Analysis of baseball behavior recognition model based on Dual-GCN improved by motion weights

Area of Science:

• Computer Vision
• Sports Analytics
• Machine Learning

Background:

• Baseball behavior recognition faces challenges with accuracy and character interaction.
• Existing models struggle to effectively capture the nuances of baseball actions.

Purpose of the Study:

• To develop an improved model for accurate baseball behavior recognition.
• To enhance the connection and contribution between characters in baseball videos.
• To overcome limitations in current baseball action recognition systems.

Main Methods:

• A novel motion weight improvement model based on a dual-graph convolutional network (DGCN) was proposed.
• The DGCN was utilized for behavior recognition and key region segmentation in baseball video images.
• Motion weights were incorporated to enhance character correlation and contribution.

Main Results:

• The model achieved optimal performance at a 1/2w frame rate, 1/2H key area width, and 2 key areas.
• The highest accuracy reached 94.84%, surpassing the hierarchical temporal depth model by 12.06%.
• Incorporating motion weights further improved accuracy by 3.45%.

Conclusions:

• The proposed dual-graph convolutional network model significantly enhances baseball behavior recognition accuracy.
• The motion weight improvement effectively boosts performance and character interaction analysis.
• This research offers valuable insights for advancing behavior recognition in sports analytics, particularly for baseball.