Advancing Multidimensional Video Analysis to Enhance Human Behavior Understanding and Steganalysis for Forensic Applications

Abstract

In this thesis, two interlinked areas of video analysis are addressed, both underpinned by computational methods yet distinct in application, in the realm of multifaceted video examination systems. The first advancement comes in the form of the OHON4D (Optimized Histogram of Oriented 4D Normals) descriptor, a geometrical method building on the foundational aspects of the HON4D (Histogram of Oriented 4D Normals) model for human activity recognition. This geometrically focused descriptor is particularly rare in computer vision and represents a critical evolution in the field, providing a streamlined solution to the recognition of human behavior in depth-captured videos. By honing in on the most essential four-dimensional projectors, OHON4D achieves the dual benefit of lowering the processing time while preserving the accuracy of the original HON4D descriptor, thus enhancing practical applications in video analysis.
In tandem with this, we address the topical shift seen in video steganalysis, with a particular focus on the High Efficiency Video Coding format (HEVC, also known as H.265). While the Advanced Video Coding format (AVC, also known as H.264) remains the prevalent format and has a mature body of steganalysis research owing to its widespread adoption in various multimedia applications, H.265 is gaining prominence. This shift comes as H.265 garners attention for its advanced compression efficacy, which also presents new challenges and opportunities in steganalysis. In response to these emerging needs, this thesis introduces an entropy-metadata-weighted feature analysis alongside machine learning classification, focusing on the forefront of H.265 steganalysis, and transcending the earlier focus on H.264. This model is not solely dependent on entropy values but integrates video metadata to dynamically tailor entropy coefficients, taking into account variables such as resolution, frame rate, and scene dynamics. Thus, this approach enhances the detection capability across various video contexts. Synthesizing the accelerated recognition capabilities of OHON4D and the precision of H.265 steganalysis, this thesis underlines a versatile and robust dual-layered system that could effectively serve to enrich forensic video analysis and surveillance measures. While the research encompassed here extends beyond the exclusive domain of forensics, the prospective forensic applications are evident, sketching out a new horizon where advanced video analysis intersects with evolving security demands in an age where video encoding standards continue to advance.

Date of Award	28 Nov 2024
Original language	English
Awarding Institution	Northumbria University
Supervisor	Ammar Belatreche (Supervisor) & Ahmed Bouridane (Supervisor)