Stefan Lueke

Towards a generic and efficient environment model for ADAS

In research projects for future ADAS functions a dense environment model covering free space is often necessary, which is obtained by complementing or replacing a common object list by a grid based environment model. The drawbacks of grid based models are their demands for memory, computational resources and bandwidth. This paper analyzes the influence of data compression on accuracy and resource demand of a grid. By using a simple compression scheme the transmission bandwidth and the required computational resources can be significantly reduced.

Evaluation method and results for the accuracy of an automotive occupancy grid

This paper presents an evaluation method for the accuracy of automotive occupancy grids and results for the influence of the discretization and pose estimation of a radar based grid mapping algorithm. An automotive centric review of evaluation methods and map quality measures developed for robotic applications is given. Based on the results of the review, an extensible toolset to create ground truth maps and to compare them against automotive grid maps using different map quality measurements is proposed. Several map quality measures are compared and the best performing method to evaluate the accuracy of a radar based occupancy grid mapping algorithm is chosen.

Generic hypothesis generation for small and distant objects

Existing approaches to object detection address the generation of object hypotheses by extracting several cues in natural and automotive images, relying on objects with sufficiently high resolution. Very little to almost no approaches, however, address the generation of hypothesis of very small or distant objects in images such as on motorways. Here, we propose a simple yet effective approach to generating hypotheses of small and distant objects in images. Our key contribution is a novel voting scheme that makes efficient use of the different appearance of small candidate objects to their environment. We model the environment as being composed of very few regions with homogeneous appearance, extracted by evaluating the inner statistics of an image in an unsupervised fashion. Small regions that can not be assigned to the environment form potential candidate locations. Experimental results on motorway scenes with cars, traffic signs, and other automotive objects based on a variety of performance evaluation metrics show that our approach provides promising results, and outperforms one of the currently leading approaches in generating hypotheses for small and/or distant objects in images.

Realtime Roadboundary Detection for Urban Areas

In the context of driver assistance systems lane detection systems are used for lane departure warning and lane centering functions. In some cases, especially in urban scenarios, no lane marks are available and for some circumstances lane marks are not sufficient to assist the driver. For novel urban functions it is necessary to know in which areas the car can operate. This problem requires detecting the free space as a traversable area for vehicles and has to take into account small objects such as curbstones. In this paper we present a real-time road boundary detection for urban areas to realize an easy lane recognition algorithm by using a series hardware stereo camera system. The algorithm is based on dense stereo information.