Bridging the Gap between Detection Algorithms and Real-World Challenges

Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Senem Velipasalar Gursoy

Subject Categories

Computer Sciences | Physical Sciences and Mathematics


In recent years, the remarkable progress in the fields of computer vision and machine learning has unleashed a wave of groundbreaking applications, transforming the way we interact with technology on a daily basis. From smartphones to complex hardware systems, these advancements have become integral components of our lives, offering unprecedented capabilities and convenience. At the core of computer vision lies the creation of advanced detection models. These models are crafted to autonomously recognize and scrutinize a diverse range of visual elements, encompassing objects, faces, and anomalies within images and videos. The significance of detection models extends across various applications, spanning from surveillance and security to healthcare and autonomous vehicles. They serve as the fundamental building blocks for comprehending and engaging with the visual realm, forming the basis for numerous innovative technologies and services. This thesis explores the development and application of detection algorithms across multiple domains, addressing real-world challenges and making significant contributions to the field. We begin by focusing on pedestrian detection, proposing novel methodologies that incorporate saliency features and thermal imaging. By leveraging saliency information, our approach enhances the accuracy and robustness of pedestrian detection algorithms, enabling reliable performance even in complex scenes with occlusions or low visibility. Additionally, we harness the power of thermal cameras, which provide valuable thermal signatures, to further improve pedestrian detection performance in challenging scenarios such as nighttime or adverse weather conditions. Next, we turn our attention to the development and reliability assessment of an occupancy sensor system tailored for residential buildings. First, we present an AI-based detection algorithm that we developed by using around 24K images. For this approach, we gathered data for various corner case scenarios, including individuals reclining on couches, individuals resting in beds while being covered by a blanket at various amounts, individuals donning hats and/or sunglasses, diverse indoor camera perspectives, and more. This data was acquired from online sources and supplemented by our own data collection efforts. On the hardware front, we have designed a stand-alone, battery-powered residential occupancy detection system. This system offers a cost-effective, high-precision solution for addressing the limitations of existing occupancy detection approaches. These limitations include (i) not being able to detect stationary occupants; (ii) not being able to classify the source of the motion (such as a pet); (iii) not allowing for embedded or onboard computation, and requiring external or cloud-based processing, especially depending on camera resolution and employed algorithms, (iv) being sensitive to lighting changes, and thus prone to missed detections or false alarms; (v) requiring adjustment of settings for different scenarios, and thus complicating self-commissioning; (vi) not providing high-enough accuracy, especially in the corner cases mentioned above; (vii) being costly; and (viii) not being battery-powered, which limits the ease of use and installation. The system is designed and built to function efficiently on standard alkaline batteries, operating autonomously without relying on cloud-based or external computing resources. It comprises energy-efficient, Artificial Intelligence (AI)-enabled IoT platforms, each equipped with multi-modal sensors for motion, audio, and video data processing. These platforms autonomously analyze sensor data locally, transmitting only binary occupancy status results to a central platform. We conducted extensive real-world testing, covering various challenging scenarios, including individuals in various postures (e.g., lying down, seated), scenes featuring pets (e.g., cats), as well as very low-light and no-light conditions. Integrated laboratory tests demonstrated exceptional accuracy. For daytime tests that lasted over 111 hours, 100% accuracy with zero false positives is achieved. Similarly, for the no-light tests conducted over 10 hours, the accuracy exceeded 99%. Further real-life testing was carried out in three different apartments, totaling approximately 412 hours, with an average accuracy of 99.37%. An extensive evaluation of the platform was conducted to gauge its reliability and performance. Inspired by established methodologies in assessing occupancy sensor systems, the critical importance of thorough evaluation in enhancing the reliability of occupancy detection systems is recognized. The platform, designed for accurate occupancy detection under various conditions, underwent rigorous testing and analysis. This evaluation not only contributes to understanding the system's performance intricacies but also aligns with broader goals of advancing reliable occupancy detection methods, thereby fostering energy efficiency and automation in diverse applications. After proposing ML-based approaches for thermal camera and visible range images, this thesis presents another advanced detection algorithm for another sensor modality, more specifically magnetic resonance imaging (MRI) scans. In the realm of healthcare, our research aims to address the early detection of Alzheimer's disease using MRI scans. By developing attention networks and applying machine learning techniques, we strive to identify preclinical stages of Alzheimer's disease. This early detection is crucial for timely intervention and treatment, potentially leading to better patient outcomes. Leveraging the rich information captured in brain MRI scans, our methodology effectively exploits the attention mechanism to highlight subtle abnormalities and patterns indicative of preclinical stage Alzheimer's disease. Through extensive experiments and comparisons with existing approaches, we demonstrate the efficacy and potential of our proposed methodology for early detection and monitoring of Alzheimer's disease, contributing to the growing body of research in neuroimaging and disease detection. This work is the first one that uses a transformer network that detects this disease in its very early stage when all the indicators, including assessments by doctors, show that the patient is healthy, with no symptoms. Collectively, this thesis delves into the intricacies of detection algorithms and their practical applications in multi-domain imagery, ranging from thermal images to visible-range images to MRI scans. Through our research in pedestrian detection, occupancy detection, occupancy sensor reliability assessment and Alzheimer's disease detection, we strive to bridge the gap between detection algorithms and real-world challenges. Our studies contribute to the field of computer vision and machine learning, offering promising solutions for critical and challenging problems in diverse domains.


SURFACE provides description only. Full text may be available to ProQuest subscribers. Please ask your Librarian for assistance.

This document is currently not available here.