Autoscout

by Colin Haine, April 2025

Description

From a video of a match, it autonomously forms the paths of each robot and how many cycles each robot does.

What it does

get_info.py

• 0: Get TBA data api and youtube video (yt_dlp) video_detection.py
• 1: the ai detects the robots in every frame using YOLOv8
• 2: combine the camera angles and plot detections as points on a top-down map
• 3: clumps points together and remove false positives
• 4: create paths from points that are close to points in the last frame
• 4.5: read bumper numbers every once in a while using easyocr cleanup.py
• 5: stitch the paths together into the full march for each robot
• 6: count how many cycles each robot

ai robot detection model dataset, trained specifically on cave2025: https://universe.roboflow.com/frcrobotfinder/ventura-2025

Next steps

• MAKE IT RELIABLE
  • Correct certain maching algorithms
  • fine tune the joining points
  • handle endgame
  • use match numbers to determine confidence when matching up paths
• Add robustness for multiple competitions:
  • have a folder for each comp, that has the data for the corner locations
  • train an ai based on general lighting, add a parameter to choose which model
  • try out more models
• ADD SCORING
  • implement scoring increments
  • handle algae
• Add speed parameters (ex. skipping every other frame)
  • See how much the reliability drops if so
  • Add checkboxes/drop downs to the ui for parameters
• add tests (useful when installing on another computer since there are so many libraries used)
  • chatgpt should do a good job of this
• try installing on windows and resolve problems
  • create steps to use nvidia GPU
• raise an army to train the ai during worlds
• make the kittens cuter
• add status box for gpu utilized

Installation

1. Install python:
   1. Mac: Brew install python@3.10
   2. Windows: Get python
2. Install git
   1. Mac: Brew install git
   2. Windows: get from git-scm.com
3. Create a venv
4. Install poetry
   1. Mac: Brew install poetry
   2. Windows: curl -sSL https://install.python-poetry.org | python3 -
5. Clone the project
   1. Git clone https://github.com/TinyElephant56/AutoScouter
6. Test it
   1. first, get the youtube video, which downloads to a dir outside of comptracker, ../Captures/
   2. run video_detection.py
7. On a computer with GPU, install cuda drivers
   1. Windows: pain.
   2. Mac: mpu is great