Deep-Learning Based Action Recognition in Basketball

Abstract

This project endeavors to create a cutting-edge method for identifying movements performed during basketball games through the use of Fast R-CNN and deep learning techniques. The proposed system takes frames extracted from videos, where each frame is labeled with the action being performed (e.g., run, dribble, defense). The system does not rely on separate annotations or label flies, making it more versatile and adaptable to new datasets. This project comprises a series of phases, commencing with the extraction of features through frame pre-processing, followed by the application of a pre-trained Fast R-CNN model utilized for object detection. The final stages of this project focus on training the model using a specially designed dataset class, which defends the input data parameters, such as image files and associated annotations. The training process involves defining the training and validation data loaders, setting up the loss function, and optimizing the model using backpropagation. Upon completion of the model training, it generates a series of bounding boxes that encompass the identified players. These boxes serve as a means of identifying and distinguishing the various actions being executed. The proposed system is evaluated using standard metrics such as accuracy, precision, and recall on a test set of frames. The results show that the system achieves high accuracy in recognizing actions in basketball, demonstrating the potential of the proposed approach for real-world applications. To summarize, the efficacy of Fast R-CNN in perceiving basketball activity is convincingly revealed by this project, underscoring the indispensable value of deep learning strategies and customized dataset classes in creating efficient and precise computer vision technology.