Teilautomatisierte Interpretation von Straßenverkehrsszenen bei Einsatz von Schwenk-Neige-Kameras

To increase safety on roads, an increasing number of cameras are being used to observe the traffic situation. For economic reasons, rotatable cameras are mostly used. An automation of the incident detection by means of video detection systems (VDS) would be a clear relief for the operation. In addition, VDS could be used to gain spatial traffic-data, which is not possible by conventional detection systems. Because the techniques of currently-used VDS are designed for static perspectives and constant lighting conditions and are therefore mostly based on a reference picture, they cannot be used with rotatable cameras and changing environmental conditions. Therefore, for VDS which are to be used with changing perspectives and all weather conditions, procedures which are not based on fixed reference pictures have to be used. With a view to these conditions, a procedure to make a situation analysis of traffic scenes for rotating cameras possible without using reference pictures was developed. The development of the two main tasks of the model - the selection of suitable procedures as well as the linking of extractable characteristics with defined previous knowledge-took human intelligence into consideration. In order to make a reliable object recognition possible, procedures for the extraction of characteristics had to be worked out. The usability of several characteristics of a digitized picture were examined in order to identify and describe an object clearly. Here the development of a concept for the extraction of object outlines was a special emphasis. An accurate determination of object positions and an interpretation of object movements can only take place using a known adjustment of the camera (direction, slope angle, focal length). Three procedures which derived the camera orientation from the information of the existing picture sequence and the given knowledge put in a scene model were examined in detail. For object tracking the Condensation algorithm was executed, which achieves reliable results even under difficult conditions. A test software was developed for the evaluation of the presented models and procedures, in which the orientation regulation, the extraction of characteristics as well as object recognition and tracking are implemented. After creating a knowledge base, the models were validated with detailed analyses of lifelike traffic scenes. At the end of the text, the results of this work are summarized, and a view for future research fields within the range of video detection is given