7 weights.
2. Train model with custom dataset.
3. Perform training on Google Cloud Platform (GPU support).

Module 4: Model Inference

1. Load trained model.
2. Perform detection on images, videos, or webcam feed.
3. Generate bounding boxes and class labels.

Module 5: Alert System (MQTT)

1. Integrate MQTT protocol.
2. Send alerts to mobile or monitoring systems when weapons are detected.

Module 6: Web Integration

1. Use Flask framework.
2. Stream live video with detection results on a web interface.

6.Software Requirements

1. Operating System                : Windows 10 / Linux
2. Programming Language       : Python
3. Libraries                                : OpenCV, PyTorch, NumPy, Pandas
4. Framework                            : Flask
5. Model                                    : YOLOv7
6. Cloud Platform                      : Google Cloud Platform (GCP)
7. Protocol : MQTT

7.Hardware Requirements

1. Processor                              : Intel i3 or higher
2. RAM                                     : 4 GB (8 GB recommended)
3. Hard Disk                              : 500 GB
4. GPU (Optional)                     : For faster training and inference