Zebracorn Labs
Here at the Zebracorns, we believe in learning and pushing boundaries. In pursuit of that, we have published papers, given talks, and have other tidbits of knowledge lying around. We hope you enjoy these as much as we do.
The Incomplete Guide to Using the TX1 for FRC
This guide is not complete. It is merely the beginning on your journey with the TX1. It will not answer all of your questions nor is it meant to (It’s “incomplete” for a reason). Before you begin you should start by looking at the available resources you have and read the documentation available for the TX1. It’s entirely possible that this guide will be out of date by the time you read it. It is not definitive and should not be treated as such.
Team Culture
Presented by Marshall Massengill and Amrita Malur at the FIRST NC workshops.
2016 FRC Workshops - Basic Vision Presentation
How to get started with vision for FRC. This will include basic techniques, GRIP, using co-processors and OpenCV.
2016 FRC Workshops - Advanced Vision Presentation
How to set up and use a Jetson TX1 with advanced object recognition techniques. Covers hardware setup such as powering the board and connecting to the robot network. Also covers vision coding techniques used to find retroreflective tape targets commonly found in FRC games.
nVidia FRC Day Presentation
Team900 was invited to nVidia's FRC First day to present this overview of their 2016 vision code
The Rise of a new HERO
This project began as a thought exercise after one of our mentors had been playing with the new HERO board from Cross the Road Electronics (CTRE). The board was lightweight and versatile enough to be a very reliable robot control system.
Positional Motion Profiling
This paper demonstrates team 900's efforts in using motion profiling on the Talon SRX. It contains a guide for motion profiling in LabVIEW with example code, but it is applicable to all languages. Questions about the paper can be posted here and someone who worked on this project will reply.
ZebraVision 4.0 - Neural Nets
Document describing the structure, implementation and utilization of neural networks for tracking game objects on the field in real time. We used the Caffe library from Berkeley Vision, the OpenCV library, and the DIGITS software from NVIDIA to create the neural network. This specific network was developed to track boulders for the FRC 2015-2016 season but can be applied to track virtually any object with proper data collection and training.
ZebraVision 4.0 - Image Capturing
Zebravision 4.0 is Team 900s vision system for the 2016 season; FIRST Stronghold. Our work was focused around recognizing the vision goals using shape and color based matching, recognizing the boulders using a neural network, and integrating the detection systems into a tracking system using the StereoLabs ZED stereo camera. This paper describes our methods of generating data for the neural networks. We took video of the ball using various lighting conditions and different, single-color backgrounds. We then identified the ball in these videos, took the clearest frames with the ball visible, and applied randomized shifts to the background in order to generate data.
ZebraVision 4.0 - Goal Detection
Zebravision 4.0 is Team 900's vision system for the 2016 season; FIRST Stronghold. Our work was focused around recognizing the vision goals using shape and color based matching, recognizing the boulders using a neural network, and integrating the detection systems into a tracking system using the StereoLabs ZED stereo camera. One of the main features of Team 900's Zebravision code this year was goal detection. This paper gives an overview of the hardware and code used. The system used a Stereolabs ZED RGB-depth camera and green LED rings to highlight the retroreflective tape around the goal. The image was filtered to look for the reflected LED color and thresholded to turn it into binary green / not green image. The code then extracted contours from the image and applied a number of simple filters to rule out blobs which were obviously not goals. The remaining contours were scored in a number of criteria and the best scoring few objects were assumed to be goals. If more than one valid goal is found, several tiebreakers were used to pick one goal to shoot at. If a valid goal was found, the angle and distance to the target was reported; if none were found, a packet with -1.0 distance and angle was returned to the roboRIO.