Vision Rig | Sujjay Karthikeyan

The Setup.

The Limelight 2 is engineered hardware: a dedicated vision coprocessor running at 90fps with onboard target detection, designed for FRC robots. But there's nothing robot-specific about its output. Pulled off the competition robot and paired with a Raspberry Pi, it becomes the cheapest high-frame-rate machine vision rig you can build.

This project explores what that rig can do outside of FRC — specifically for drone-adjacent work: landing pad detection and visual odometry experiments as a complement to the VTOL project's perception stack.

Vision Coprocessor

Limelight 2

Host Compute

Raspberry Pi

CV Library

OpenCV

Language

Python

Frame Rate

90fps

Vision Pipeline.

01 Capture

Limelight 2

90fps @ 320×240, hardware-accelerated target pipeline

02 Network

NetworkTables / HTTP

Raw target data streamed to Pi over LAN — tx, ty, ta, ts values

03 OpenCV

Frame Processing

Contour detection, pose estimation, homography transforms

04 Output

Pose + Odometry

Camera-frame position relative to landing pad / tracked feature

05 Log

Timestamped Data

Frame-by-frame pose log for replay and algorithm tuning

FRC vs. Non-FRC Use.

Original — FRC Robot

Target: AprilTag fiducials on field elements
Goal: Robot-to-target alignment for shooter
Output: Angle + distance fed to PID
Environment: Indoors, controlled lighting
Latency budget: <20ms to motor command

Repurposed — Vision Testbed

Target: Landing pad markers + natural features
Goal: Relative pose estimation for UAV descent
Output: Homography-derived 6-DOF pose
Environment: Outdoor, variable lighting
Latency budget: Logged for offline analysis

Experiments.

Landing Pad Detection ArUco marker detection on a custom landing pad, computing the camera-to-marker transform at various descent angles and altitudes. Baseline data for eventual integration with the VTOL's approach phase.

Optical Flow Odometry Frame-to-frame feature tracking using Lucas-Kanade optical flow to estimate camera velocity. Evaluating feasibility as a dead-reckoning backup when GPS is degraded.

Limelight vs. Pi Camera Comparison Side-by-side latency and accuracy comparison between the Limelight's dedicated pipeline and a Pi Camera 3 running the same OpenCV pipeline — quantifying the hardware advantage of a dedicated coprocessor.

Perception Stack.

This rig is the physical testbed for the perception layer that would pair with the VTOL UAV. The Limelight's 90fps rate and low-latency pipeline make it unusually capable for a sub-$300 hardware stack. The non-FRC experiments here feed directly into what a real autonomous landing system would need.