3D Object Recognition in LIDAR Point Cloud Sequences

OData support
Dr. Szécsi László
Department of Control Engineering and Information Technology

Today’s people have high expectations for technological and technical improvements that can be used for making everyday life easier, saving time or increasing our sense of security. Think of the increase of the reliability of safety systems, the high degree of computerization of transport, or the development of medical imaging systems. In these fields there are a number of algorithms that are based on computerized imaging techniques. Some of these algorithms implement 2D image processing while others process 3D measurements. Take the LIDAR technology, one of the well-known three-dimensional devices.

The thesis provides an overview of the importance of the LIDAR technology and also of the steps of the processing of point clouds. It deals with the separation of static and moving objects and with the importance of its methods in the street. It gives special emphasis to the separation of shapes as it has a key role in the area of security systems as well as in urban modelling. It gives an overview of some of the implemented procedures such as the separation of columns, the segmentation of billboards, the extraction of pedestrians, which are mainly based on geometric properties. It sums up some scientific results in connection with ground measurements, and gives a detailed introduction to a system that consists of two parts.

One of these parts, which has most of its functions already been put into practice, is able to read and label a point cloud sequence, and the result can be saved and read again. I will discuss the new function that during annotation it will be possible to classify the different shapes and put them in a training set. The other part is a new frame system into which the point cloud used in the above mentioned system and the annotation file will be loaded, and with their help I will define statistical data. By using the data defined in the process I will classify the shapes that we did not put into the training set. My aim is to study how good the statistical methods prove to be in the process. Their applicability depends on the number of objects, the degree of identity of similar objects, and also the number of classified sets, so they can be used in different ways in different environments.


Please sign in to download the files of this thesis.