Biomimetic approach to pedestrian recognition Andrej Lucny Faculty of Mathematics, Physics and Informatics, Comenius University Bratislava and MicroStep-Mis (http://www.microstep-mis.com/) We introduce our conception of pedestrian recognition. We consider this task as a complex problem and we propose how to put together various particular methods to support each other rather than looking for a best single method or a pipeline. In our opinion it is necessary to deal with all aspects which are parts of human sensation involved in pedestrian recognition. Thus our approach is bio-mimetic. In practice it means that we need to reconstruct whole 3D model of the scene rather than to highlight shape of pedestrian on 2D image or similar. We introduce so-called agent-space architecture which is able to capture such relations among processing modules like mutual cooperation, e.g. the fact that either we are able to recognize a pedestrian due to good segmentation results, either we can improve segmentation process according to successful recognition. Moreover some modules can work better but slowly, some quickly but worse; thus overall reaction can be fast and gradually improved. Finally some modules can compete which is useful when e.g. we need to follow several pedestrians on scene in parallel. Particularly we implemented reconstruction of 3D model containing pedestrian walking through zebra crossing from incomplete individual detections of figures or their parts like heads. This reconstruction was able to overcome such phenomena as occlusion. Anyway our approach is quite far from a product, it is rather idea how to treat this problem. personal web-page of author: www.microstep-mis.com/~andy Host of the seminar: Vasek Hlavac