Search for a command to run...
Recent advancements in deep learning have significantly enhanced body gesture recognition, enabling real-time interaction between humans and machines through the modeling of spatial–temporal features. However, many existing approaches primarily rely on frame-based or visual feature representations and are often evaluated in offline settings, which limits their stability and responsiveness when applied to real-time 3D game environments that require continuous and dynamic player movement. In this paper, we develop a gesture-controlled endless runner game using a skeleton-based Graph Neural Network–Long Short-Term Memory (GNN–LSTM) model. The proposed system enables real-time interaction without the need for conventional input devices and is directly integrated into a Unity-based game environment. A dataset of 1,000 gesture videos across five classes (Jump In Place, Jump Left, Jump Right, Looking Down, and Still Pose) is processed using MediaPipe Pose to extract 33 body keypoints per frame, which are then normalized and represented as graph structures to capture spatial and temporal motion patterns. Experimental results show that the GNN–LSTM model achieves a validation accuracy of up to 97.5% and a test accuracy of 96%. Although CNN–LSTM attains slightly higher test accuracy, the GNN–LSTM model demonstrates more stable validation performance and robustness by leveraging skeleton-based representations, making it more suitable for real-time gesture control in interactive gameplay. Integrated with Unity, the proposed system allows intuitive and responsive control of character movements during gameplay. These findings highlight the effectiveness of temporal graph-based representations for stable and natural gesture-based human–computer interaction in real-time 3D games.