Prediksi Arah Tendangan Penalti Berbasis Estimasi Pose dan Run Up Sequence Penendang Menggunakan Algoritma Long Short Term Memory (LSTM) dan 1D Convolutional Neural Network (CNN)

Abstract

Penalty kicks are decisive moments that can determine football match outcomes. For goalkeepers, the ability to read kick direction before the ball is struck provides an invaluable advantage. This research develops a penalty kick direction prediction system as a goalkeeper training tool by analyzing the kicker's body movements during the run-up phase. The system employs 3D pose estimation from 429 penalty kick videos and a hybrid 1D CNN-LSTM architecture that combines frame-level spatial pattern analysis with overall movement dynamics understanding. The development process faced significant challenges, particularly limited dataset size and inherent ambiguity in center-directed kicks. Ablation study revealed that the hybrid architecture achieved 52.33% accuracy, improving 11.63% over spatial-only models. A key finding is that temporal component is approximately 2.5 times more critical than spatial component confirming that "how the kicker moves toward the ball" is more informative than "body position at any single moment". While accuracy has not yet reached ideal levels for direct implementation, this research provides scientific and technical foundations for developing computer vision-based goalkeeper training tools, and opens future research directions through integration of additional visual information such as ball trajectory and field context.