Section: New Results

Crowd Data Collection from Video Recordings

Participants : Jihed Joobeur, François Brémond.

The aim of this work is to analyze crowd behaviors by detecting specific situations : panic, congestion, fighting etc. We validate our work with subway station videos from VANAHEIM project. We use Mixture of Gaussian based segmentation to extract moving point and then detecting moving objects. Subsequently inside these moving objects we detect FAST feature points and compute HOG descriptors for tracking these points. We compute different features based on these points like speed and orientation. To estimate the crowd density we use features based on Grey-Level Co-occurrence Matrix. As these features depend on the distance of people from the camera, we divide the scene into different zones which have each zone same distance from the camera. In each area, compiling all the information on speed, direction and learned over a threshold density of the crowd, we can learn and detect different situations. For example, if the density increases and the average speed decreases in a pre-defined zone, that may correspond to a congestion situation.

On figure 17 the FAST feature points are shown in blue points, while the tracking of these points is shown in yellow.

Figure 17. Feature points detection and tracking in different zones.