Project Overview
The AI-Based Weapon Detection System is a computer vision project that uses deep learning to automatically detect guns and knives in images, video files, or live camera feeds. The system is developed using YOLOv8 (You Only Look Once) for real-time object detection and is optimized to run on CPU-only environments, making it suitable for low-resource systems.
This project is intended for academic, research, and educational purposes, such as Final Year Projects (FYP), security system simulations, and AI learning demonstrations.
Key Challenges
-
Low accuracy in knife detection due to small object size and background similarity
-
Class imbalance between gun and knife datasets
-
High false positives in complex real-world environments
-
Limited dataset diversity (lighting, angles, occlusion)
-
Running on CPU-only systems, restricting performance and speed
-
Real-time detection optimization without GPU acceleration
-
Fine-tuning confidence thresholds to balance precision vs recall
Key Achievements
-
Successfully built a real-time weapon detection system using YOLOv8
-
Achieved high accuracy for gun detection (~0.99 mAP@50)
-
Implemented live detection via webcam with bounding boxes and labels
-
Designed a custom dataset pipeline (labeling, training, validation)
-
Optimized the model to run efficiently on low-resource (CPU) systems
-
Reduced false detections using threshold tuning and dataset balancing
-
Built a modular and extensible AI pipeline for future improvements
Performance Metrics
-
Precision (Gun): High (~0.95 – 0.99)
-
Recall (Gun): High (~0.95 – 0.99)
-
mAP@50 (Gun): ~0.99
-
Precision (Knife): Moderate
-
Recall (Knife): Low
-
mAP@50 (Knife): ~0.47
-
mAP@50-95: Moderate overall (due to knife class)
-
Inference Speed: Real-time on CPU (optimized)
-
Epochs Trained: 100
-
Image Size: 640x640
-
Batch Size: 8
