Zhengbo Zou

The hazardous working environment of construction workers leads to fatal accidents, whereas it is challenging to balance worker safety and productivity. Egocentric (i.e., first-person view) pose estimation is effective in identifying, localizing, and tracking worker poses under construction scenarios. It can be used as a human-centric approach to reduce fatalities and increase productivity through various downstream applications within Intelligent Construction, such as safety monitoring and human-robot collaboration. However, most of the egocentric pose estimation models are trained with daily life poses and have limited generalization ability to construction poses. The scarcity of publicly available datasets designated for egocentric worker pose estimation further impedes the training and evaluation of existing and newly developed egocentric worker pose estimation models. In this regard, this thesis presents a dual-modules egocentric worker pose estimation approach, which effectively addresses the major issues of self-occlusion and depth ambiguity in worker pose estimation. The proposed approach employs a supervised 2D module to address self-occlusion in construction poses and an unsupervised 3D module to handle depth ambiguity using only monocular-view input. It not only achieves a PCKh@0.5 of 98.7% in 2D pose estimation and a N-MPJPE of 96.43mm in 3D pose estimation, but also demonstrates strong generalization ability in the construction domain. To address the significant gap in egocentric pose estimation dataset for construction environments, this thesis also introduces ICON-Pose, the first open dataset includes thousands of egocentric images and corresponding 2D and 3D worker poses data across 63 action types. ICON-Pose accurately depicts complex construction worker poses with the proposed approach and shows a high level of diversity in action types, data forms, participants, and construction site settings. This dataset has the potential to serve as a general benchmark dataset not only for subsequent analyses in Intelligent Construction but also in the broader field of egocentric vision, particularly for capturing and analyzing out-of-distribution poses. It is expected to support artificial intelligence research in the construction domain along with the egocentric worker pose estimation approach.

View record

This thesis introduces an integrated approach aimed at boosting energy efficiency and advancing sustainability in buildings via innovative Demand Response (DR) programs and the intelligent management of Heating, Ventilation, and Air Conditioning (HVAC) systems. By addressing the fragmented efforts in current DR initiatives and the limitations of traditional HVAC control methods, this study introduces two groundbreaking frameworks that collectively provide an economically viable pathway towards reducing carbon emissions, elevating energy efficiency, and improving comfort for occupants in the built environment.First, I propose an integrated DR-based framework that utilizes medium-term electricity Demand Forecasting (DF) and optimal design and management of Battery Energy Storage Systems (BESS). This approach, leveraging robust 30-day ahead DF based on a state-of-the-art (SOTA) Transformer model, delivers significant electricity cost savings of C$ 311K and a reduction of 471 tonnes of CO₂-equivalent (CO₂-e) for 72 target buildings over winter months. These results underscore the framework's potential to revolutionize energy management and sustainability on an urban scale.Concurrently, I address the rigidity of traditional Rule-Based Feedback Control (RBFC) systems in HVAC control by introducing a novel Deep Reinforcement Learning (DRL) framework, which is expected to respond to unseen system dynamics effectively. This framework incorporates the same Transformer model for superior Time-Series forecasting (TF), enabling a more accurate RL training environment. The study demonstrates the framework's effectiveness in HVAC system modeling and control, achieving an average of 23.8% higher prediction accuracy of HVAC system operations over baseline models. Moreover, the proposed past observable RL agent significantly enhances performance, yielding a 44.2% and 39.6% improvement in synthetic reward metrics (corresponding to energy consumption and thermal discomfort), compared to RBFC and standard RL agents.Together, these integrated frameworks highlight the synergy between advanced DR strategies and intelligent HVAC control, facilitated by cutting-edge machine learning (ML) techniques. By combining precise DF, optimal energy storage management, and adaptive HVAC control, this thesis contributes to the fields of sustainable building design and operational optimization. The findings not only showcase the potential for substantial economic and environmental benefits, but also pave the way for future research in applying advanced computational methods for building management.

View record