Ein generisches System zur automatischen Detektion, Verfolgung und Wiedererkennung von Personen in Videodaten

An important area in computer vision is the person-centered video analysis. Applications cover many areas of today's life like driver assistance, human-machine-interaction, threat assessment in military context and specifically visual surveillance. The basis of this person-centered analysis is person detection and tracking in video data. This is a precondition for all subsequent analysis or interpretation approaches. Moreover, person reidentification is a substantial component of many applications. Such a reidentification of persons is necessary in cases where a long time period or a large spatial area is considered. In these cases, connections between the occurrences of people that are not directly temporally or spatially connected are to be established. A typical example of this is the surveillance of large public spaces like airports where multiple networked cameras are utilised and a long time period is relevant. Due to the diversity of application areas for person detection, tracking, and reidentification, it is desirable to develop a generic system that is most independent of certain aspects of application scenarios and thus universally applicable. In this work, such a system for person detection, tracking and reidentification is introduced. This system is generic regarding different aspects. The system is independent of the application scenario, meaning that no assumptions on the application environment are made. For instance, it is not assumed that the scene background is known or other information regarding the scene is available. It is also not assumed that the recording sensor is stationary, which means the system introduced in this work is applicable in the case of a moving camera. Equally, the system is not limited to certain object classes since no object class specific knowledge other than a set of training samples is used. In addition, the system is mostly independent of the used sensor since no other than the intensity-gradient based local features are used. Thus, the overall system is applicable in the visible and the infrared spectral range since no features like color or depth are employed. The system generality is specifically accomplished by the exclusive use of the Implicit Shape Model approach and local image features for all three system levels, whereby the levels are closely connected and merge in an integrated approach. For person tracking, an extension of the Implicit Shape Model, which combines bottom-up tracking-by-detection w