Videobasiertes Multi-Personentracking in komplexen Innenräumen

Video-based detection and tracking of persons in different environments gets increasingly more attention in computer vision research. Currently, the most prominent application for people tracking systems is automatic video surveillance, other usages are motion capturing and man-machine interaction. The majority of existing tracking systems are created for the surveillance of wide areas and work with high and distant camera positions. Narrow indoor environments, however, require cameras that are low and relatively close to the observed persons. Existing methods for such environments use either multiple cameras with overlapping fields of view or stereo cameras. This thesis presents the development of a people tracking system that extracts the floor trajectories of multiple persons in a narrow, cluttered indoor environment from the image sequence of a single monocular colour camera. Additional cameras with overlapping or non-overlapping fields of view can be easily added thanks to the modular structure of the system. The background of this work is a project with the goal of detecting the positions and trajectories of all passengers in an aircraft cabin. Difficulties in those environments are partial occlusions of persons by scene objects or the frequent overlapping of two or more persons in the image plane. The system also has to deal with dynamic illumination changes and sitting persons in the field of view. To cope with these requirements, novel approaches for the individual processing steps were developed that use extensive knowledge about the current scene status and additional a-priori knowledge about the monitored environment. A three-dimensional model of the scene is used to handle the occlusion by scene objects. The first processing step is image segmentation, which classifies each pixel either as background scene, a specific person or unknown foreground object. To this end, an illumination adaptive background model and colour distributions of the tracked persons are used in combination with a prediction of their current positions. The person colour models consist of histogram-based representations of multiple colour clusters, which can be translated and scaled in colour space while keeping the general colour topology. So, each colour model can be flexibly adapted to varying illumination in time or space. Based on this, a method for illumination robust person identification by the colour appearance was developed. To calculate the position of a person, first the complete silhouette is reconstructed by aligning a model of the human silhouette to the visible part of the person in the image. After that, coordinate transformation is used to calculate position candidates from the detected head and feet coordinates, which are weighted according to their reliability resulting from the occlusion and camera perspective. Additionally, the valid depth interval of the person is calculated from the occlusion situation. Another step builds the filtered trajectory from the data of one or more cameras. The tracking system works in real-time and was evaluated in multiple different scenes (office environments and aircraft cabin mock-ups) with 1 to 3 cameras and up to 4 persons. The result of the visual evaluation yields an average percentage of correct positions of 96%