Group Emotion Detection Based on Social Robot Perception.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Social robots are increasingly deployed in public spaces, necessitating advanced Human-Robot Interaction (HRI) designs.
• Current HRI lacks efficient methods for robots to perceive and react to group emotions in social settings.
• Recognizing group emotions from a robot's perspective, considering its sensory limitations, is an underexplored area.

Purpose of the Study:

• To develop a system for robots to recognize social scenes, identify groups of people, and detect their prevailing emotions.
• To propose a robocentric approach for emotion recognition, utilizing visual cues like face size.
• To create validated datasets for training and testing emotion recognition in simulated social environments.

Main Methods:

• Utilized the Visual Geometry Group (VGG) neural network (VGGFace) for individual emotion recognition from video frames.
• Implemented a fusion method to aggregate individual emotions for frame-level emotion detection.
• Developed a strategy for creating simulated datasets using the Robot Operating System (ROS) and Gazebo for robot-centric data collection.

Main Results:

• Achieved high accuracy in individual emotion detection (99.79%).
• Demonstrated effective group emotion (scene emotion) detection with 90.84% accuracy in a cafeteria and 89.78% in a museum simulation.
• Validated the proposed approach in realistic simulated social environments (museum, cafeteria) within ROS/Gazebo.

Conclusions:

• The developed system successfully enables social robots to recognize group emotions from a robocentric viewpoint.
• The findings contribute to more sophisticated and context-aware HRI by allowing robots to understand group dynamics.
• The proposed dataset creation strategy and robocentric approach offer a foundation for future research in social robot perception.