Design

Project API

Our project is split into two key parts: Pursuit-Evasion and Classification.

Blue team: Aditya, Anthony, and Rohan worked on creating the Pursuit-Evasion module and integrating the Classifier unit into the game.

Gold team: Joel and Jason worked on developing the stick classification unit that enabled a user to point to a turtlebot and start the game.

Pursuit-Evasion

Goal: Create the core tag logic and enable bots to evade and pursue each other using AR tags in reference to a Realsense camera. Key portions included:

Using ROS multi-master to handle turtle-bot communication via shared topics.
Each turtle bot would have an associated AR tag, and the bots would be tracked by the Realsense camera via the ar_track_alvar package.
The evader bots will effectively use a "rotate-and-mimic" evasion strategy to escape the chasing bot.
The pursuer bot will use dynamically updating transform information to control its movements to reach the evaders.
Game logic: the game ends once the pursuer makes contact with the evader. This is done via a distance threshold.
Set boundaries so that turtle bots don’t chase each other out of the range of the real sense camera.
Experiment with pursuit-evasion algorithms as an extension or alternative to improve performance or incorporate new game features once we are able to get everything else working.

Pursuit

Define:
- P := Pursuer position relative to camera.
- E := Evader position relative to camera.
Objective:
- min || P - E ||
Secondary objective:
- Coerce evader(s) towards boundaries to get them out

Evasion

Objective:
- Max || P - E ||
Strategies:
- Approach: Rotate Away and Mimic Pursuer Transforms
- When at the boundaries, project evasion trajectory on arena borders. (Move along boundary away from pursuer)
- Boundaries would be specified by AR Tags located at the upper left and lower right corners of our rectangular grid.

Computer Vision

Goal: Be able to recognize which turtle bot the person is pointing to by dissecting point clouds from a real sense camera. Also have a threshold that recognize when the person is not pointing at any turtle bot, in other words, the point has to be close enough to the turtle bot for it to know.

Potentially use a stick to point instead of a hand to make detection and extrapolation easier.
The pointing direction is the first principle component of the pointing object.

Computer Vision Pipeline: (Specifics in Implementation Section)

Color Space Segmentation: RGB → HSV
RGB to Grayscale
Gaussian Blur
Threshold
Morphological Dilation
Contour Detection
Fit Rectangles
Max Area
Apply Segmentation
Morphological Erosion

Integration

Goal: Determine which turtle bot the stick is pointing at.

Integration Pipeline: (Specifics in Implementation Section)

Transform turtle bots from camera frame to stick frame.
Create likelihood model to determine which bot we are most likely pointing at.
Choose the most likely bot!

Page updated

Report abuse