Motion behaviour recognition dataset collected from human perception of collective motion behaviour

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Collective motion in nature (bees, birds, fish) inspires computer-based swarming systems.
• Swarming systems are used in robotics for tasks like formation control and exploration.
• Detecting collective motion behavior is challenging for computers, despite human ease.

Purpose of the Study:

• To enable machine learning models to mimic human perception of collective motion.
• To create a ground truth dataset of human perception for collective behavior recognition.

Main Methods:

• Collected ground truth data via an online survey.
• Participants labeled short videos of simulated 'boid' movements.
• Videos were classified as 'flocking'/'not flocking', 'aligned'/'not aligned', or 'grouped'/'not grouped' using slider-based input.

Main Results:

• Averaged participant responses to create three binary labels per video.
• Analyzed the dataset to confirm machine learning feasibility.
• Demonstrated high accuracy in binary classification learning from human perception data.

Conclusions:

• It is possible for machines to learn binary classification labels from human perception of collective behavior.
• This approach enables AI to recognize complex group dynamics.
• The collected dataset can advance research in agent-based systems and AI perception.