Automatic crowd density and motion analysis in airborne image sequences based on a probabilistic framework

Real-time monitoring of crowded regions has crucial importance to avoid overload of people in certain areas. Understanding behavioral dynamics of large people groups can also help to estimate future status of underground passages, public areas, or streets. In order to bring an automated solution to the problem, we propose a novel approach using airborne image sequences. Our approach depends on extraction of local features from invariant color components of the images. Using extracted local features as observations, we form probability density functions (pdf) for each image of input sequence which holds information about density of people. We introduce four measures to extract information about pdf characteristics. A change within the four measures over the image sequence gives important information about status of the crowds. Besides, we also use obtained pdfs to estimate main crowd motion directions. To test our algorithm, we use a stadium entrance image data set, and two festival area data sets taken from an airborne camera system. In order to be later able to reach real-time performance the algorithms use parameters which can be extracted directly from the image data. Experimental results indicate possible usage of the developed algorithms in real-life events.