Machine Learning-Based Human Action Estimation for Interactive Spatial Lighting Control

This study investigates the integration of machine learning and architectural lighting design by proposing a proof-of-concept adaptive lighting system driven by human actions and spatial position. A custom video dataset was created based on five actions—standing, sitting, walking, running, and dancing— and three positional categories within a defined space. Two different machine learning approaches were evaluated for human action recognition: a skeletonbased model using MediaPipe pose extraction with an LSTM architecture, and a pixel-based approach combining feature extraction from raw video frames with an MLP classifier. The classified action and position data were mapped to predefined lighting schemes generated parametrically using Grasshopper, enabling context-aware lighting recommendations. The results show that while action classification accuracy is limited due to dataset size, position recognition achieves high reliability. The study highlights the potential of action-oriented, human-centered lighting systems and outlines directions for future research involving larger datasets and user-centered evaluations.