LSTM-Based Soccer Video Summarization Via Event Classification for Highlighting Key Moments

Abstract

Video summarization creates brief synopses of video content by selecting the most informative segments, either as key-frames or key-fragments. This study introduces an advanced method for event classification in soccer videos using a modified stacked Long Short-Term Memory (LSTM) model. By utilizing the Soccer Action Detection Compilation (SADC) dataset, which includes detailed annotations of football events, our method combines VGG16 for feature extraction with LSTM for event classification. The model efficiently identifies crucial events such as goals, goal attempts, and yellow cards, while also filtering out non-essential segments labelled as "No Events." When compared to the Bi-LSTM model, the modified LSTM demonstrates superior performance in terms of precision, recall, and F1-score for several key event classes. Specifically, at epoch 150, the modified LSTM achieves an F1-score of 0.87 for "No Event" segments and perfect precision for "Goal" events, surpassing the Bi-LSTM. These findings underscore the model's stability and accuracy, which ensure the production of high-quality highlight reels by emphasizing important events and minimizing irrelevant content. In the future, by removing non-essential segments recognized as "No Events" from any football match, one can generate perfect highlight reels that capture all crucial moments, providing viewers with a more focused and enjoyable experience.