Subteam · 03

Perception.

Computer vision for autonomous navigation and real-time object detection from the air.

About

Giving the drone
its eyes.

The Perception subteam develops the computer vision pipelines that let the drone see the world below, identify competition targets, and feed spatial information back to the autonomy stack.

We work with onboard cameras, train detection models, and implement real-time inference pipelines that run reliably at altitude during fully autonomous flight.

Focus Areas

What we work on.

Target Detection & Classification

Training and deploying YOLO11n to locate and classify competition targets — shapes, letters, colors — from aerial imagery, running under TensorRT on the onboard Jetson.

Dataset Collection & Training

Building labeled datasets in Roboflow and supplementing them with synthetic aerial imagery rendered from Blender assets and environments.

Geolocation

Computing GPS coordinates of detected targets by combining camera geometry, drone pose (GPS + attitude), and ground-plane projection.

Model Training Pipeline

Labeling aerial imagery, augmenting datasets, and training YOLO11n to detect and classify competition targets with high accuracy under real flight conditions, then exporting to TensorRT for deployment.

CAM-01 TARGET · 94.2% TARGET · 91.5% TARGET · 88.3% TARGET · 96.7% ALL TARGETS ACQUIRED

Technology

Our stack.

Framework Python and PyTorch for training, TensorRT for optimized inference on NVIDIA Jetson
Detection Model YOLO11n
Camera Arducam AR0234, 2.3 MP global shutter — 1920×1200 @ 30 fps
ROS 2 Integration Control parameters are mediated over uXRCE-DDS (Micro XRCE-DDS) between the Jetson and the Pixhawk
Ground Truth Validation Roboflow datasets paired with simulated assets and environments generated in Blender

Meetings

When we meet.

Vision Dev Session

Mondays 2:00–3:00 PM

SlugWorks / Remote

No prior CV experience needed — we'll get you up to speed.