Analysis of Movement and Activities of Handball Players Using Deep Neural Networks

Area of Science:

• Computer Vision
• Deep Learning
• Sports Analytics

Background:

• Handball is a dynamic team sport with complex player movements and actions.
• Existing computer vision algorithms face challenges in accurately detecting, tracking, and recognizing actions in unconstrained sports scenes.
• There is a need for robust, sensor-independent computer vision solutions for handball analysis.

Purpose of the Study:

• To develop and evaluate deep learning models for player and ball detection, player tracking, and action recognition in handball videos.
• To create a custom handball action dataset for training and validating computer vision models.
• To enable broader adoption of computer vision in handball analysis for both professional and amateur levels.

Main Methods:

• Fine-tuning You Only Look Once (YOLO) and Mask Region-Based Convolutional Neural Network (Mask R-CNN) for player and ball detection.
• Comparing DeepSORT and Bag of tricks for SORT (BoT SORT) algorithms for player tracking.
• Utilizing Inflated 3D Networks (I3D) for handball action recognition, including multi-class and ensemble models.

Main Results:

• Evaluated various object detection models (YOLO, Mask R-CNN) and tracking algorithms (DeepSORT, BoT SORT).
• Developed Inflated 3D Networks (I3D) models for action recognition achieving average F1 scores of 0.69 (ensemble) and 0.75 (multi-class).
• Demonstrated the models' capability to recognize nine distinct handball action classes.

Conclusions:

• Deep learning, specifically I3D networks, offers effective solutions for handball action recognition and video indexing.
• The developed methods provide a foundation for automated analysis of dynamic team sports.
• Further research is needed to address challenges in applying deep learning to complex sports environments.