SimulSilat: A Deep Learning-Based Simulator for Silat Training

This paper presents an investigation into the application of deep learning techniques and a simulated user interface system for Silat movement recognition. Silat, a traditional martial art originating from Southeast Asia, offers numerous physical and mental health benefits. To enhance the training experience and improve skill development, this study developed graphical user interfaces using MATLAB App Designer and employed a web camera connected to a laptop as the main hardware setup. The SimulSilat simulator facilitated video recording and registration of specific Silat stances, while various convolutional neural network (CNN) architectures, namely ResNet-18, ResNet-50, ResNet-101, DarkNet-19 and DarkNet-53, were evaluated for movement recognition. The simulation setup involved a white screen as the training area background, with portable LED lights strategically placed for optimal ambient lighting. The system operated in two modes: Live mode, which identified Silat movements and provided real-time accuracy, and Demo mode, which performed classification on previously saved video recordings. A dataset of 1,500 images captured from 50 Silat practitioners was utilized, with 70% allocated for training and 30% for validation. The results demonstrated the performance and generalization capabilities of the CNN architectures. DarkNet-53 emerged as the most promising architecture, exhibiting superior accuracy and generalization over the training epochs. ResNet-50 and ResNet-101 also showcased competitive performance, capturing relevant features from Silat movements. The findings emphasize the potential of deep learning and simulated user interfaces in advancing Silat training methodologies.