Recently, researchers in video-surveillance shifted the attention from the monitoring of a single person in a camera-monitored environment to that of groups and their behavior. This novel level of abstraction provides event descriptions which are semantically more meaningful, highlighting barely visible relational connections among people.
Automatic Crowd Analysis is a research area which can be used for anomaly detection: panic scenarios, dangerous situations, illegal behaviors, etc. We are currently working on anomaly detection in crowded scenarios using a particle-based paradigm, in which a large set of virtual particles simulates the crowd behavior using visual cues such as the optical flow.
Following the paradigm of people detection and tracking, small group analysis is split in group detection and tracking.
The objective of group detection is to find collection of people who share certain aspects, interact with one another, accept rights and obligations as members of the group and share a common identity. To this end, we aim to investigate novel models and technologies that embeds notions of social psychology into computer vision techniques, offering a novel research perspective for the video surveillance community.
On the other hand, group tracking consists in following tight formations of individuals while they are walking or interacting. One of the major difficulties of group tracking lies in the high variability of the group entity: splitting, merging, initialization and deletion are frequent events that characterize the life of a group, and that are usually modeled by heuristic rules, yielding to a scarce generalization. Our idea is to perform, at the same time, tracking of individuals and groups. The problem, dubbed as joint individual-group tracking, introduces novel issues that need to be investigated (e.g.: high-dimentional spaces, and highly non-linear dynamics).