Occupancy information is a fundamental resource in smart building management, as it provides insight into how spaces are used and allows building systems to respond dynamically. It can capture different levels of detail, ranging from simple presence detection to more advanced forms such as activity recognition. Within this hierarchy, people counting represents a critical level of occupancy information, focusing on estimating the number of individuals in a given space. Accurate people counting enables a wide range of applications, including demand-driven Heating, Ventilation, and Air Conditioning (HVAC) control, optimization of space utilization, compliance with safety and security regulations, and the enhancement of occupant comfort and experience. 

Despite its importance, people counting with non-privacy-invasive sensors remains a challenging task.Vision-based and device-tracking approaches often achieve high accuracy but raise concerns over privacy, cost, and scalability, limiting their practical use. While non-privacy-invasive methods have been widely explored, they still lack low-cost and scalable solutions that leverage sensors already embedded in smart buildings. Binary PIR devices, for example, are common in building automation systems for presence detection, yet their potential for multi-person counting remains underexplored. No comprehensive review has mapped the role of binary and signal-based PIR sensors across different levels of occupancy information or examined their integration with machine learning in real-world settings. Previous studies have typically focused on controlled lab setups or dense multi-sensor deployments, which limit generalizability. In addition, contextual booking data has rarely been incorporated directly into occupancy estimation models. Finally, advanced temporal deep learning models such as Transformers, along with systematic analyses of historical window size and feature importance based on environmental data, are still underexplored, leaving open opportunities to optimize both modeling strategies and sensor deployment in real-world smart buildings.

 This thesis addresses these gaps through four studies. First, a systematic literature review synthesizes the role of binary and signal-based PIR sensors across different levels of occupancy information, identifying research gaps and motivating new approaches. Second, an event-based framework demonstrates that reliable occupancy estimates can be achieved using only two binary PIR sensors combined with machine learning. Third, a context-aware extension integrates PIR data with booking records, improving estimation accuracy and exposing patterns of underutilization, overbooking, and mismatches between planned and actual usage. Finally, a fixed-interval time-series framework evaluates advanced deep learning architectures, including RNNs, LSTMs, GRUs, CNN-LSTMs, and Transformers, showing how temporal context length and multimodal feature sets influence performance.

Taken together, these studies advance the state of the art by demonstrating that accurate, scalable, and privacy-preserving people counting does not require invasive or costly infrastructures. By leveraging PIR sensors already present in many buildings, enriched with contextual and environmental data and analyzed through modern machine learning techniques, smart buildings can achieve reliable occupancy estimation. Beyond technical performance, the findings highlight the broader operational value of occupancy information for efficient space management, energy savings, and improved occupant experience, while laying the foundation for next-generation intelligent building systems.