Adaptive museum lighting using CNN-based image segmentation

Abstract

Modern smart lighting systems have the potential to improve lighting conditions by employing adaptive dimming. However, the light intensity pattern of the individual luminaires is often not optimally matching with the region(s) of interest. By using adaptive lighting fixtures with a tunable radiation pattern, much more targeted illumination can be enabled. The already demonstrated adaptive lighting fixtures are typically quite bulky, and require user input to optimize the emitted light. In this paper, a compact lighting fixture is proposed that consists of a high-power LED, a sequence of lenses, an adjustable mirror and tunable diffuser, to achieve an adaptive lighting functionality. This adaptive luminaire is integrated with an embedded system and camera, to realize autonomous beam adjustment based on computer vision in a museum setting. The embedded system performs semantic segmentation to detect painting masks, and adapts the light beam direction and spot size to the detected painting frame via a feedback loop. The system can be considered as a first demonstration of the integration between state-of-the-art computer vision and adaptive light fixtures for indoor lighting.