Computer Vision News - August 2021

Human behavior analysis from video data is one of the most complex challenges in the computer vision community as movements are difficult to define and lack clear semantic structures. As automatic human behavior understanding is a research topic that can potentially be used to support several fields of our society, during my PhD, with my promoter Stelios Asteriadis and co-promoter Mirela Popa, we focused on two research fields: Video Surveillance and Affective Computing. Dario Dotti completed his PhD in the Department of Data Science and Knowledge Engineering at Maastricht University (The Netherlands) in June 2021. His research focused on developing novel computer vision models for behavior understanding in different scenarios. His passion for this topic began in the framework of the European project ICT4Life, where, in a multidisciplinary team, they built smart services to support elderly living independently. Congrats, Doctor Dario! Automatic human behavior understanding for Video Surveillance. Human movements are generated in different forms and levels of complexity. Movements can be generated by the full body that moves coherently while performing an activity such as running or jumping, or movements can be seen solely as motion through space towards a destination. In the video surveillance context, we considered the latter definition of movements to build a hierarchical framework for modeling real-time motion trajectories (Figure 1). Figure 1. Overview of our trajectory prediction (colored in yellow) in the New York Grand Central dataset. Our framework embeds direction, orientation, as well as speed information from trajectory data (using our new feature representation on the right). Orientation and speed are critical behavior cues in surveillance scenarios. The hierarchical architecture of Autoencoders was designed to capture short spatio-temporal trajectory patches in the lower levels. Short motion patches are highly variant, containing short deviations and turns caused by moving obstacles all around. Then, in the higher levels, we combined motion patches 38 Congrats, Doctor!