YOLOv8n based Threat Detection Model

CNNs for Object Detection

YOLOv8, developed by Ultralytics, continues the legacy of the highly popular YOLO (You Only Look Once) series. This version brings significant improvements in both speed and accuracy, making it a top choice for real-time object detection tasks. Its efficient CNN-based architecture is optimized for performance on both CPUs and GPUs.

This repository features a fine-tuned YOLOv8 Nano model specifically trained for Threat Detection, designed to identify four critical threat categories with high precision and speed.

Model Overview

YOLOv8n Threat Detection is a specialized computer vision model for security and surveillance. Leveraging the speed and efficiency of the YOLOv8 Nano architecture, this model accurately detects potential threats in real-time scenarios.

The threat categories are:

Training Dataset

The model was trained on a custom threat detection dataset, meticulously curated and annotated for robust performance across various scenarios.

Class Distribution

Sample Annotations

Performance Metrics

Training performance

Confusion matrix

Validation Results

Test Results

Key Performance Highlights

• High Overall Accuracy: Achieved a strong 81.3% mAP@50 on the validation set, showing the model is highly effective.
• Exceptional 'Grenade' Detection: The model excels at identifying grenades, with an outstanding 92.1% mAP@50 and an extremely high 96.5% precision. This indicates a very low rate of false positives for this class.
• Strong Generalization: Reached a peak mAP@50-95 of 55.3%, demonstrating a good ability to predict bounding boxes with high precision (IoU > 0.95).
• Balanced Learning: The steady decrease in box_loss, cls_loss, and dfl_loss over 50 epochs indicates stable and balanced learning across localization, classification, and distribution focal loss tasks.

Model Architecture

• Base Architecture: YOLOv8 Nano (yolov8n.pt)
• Parameters: ~3 Million (3,006,428 fused)
• Computational Cost: ~8.1 GFLOPs
• Layers: The final architecture consists of 129 layers, with the final detection head (Detect layer #22) customized for 4 output classes.

Training Details

Training Configuration

• Epochs: 50
• Image Size: 640x640 pixels
• Optimizer: AdamW
• Learning Rate: 0.00125 (automatically determined by the Ultralytics framework)
• Momentum: 0.9 (automatically determined)

Training Strategy

• Transfer Learning: The model was initialized with pre-trained weights from the COCO dataset, transferring knowledge from 319 of the 355 original layers. This significantly accelerated learning.
• Automatic Hyperparameter Optimization: The framework automatically selected the best optimizer (AdamW) and its corresponding learning rate and momentum, removing the need for manual tuning.
• Dynamic Augmentation Strategy: For the first 40 epochs, a mosaic augmentation was used to expose the model to a wide variety of object contexts. This was strategically turned off for the final 10 epochs to allow the model to refine its performance on whole, un-altered images, leading to a final performance boost.

Key Performance Highlights

• 81.1% mAP@50 on the test set.
• Fast inference thanks to the optimized YOLOv8n architecture.
• Excellent precision for Gun (96.7%) and Grenade (93.1%) detection on the test set.

Model Architecture

• Base Architecture: YOLOv8 Nano (YOLOv8n)
• Input Resolution: 640×640 pixels
• Backbone: Optimized CNN
• Detection Head: Custom 4-class threat detection

Model Files

• best.pt - Main model weights

Inference Instructions

!pip install ultralytics

# process video in batches
import cv2
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
import torch
from tqdm import tqdm

# Configuration
MODEL_REPO = "Subh775/Threat-Detection-YOLOv8n"
INPUT_VIDEO = "input_video.mp4"
OUTPUT_VIDEO = "output_video.mp4"
CONFIDENCE_THRESHOLD = 0.4
BATCH_SIZE = 32  # Adjust based on GPU memory

# Setup device
device = 0 if torch.cuda.is_available() else "cpu"
print(f"Using device: {'GPU' if device == 0 else 'CPU'}")

# Load model
model_path = hf_hub_download(repo_id=MODEL_REPO, filename="weights/best.pt")
model = YOLO(model_path)

# Process video
cap = cv2.VideoCapture(INPUT_VIDEO)
frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(cap.get(cv2.CAP_PROP_FPS))
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))

fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(OUTPUT_VIDEO, fourcc, fps, (frame_width, frame_height))

frames_batch = []
with tqdm(total=total_frames, desc="Processing video") as pbar:
    while cap.isOpened():
        success, frame = cap.read()
        if success:
            frames_batch.append(frame)
            
            if len(frames_batch) == BATCH_SIZE:
                # Batch inference
                results = model(frames_batch, conf=CONFIDENCE_THRESHOLD, 
                              device=device, verbose=False)
                
                # Write annotated frames
                for result in results:
                    annotated_frame = result.plot()
                    out.write(annotated_frame)
                
                pbar.update(len(frames_batch))
                frames_batch = []
        else:
            break

# Process remaining frames
if frames_batch:
    results = model(frames_batch, conf=CONFIDENCE_THRESHOLD, 
                   device=device, verbose=False)
    for result in results:
        annotated_frame = result.plot()
        out.write(annotated_frame)
    pbar.update(len(frames_batch))

cap.release()
out.release()
print(f"Processed video saved to: {OUTPUT_VIDEO}")

Acknowledgments

• Ultralytics for the YOLOv8 architecture and framework.
• Hugging Face for model hosting and community support.
• Roboflow for the dataset.

Disclaimer: This model is for research and educational purposes. It should not be used for deployment in real-world security applications without further extensive validation.