Zhang, Wuxia
(2025)
Application of vision-based deep learning method for building occupancy and thermal comfort predictions.
PhD thesis, University of Nottingham.
Abstract
Buildings are responsible for approximately 40% of global energy consumption and 30% of greenhouse gas emissions, highlighting their central role in sustainability efforts. Heating, ventilation, and air-conditioning (HVAC) systems are typically operated based on fixed schedules or occupant input, often misaligned with actual occupancy patterns. This mismatch leads to energy inefficiencies and occupant discomfort. Vision-based sensing technologies, particularly those employing cameras, offer the potential for real-time occupancy detection, providing a foundation for more adaptive building control strategies.
This thesis investigates the integration of vision-based deep learning methods for occupancy prediction and personalised thermal comfort modelling, aiming to enhance building energy performance and occupant well-being. Following a comprehensive review of the literature and identification of key research gaps, three main studies were conducted.
The first study develops and compares eight deep learning algorithms for vision-based occupancy detection. Annotated datasets were created from images collected in controlled environments. The models were evaluated using detection accuracy, mean average precision (mAP), and inference speed. Among the models tested, YOLOv8x achieved the highest accuracy (F1 score 0.87), while YOLOv8n offered a balance between accuracy and processing speed. When integrated into building simulations using IESVE, the predicted occupancy profiles led to significant improvements in HVAC energy estimation. A daily heating demand deviation of 13.45% in the base case was reduced to under 7% using deep learning models.
The second study compares the performance of thermal and standard RGB cameras for occupancy detection. Both modalities achieved comparable accuracy—around 70% with YOLOv8 and 80% with YOLOv10—given adequate training data. RGB cameras provide high-resolution detail but are susceptible to privacy concerns and visual interference. Thermal cameras, while offering better privacy and low-light performance, face limitations in scenarios involving overlapping occupants and residual heat. The results support thermal imaging as a viable, privacy-preserving alternative in suitable contexts.
The third study proposes a vision-based thermal comfort prediction model using deep learning and thermal imagery, offering an alternative to the conventional Predicted Mean Vote (PMV) approach. Personalised models achieved up to 68.49% accuracy in intra-subject tests, indicating potential for individual comfort prediction. However, reduced performance in cross-subject testing underscored the challenge of generalising thermal comfort models across diverse users.
Through systematic evaluation of algorithms, camera types, and comfort prediction strategies, this research advances the development of intelligent building systems. The findings suggest that vision-based approaches can support real-time, occupant-centred control of HVAC systems, contributing to improved energy efficiency and thermal comfort. Future work should focus on expanding datasets, refining model generalisability, and validating performance in real-world conditions.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
