Edge–Cloud Inference Deployment for Perception-Driven Automation - A systems evaluation of stereo-vision assisted zone-occupancy for building automation control under practical accuracy, latency, and energy constraints

Abstract

This thesis evaluates how edge and cloud inference affect a stereo-depth-assisted, zone-based lighting controller for building automation. The system combines person detection with depth-based spatial assignment to estimate whether predefined lighting zones are occupied. Several model configurations are compared under edge and cloud deployment using controlled scenarios and full-day analytical lighting reconstructions. The results show that deployment placement affects both perception and system behaviour. Cloud inference improves recall in more difficult conditions, such as low light and obstruction, but introduces transport delay, queue behaviour, higher measured processing energy, and greater privacy exposure. Edge inference provides lower processing-energy use, avoids remote communication overhead, and keeps more data local, but is less robust in difficult visual conditions. Across the evaluated configurations, the projected zone-aware controller reduces lighting energy compared with less spatially specific baseline strategies.