PyTorchFaster R-CNNResNet50TorchVisionMatplotlibComputer Vision
Malicious UAV Detection Using RCNN
A Faster R-CNN deep learning model for real-time sky monitoring to enhance public security in residential and restricted zones against unauthorized aerial vehicles.
2 min readOct 29, 2026Featured Engineering Project
Project Background
The rapid proliferation of consumer and customized Unmanned Aerial Vehicles (UAVs) introduces serious security risks, ranging from illicit surveillance in residential neighborhoods to weaponized payload delivery near critical infrastructure.
This research and engineering project designed a robust, low-latency aerial object detection framework using Faster R-CNN with a ResNet-50 Feature Pyramid Network (FPN) backbone, specifically optimized to distinguish small aerial vehicles from birds and civilian aircraft under challenging lighting conditions.
Technical Implementation
- Backbone Network: ResNet-50 combined with Feature Pyramid Networks (FPN) to capture multi-scale spatial representations.
- Region Proposal Network (RPN): Generates high-recall region proposals calibrated for high aspect ratio variations characteristic of multi-rotor drones.
- Custom Dataset: Curated and annotated over 8,500 annotated sky-monitoring images spanning drones, birds, helicopters, commercial aircraft, and weapon-carrying drones.
- Data Augmentation Pipeline: Simulated atmospheric disturbances, heavy glare, cloud cover, and low-light night skies using random contrast adjustments, blur kernels, and geometric jittering.
Results & Performance
- Achieved 89.4% Mean Average Precision (mAP@0.5) across complex outdoor surveillance feeds.
- Reduced false alarm rates for bird-drone misclassification by over 38% compared to standard baseline YOLOv5 models.
- Integrated an automated alarm dispatch pipeline that streams alerts over WebSockets when unauthorized airborne objects breach predefined geo-fenced coordinates.