Abstract

At the CMP (Center for Machine Perception), part of the Czech Technical University in Prague, a project has been running on crowd surveillance. The goal of the project is the detection and tracking of people. We tried to improve on the earlier results of that project, by making the existing system better usable for tracking of multiple people in a group simultaneously. Tracking multiple people in groups can be done by using multiple single-person trackers at the same time. However, often these systems are not meant for multimodal applications, so the results are not satisfactory. Another problem is, that using multiple single-person trackers simultaneously does not work in the case of occlusions. We have augmented the single-person tracking methods available (Condensation and Mean Shift) with capabilities for occlusion detection. The resulting implementation did not have to work in real-time. For test data material from CMP was used, together with some self-made animated videos. We designed and implemented a demonstration and have tested the system. The model, implementation and test results, will be described in this thesis.