Related Experiment Video
Updated: Jul 6, 2026
12:39
A Methodology for Capturing Joint Visual Attention Using Mobile Eye-Trackers
Published on: January 18, 2020
8.1K
Dynamic graph neural networks for UAV-based group activity recognition in structured team sports
Ishrat Zahra1,2, Yanfeng Wu1, Haifa F Alhasson3
1Guodian Nanjing Automation Co., Ltd., Nanjing, China.
Frontiers in Neurorobotics
|September 24, 2025
Summary
This study introduces a deep learning system for recognizing group activities in sports using multi-modal features and a Dynamic Graph Neural Network. The framework achieves high accuracy on diverse datasets, demonstrating robust performance in real-world applications.
Area of Science:
- Computer Vision
- Machine Learning
- Robotics
Background:
- Group activity recognition is crucial for surveillance, robotics, and autonomous systems.
- Sports scenarios pose unique challenges due to dynamic interactions and occlusions.
Purpose of the Study:
- To develop a deep learning system for robust multi-person behavior recognition in dynamic environments.
- To integrate appearance, skeletal, and motion features for enhanced group activity recognition.
Main Methods:
- Utilized YOLOv11 for object detection and SORT for tracking.
- Extracted multi-modal features (HOG, LBP, SIFT, skeletal, motion context) and optimized them using genetic algorithms.
- Employed a Dynamic Graph Neural Network (DGNN) with Bi-LSTM to model spatio-temporal dynamics.
Main Results:
- Achieved 94.5% accuracy on a volleyball dataset, 91.8% on SoccerTrack UAV, and 91.1% on an NBA basketball dataset.
- Demonstrated efficient inference times (0.18-0.20s per frame).
- Validated adaptability across conventional and UAV-based video sources.
Conclusions:
- The proposed framework offers high performance and computational efficiency for group activity recognition.
- The system is adaptable to diverse sports scenarios and video perspectives, including drone footage.
