Daniel Streller

Lane detection and street type classification using laser range images

This paper describes a method for the lane detection and street type classification from range images generated by a single line laser range finder mounted at the front of a car. Additionally, an environment reconstruction is done. Objects like cars and cycles are classified and tracked. For lane detection, they are removed from the range image in order to reduce disturbances to the algorithm. The lane is described by its type, width, curvature, and relative position to the car. This information can be used for driver assistance systems like lane departure warning.

Vehicle and object models for robust tracking in traffic scenes using laser range images

Detection and modeling of dynamic traffic scenes around a, driving passenger car is the long-term aim of the research project ARGOS at the University of Ulm. Each object close to the own car should be modeled and tracked using a specific individual dynamic model. The object classification is based on the geometric outlines and the dynamic behavior. For any sensor combinations usable to detect the environment, the velocity of the objects can be measured relatively to the movement of own vehicle. To. get the absolute velocity of the objects, the motion of the own vehicle must be measured for which the well know bicycle model is used. This ego-model is fed by sensor signals provided anyway by ABS, ASR or ESP. To eliminate the own motion from the object measurements, several coordinate transformations are required in the different stages of data processing. A proposal is given on how to solve this problem when using a laser range finder as a sensing device. Moreover, a simple object model is introduced for this task in order to save processing power. The algorithms can extended towards a multihypothesis approach which will result a more robust classification and tracking algorithm.

Multiple hypothesis classification with laser range finders

An algorithm for tracking and classifying objects in urban areas using a multi-layer laser range finder is presented. Due to object disintegration caused by occlusions and fine segmentation of the data in order to separate close objects, classification is not always simple. Therefore, a multiple hypothesis approach is proposed, which keeps track of all feasible combinations of segments. The algorithm takes all segment combinations to create hypotheses and these are tracked over time using a Kalman filter. Due to the large number of hypotheses, restrictions are applied to reduce the number of hypotheses. Since applications need a description of the environment, which is described by objects, hypotheses are selected by their qualities and are provided as objects.