Repository of DUTH

Computer vision and artificial intelligence methods for video analysis and 3d scene synthesis (Doctoral thesis)

Βροχίδης, Αλέξανδρος/ Vrochidis, Alexandros

The rapid growth of video content across domains such as surveillance, live streaming, remote education, and virtual events has created a need for intelligent systems capable of efficiently analyzing and summarizing complex visual data. At the same time, high-fidelity scene synthesis has become critical for environmental monitoring, cultural heritage preservation, and immersive applications. Traditional approaches often struggle with multi-modal data, environmental variability, and dataset-specific challenges, limiting their effectiveness in real-world scenarios. This thesis addresses these challenges by developing multiple complementary frameworks that enable automated video understanding, accurate scene synthesis, and enhanced visual data quality across diverse environments, including both terrestrial and underwater contexts. The primary aim of this research is to advance the understanding and practical application of video analysis and scene synthesis. Specifically, the study develops frameworks for automated video summarization that integrate audio, visual, and textual cues, high-fidelity 3D reconstruction methods that leverage sensor fusion and fine-detail modeling, and image enhancement techniques that optimize quality according to dataset characteristics. By treating these components as distinct yet interoperable frameworks, the research provides scalable, environment-aware solutions that can be deployed in real-world applications with varying computational and operational constraints. The video analysis framework focuses on extracting meaningful information and generating concise summaries from complex video streams, integrating multi-modal cues to improve comprehension and accessibility. Scene reconstruction methods address challenges in both terrestrial and underwater environments, leveraging sensor data to produce high-fidelity, environment-aware 3D models that capture fine structural details. The image enhancement framework selects the most suitable enhancement method for each dataset, ensuring optimal visual quality for downstream analysis and practical applications. All frameworks are validated through experiments on diverse real-world datasets, demonstrating robust performance, scalability, and applicability across different conditions and environments. The results of this research demonstrate significant improvements in both efficiency and quality. The video summarization framework enables rapid extraction of meaningful events from lengthy video streams, reducing the need for in situ observation and providing remote access to critical information. High-fidelity 3D reconstruction captures detailed scene structures in both terrestrial and underwater environments, supporting cultural heritage preservation and environmental monitoring. Adaptive image enhancement ensures that visual data from heterogeneous sources is suitable for downstream analysis, improving usability and reliability across applications. Collectively, these outcomes offer practical tools that enhance decisionmaking, operational efficiency, and accessibility. Beyond technical contributions, this thesis has substantial academic, societal, and environmental significance. It provides a foundation for future research in multimedia processing, computer vision, and related fields, offering novel methodologies and evaluation frameworks. By enabling remote observation and virtual exploration, the frameworks reduce environmental impact, expand accessibility for individuals 6 Computer Vision and Artificial Intelligence Methods for Video Analysis and 3D Scene Synthesis unable to attend sites in person, and promote the preservation of cultural and natural resources. Furthermore, they support safer operations in underwater and industrial environments and contribute to more efficient monitoring and management practices. Overall, this research demonstrates how advanced computational frameworks for video understanding, scene reconstruction, and image enhancement can bridge gaps between theory and practical, real-world applications