Dr. Alexander Gutev is a Computer Vision and Machine Learning researcher and software development professional specializing in human action recognition, multimodal AI, and efficient video analytics. He holds a PhD in Communications and Computer Engineering from the University of Malta, where his research focused on computationally efficient human activity recognition using 3D video, motion saliency, and privacy-preserving depth-based representations. His expertise spans deep learning, computer vision, anomaly detection, multimodal fusion, time-series analysis, and real-time algorithm design, with strong proficiency in Python, C/C++, PyTorch, OpenCV, and related technologies. Dr. Gutev has published research in IEEE conferences and journals and has presented his work internationally. He currently works as a Project Manager at Viditrack Ltd., leading multidisciplinary software development projects and combining technical expertise with project and product leadership.

Title: Energy-Efficient Computer Vision for AI-Assisted Living Applications                                                                                                                                                                                                                             Artificial intelligence and computer vision are increasingly enabling intelligent systems capable of understanding human activities and supporting applications in healthcare, assisted living, and smart environments. However, the growing computational requirements of modern AI models raise important challenges related to processing efficiency, energy consumption, and sustainability. This presentation explores the use of efficient computer vision and machine learning techniques for human activity recognition, with a particular focus on applications in assisted living and privacy sensitive environments. The work presented investigates human activity recognition using 3D video data, including depth and skeleton-based representations, to identify relevant human movements while reducing the volume of data processed by computationally intensive recognition models. A motion saliency-based approach is introduced to focus processing on the most relevant regions and temporal information within video sequences. This enables the development of more computationally efficient activity recognition systems while maintaining recognition performance. Multimodal approaches combining complementary 3D video representations are also explored to improve robustness and efficiency. The proposed approaches have potential applications in assisted living environments, where continuous and unobtrusive monitoring of human activities could support the detection of daily activities and potentially assist in identifying unusual or potentially concerning events. The use of depth-based information also provides an opportunity to address privacy concerns associated with conventional RGB video surveillance by reducing reliance on identifiable visual information. In addition to the healthcare and assisted living perspective, the presentation considers the role of Green AI and sustainable machine learning. By reducing the amount of data presented to computationally intensive models, the proposed methods achieve substantial reductions in training and inference time, with experiments demonstrating approximately four- to five-fold reductions in computational processing requirements and significant reductions in energy consumption during model training. These findings highlight how algorithmic and data-processing strategies can contribute to more sustainable AI systems without necessarily requiring reductions in recognition capability. The presentation concludes by discussing the broader potential of efficient computer vision for developing sustainable, privacy-aware AI technologies for healthcare, assisted living, and other real-world applications where continuous intelligent monitoring is required.