Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Pedestrian behavior analysis in crowds using image-based methods

: Pathan, Saira; Richter, Klaus


Chraibi, M.:
Traffic and Granular Flow '13 : Tenth edition, the international conference "Traffic and Granular Flow" (TGF), 27 September 2013, Jülich
Cham: Springer International Publishing, 2015
ISBN: 978-3-319-10628-1 (Print)
ISBN: 978-3-319-10629-8 (Online)
ISBN: 3-319-10628-7
International Conference "Traffic and Granular Flow" (TGF) <10, 2013, Jülich>
Fraunhofer IFF ()

In this paper, we aim to investigate the image-based approaches and propose a framework to examine the pedestrian flow in crowds on various real situations. This research is inclined on two main aspects: first, an in-depth analysis of image-based approaches is given particularly for the situation containing large number of pedestrians (i.e., crowds) and second, our proposed approach which mainly focuses on computing the flow data, modeling, and classifying the corresponding behaviors of pedestrians in a crowd. The dynamic data of underlying crowd scenes establish a large cloud of information (i.e., correlated or un-correlated data). Therefore, it is essential to extract the meaningful information from the data cloud however the selection of criteria is a crucial task which is answered in the first part of the paper. Moreover, in crowded scenes, it is challenging to extract individual characteristics (e.g., head, torso, or leg count) of every pedestrian forming the crowd. Because, the pedestrians do not own these characteristics while moving in the form of groups. Therefore, we can not rely on such individual information of every pedestrian for longer time instances. Based on this fact, in this research, we measure the dynamic contents over consecutive frames. After this, we model this information by computing the Histogram of Flow (HOF) for each time instance. Later, we classify these HOF features according to our behavior-specific classes. We have tested the proposed approach on the dataset recorded with the help of approximately 30 volunteers. In the context of pedestrian behaviors characterization, we have employed Support Vector Machines on our recorded dataset and achieved 91 % classification rate.