Program Type

Graduate

Faculty Advisor

Dr. Robin Ghosh

Document Type

Presentation

Location

Face-to-face

Start Date

18-4-2024 11:00 AM

End Date

18-4-2024 11:30 AM

Abstract

The field of sports video analysis using deep learning is rapidly advancing. Proper classification and analysis of sports videos are essential to manage the growing sports media content. It offers numerous benefits for the media, advertising, analytics, and education sectors. Soccer, also known as football, worldwide, is among the most popular sports. This research study used a deep learning-based approach for soccer action detection. Deep learning has become a popular machine learning technique, especially for image and video classification. We have used the SoccerAct dataset, which consists of ten soccer actions like corner, foul, freekick, goal kick, long pass, on target, penalty, short pass, substitution, and throw-in. Our study analyzes ConvLSTM and LRCN, two different methods for soccer action detection in video clips. ConvLSTM is an extension of the Long Short-Term Memory (LSTM) architecture that allows the modeling of both spatial and temporal dependencies in sequential data by integrating convolutional operations into recurrent neural networks (RNNs).On the other hand, convolutional neural networks (CNNs) and recurrent units -usually LSTM or GRU-combine in Long-term Recurrent Convolutional Networks (LRCN) to capture temporal relationships through RNNs and spatial characteristics through CNNs. The ConvLSTM model achieved 70%, and the LRCN model achieved 71% accuracy on its first iteration. The model's performance is enhanced through fine-tuning the pre-trained model, incorporating batch normalization to mitigate overfitting, and conducting hyperparameter tuning. Our study highlights how deep learning techniques can improve soccer action detection systems significantly. A soccer action detection model could be used in various real-world contexts, including sports performance analysis, journalism and reporting, broadcasting, and many more.

Share

COinS
 
Apr 18th, 11:00 AM Apr 18th, 11:30 AM

League of Learning: Deep Learning for Soccer Action Video Classification

Face-to-face

The field of sports video analysis using deep learning is rapidly advancing. Proper classification and analysis of sports videos are essential to manage the growing sports media content. It offers numerous benefits for the media, advertising, analytics, and education sectors. Soccer, also known as football, worldwide, is among the most popular sports. This research study used a deep learning-based approach for soccer action detection. Deep learning has become a popular machine learning technique, especially for image and video classification. We have used the SoccerAct dataset, which consists of ten soccer actions like corner, foul, freekick, goal kick, long pass, on target, penalty, short pass, substitution, and throw-in. Our study analyzes ConvLSTM and LRCN, two different methods for soccer action detection in video clips. ConvLSTM is an extension of the Long Short-Term Memory (LSTM) architecture that allows the modeling of both spatial and temporal dependencies in sequential data by integrating convolutional operations into recurrent neural networks (RNNs).On the other hand, convolutional neural networks (CNNs) and recurrent units -usually LSTM or GRU-combine in Long-term Recurrent Convolutional Networks (LRCN) to capture temporal relationships through RNNs and spatial characteristics through CNNs. The ConvLSTM model achieved 70%, and the LRCN model achieved 71% accuracy on its first iteration. The model's performance is enhanced through fine-tuning the pre-trained model, incorporating batch normalization to mitigate overfitting, and conducting hyperparameter tuning. Our study highlights how deep learning techniques can improve soccer action detection systems significantly. A soccer action detection model could be used in various real-world contexts, including sports performance analysis, journalism and reporting, broadcasting, and many more.