Supervision of dynamic urban scenes is a crucial task in traffic control, environmental preservation, crime prevention and traffic accident prevention. Automatic analysis of conventional imagery data is challenging because of the limited field of view of the street surveillance cameras, the high variety of the objects, the multiple occlusions and the quality variances caused by the light and weather condition changes.
A LIDAR sensor mounted on a moving car provides a data stream in which every frame is a three dimensional point cloud with a constantly changing coordinate system that moves along with the car. From this data, we can easily extract accurate 4D (space & time) geometrical information about the scene but the analysis and interpretation of this sparse data needs more complex approach than a static LIDAR measurement or a conventional 2D image sequence processing. In this thesis, the student’s task is to develop algorithms to analyze point cloud data stream. These algorithms should help the later development of complex scene interpretation functions.
In the thesis the tasks will be: to introduce the LIDAR sensing technology and to research state of the art accomplishments in the field of analysis of ground LIDAR data. The student will be required to implement LIDAR sensory data segmentation methods (the task is to identify semantic classes in point clouds such as vegetation, road, walls, objects on the streets) and to implement an automatic procedure to register separate clouds in the data stream against each other. Also, he should investigate the possible usages of the intensity channel for recognition tasks and he should analyze the efficiency of the implemented tasks and suggest possible improvements on his methods.