Optimization of table tennis target detection algorithm guided by multi-scale feature fusion of deep learning

Area of Science:

• Computer Vision
• Sports Technology

Background:

• Accurate ball tracking is crucial for table tennis analysis and athlete training.
• Existing target detection methods often struggle with the speed and complexity of table tennis matches.

Purpose of the Study:

• To develop a deep learning-based table tennis target detection (TD) method using multi-scale feature fusion (MFF).
• To enhance detection accuracy for the table tennis ball, optimize athlete training, and improve technical skills.

Main Methods:

• Utilized FAST Region-based Convolutional Neural Network (FAST R-CNN) for initial target detection.
• Implemented a multi-scale feature fusion (MFF) guidance method to integrate information from different feature levels.
• Focused on improving the accuracy of table tennis TD through DL and MFF.

Main Results:

• The proposed target detection algorithm (TDA) achieved a mean Average Precision (mAP) of 87.3% on the test set.
• Demonstrated superior performance and robustness compared to other existing TDAs.
• Validated the effectiveness of the DL TDA combined with MFF.

Conclusions:

• The proposed DL-based TDA with MFF significantly improves table tennis ball detection accuracy.
• The method shows high robustness and potential for application in various detection fields.
• This technology can aid in the practical application of target detection in real-world scenarios, including sports analytics.