| dc.description.abstract | The martial art of Shorinji Kempo is a discipline that integrates physical techniques with character development, placing a strong emphasis on precision and flexibility in every movement. The complexity of its techniques requires direct supervision from an instructor to ensure proper mastery and minimize the risk of injury. However, limited access to expert supervision often becomes an obstacle, thereby hindering progress and increasing the potential for technical errors among practitioners. This research aims to develop an automated system based on deep learning to detect and classify basic Shorinji Kempo movements, serving as a self-training aid.This system utilizes MediaPipe to extract body Pose features from training videos. Specifically, it extracts keypoints as 2D coordinates from 59 key points (17 Pose, 21 left hand, 21 right hand) to capture posture in detail. This sequential data is then processed using a Bidirectional Long Short-Term Memory (BiLSTM) architecture with Attention mechanism. BiLSTM was chosen for its ability to understand the context of movements from both forward and backward directions, while the Attention mechanism allows the model to focus on the most significant frames. The dataset used comprises 540 videos of nine basic movements, which underwent augmentation to enhance the model's learning. The test results show that the system achieved a very high training accuracy of 82.22%. In addition to classification, the system also provides corrective feedback through a geometric analysis of the user's body position against an ideal standard. Thus, this system has the potential to make Shorinji Kempo training more effective, safe, and accessible. | en_US |
