Crowd Analysis

Crowds are a main source of events for SMART. At a first level of analysis, we are interested in their static and dynamic analysis. From an algorithmic point of view, static analysis is based on dynamic foreground/background segmentation:

On the other hand, dynamic analysis is based on motion vector extraction.

Static analysis of crowds

In SMART, following data protection principles, crowds are analysed as groups of foreground pixels derived by an adaptive foreground segmentation algorithm featuring spatio-temporal adaptation of its learning rate. From the foreground pixels we determine:

• Crowd density, i.e. a measure of crowd formation. It is derived by processing the foreground pixels in every frame.
• Prominent colours in crowds, that are of interest from a fashion point of view.
• Change detection, i.e. identification of medium-term changes in the background. Such changes are generated by parking or unparking vehicles, by left objects, changes in advertisment boards, etc.

Dynamic analysis of crowds

This analysis yields crowd behavior. We detect the following behaviors:

• Conformance crowd motion to some prminent direction of flow.
• Concentrating towards a point.
• Dispersing away from a point.

Demonstration

Analysis of crowds in different zones: the streed (top) and the plaza (bottom) based on the SMART November 2012 crowd analysis system. Adaptive foreground segmentation yields foreground blobs (shown in colour over the grayscale background). The blobs are analysed to yield crowd density (red and green horizontal bars for the different zones) and prominent crowd colours (coloured rectangles). Motion analyis detects conformance to interesting directions of motion (shown in red when significant).

Traffic Analysis

Traffic is analysed either as a crowd, only from the city traffic cameras, or as individual cars, using trackers.