Meir Cohen, Ph.D. Thesis Seminar
Wednesday, 10.9.2014, 12:00
It is quite common that multiple human observers attend to a single point of interest. Mutual awareness activity (MAWA) refers to the dynamic of this social phenomena. A peak of a MAWA is known as a mutual awareness event (MAWE) and can be interpreted as a "buzz" event, which draws the attention of many observers. A preferred way to monitor those social phenomenon is with a camera that captures the human observers while they observe the activity in the scene. Our work studies the underlying geometric constraints of MAWEs (joint work with Amit Adam) and the related dynamics of MAWAs. Those constraints are reformulated in terms of image measurements, which are collected using existing face detection and head pose estimation algorithms. Those constraints are then used in a method that (1) detects how many such points of interest exist if any, (2) determines where each point is located, (3) identifies which observer attends to what, (4) reports where and when each observer was while attending, and (5) tracks the above quantities over a long time in an on-line manner.
The suggested method is unsupervised and can deal with the general case of an uncalibrated camera in a general environment and an unconstrained activity in the scene. This is in contrast to other work on similar problems that inherently assume a known environment or a calibrated camera or a restricted occurrence in the scene.
In addition, the current work attaches a social semantics to the detected MAWA. A deeper social interpretation is suggested by exploiting and analyzing the spatiotemporal correlations between the MAWA and the activity in the scene, i.e. the dynamics of the events which occur in the visual field of view of the observers. The statistics of those social interpretation are aggregated over a long time yielding social characteristics of an individual observer, the entire group, and of the activity in the scene.
The method was tested on about 75 images from various scenes. In addition, the method was tested on videos of interesting activities, including: a single moving human observer that fixates on a single static interest point, a panel of several participants, and a classroom with many observers. The method robustly detected MAWEs and MAWAs, estimated their related attributes, and linked them with various social interpretations.