Alexander
Gutev

Understanding Activity in Private and Public Setups using 3D Video Content

My doctoral research investigated machine learning approaches to human action recognition, with a dual focus on efficiency and privacy. The first strand developed a novel motion saliency detection technique that identifies the most relevant spatial regions within video sequences, allowing action classification models to operate on a reduced subset of the data. This yielded substantial gains in both training and inference speed, alongside a measurable reduction in energy consumption — all without sacrificing classification accuracy. The second strand addressed deployment in privacy-sensitive environments where conventional colour video cannot be used. Building on the saliency-based methods established earlier, a multimodal framework was devised that draws exclusively on depth data and skeleton joint information. This approach not only preserved the efficiency improvements of the first phase but achieved classification accuracy surpassing existing state-of-the-art methods. The research produced five peer-reviewed publications across IEEE venues.