VisionReasoningSecuritySystem

Category	Tool / Tech	Purpose / Description
🧠 ML Model	YOLOv8 (Ultralytics)	Object detection model fine-tuned for weapon detection
	PyTorch .pt	Model format used during inference
🏋️ Training	Transfer Learning	Trained with custom dataset (Roboflow)
📁 Dataset Format	YOLO (labels .txt, data.yaml)	Used for training, validation, and inference
🧪 Inference Engine	ultralytics Python package	Handles model loading and inference pipeline
🎥 Video Input	OpenCV (cv2.VideoCapture)	Captures webcam or mobile stream
	Custom cam selector (list_available_cameras())	Lists available webcams and IP streams
📱 IP Streaming	IP Webcam	Sends mobile camera feed over local network
🌐 UI Framework	Streamlit	Real-time interactive web app with live video + controls
📦 Deployment	Docker	Containerized deployment of the app
	Docker Compose	Orchestrates multiple services (e.g., app + model + others)
	--device /dev/video0	Grants container access to physical webcam
🤖 Notifications	Telegram Bot (telebot / python-telegram-bot)	Sends real-time alerts and status updates
	Bot token + Chat ID	For secure messaging
📊 Logging	runs/detect/train*/, results.csv, labels.jpg	YOLO training results, validation performance, and loss curves
🧠 Monitoring	TensorBoard (optional)	Training insights via logs (if enabled properly)
⏱️ Alert Control	Python time.time()	Prevents alert flooding using cooldown timer
🔗 LLM Reasoning	LangChain	Manages LLM pipeline and prompt engineering
🤖 Prompting Agent	Open-source LLM (e.g. Mistral via Ollama)	Generates human-readable descriptions of detected scenes

---

📊 Model Performance Summary

Class	Images	Instances	Precision	Recall	mAP@0.5	mAP@0.5:0.95
all	327	1042	0.645	0.496	0.528	0.293
Armed-Person	260	376	0.852	0.522	0.720	0.436
Rifle	274	405	0.855	0.306	0.484	0.220
Person	143	257	0.575	0.405	0.468	0.282
Knife	4	4	0.299	0.750	0.439	0.234

⚡ Inference Speed

• Preprocess: 0.2ms
• Inference: 2.0ms
• Loss: 0.0ms
• Postprocess: 1.7ms

---

🔗 Transfer Learned Model

Download Model: MEGA