Edge-Optimized Vision Model

Real-Time UAV Object Detection & Localization

A lightweight deep neural network fine-tuned on 1,359 aerial drone images. Engineered with a 1500px input resolution and anchor-free detection head to identify distant micro-UAVs and multi-rotors with 86.8% mAP@50 at ~7.7ms latency.

Test Detection Sandbox Google Colab Kaggle Dataset

Precision mAP@50 86.8%

Inference Latency ~7.7 ms

Input Resolution 1500 px

Model Size 6.5 MB

Inspection Workspace

Live Inference Sandbox & Detection Viewer

Inspect fine-tuned YOLOv8 model outputs with interactive bounding boxes, confidence filters, and telemetry metrics.

File Upload & Ingestion

Target image and video stream gateway

LOCKED

Upload Target Image / Video

Accepts: .JPG, .PNG, .MP4 (Resolution: 1500px)

Upload Disabled in Demo: Live inference requires an active GPU server. File upload is currently disabled in this public static demo. Please select from the pre-processed sample detections below to inspect model output.

YOLOv8 Output Quadcopter in Twilight Sky

Min Conf: 50%

High-altitude tactical surveillance quadcopter localized in mid-air against overcast evening sky. Anchor-free detection head retains rotor and landing skid features with 94.6% confidence.

Confidence Score

94.6%

Inference Latency

7.7 ms

Detected Target

Quadcopter Drone

Frame Resolution

1500 × 844 px

Detection Archive

Pre-Processed Sample Detections

Click any sample tile below to load the image into the interactive detection canvas and inspect the model's bounding boxes, confidence tags, and inference telemetry.

Quantitative Evaluation

Model Performance & Benchmarks

Evaluated across 1,359 high-resolution aerial validation benchmark images from the Kaggle UAV dataset.

Model Architecture	mAP@50 (%)	Precision (%)	GPU Latency (ms)	Model Size (MB)	Edge Feasibility
DroneGuard YOLOv8n (Ours)	86.8%	85.0%	7.7 ms	6.5 MB	OPTIMAL (95+ FPS)
YOLOv5s Baseline	79.4%	78.2%	14.2 ms	14.8 MB	BASELINE
SSD-MobileNet V2	68.1%	71.0%	11.5 ms	19.2 MB	BASELINE
Faster R-CNN ResNet50	84.2%	81.5%	52.0 ms	168.0 MB	HEAVY / NON-REALTIME

Architecture Highlights

Key Engineering Advantages

Designed specifically to solve small aerial target detection challenges in open airspace.

Ultra-Low Latency (~7.7 ms)

Sub-10ms inference allows continuous 95+ FPS tracking on edge GPUs, enabling real-time detection and interception feeds.

High-Res Head (1500 px)

Preserves micro-UAV structural features, drastically minimizing false positives against birds, clouds, and distant objects.

Compact Footprint (6.5 MB)

Lightweight weights fit directly onto embedded hardware including NVIDIA Jetson Nano/Orin and on-board drone microcomputers.

Anchor-Free Detection Head

Direct box regression with decoupled classification and localization heads for superior convergence across diverse aspect ratios.

Real-Time UAV Object Detection & Localization

Live Inference Sandbox & Detection Viewer

File Upload & Ingestion

Pre-Processed Sample Detections

Model Performance & Benchmarks

Key Engineering Advantages

This is a Live Demo