Kristin Schönherr

Evaluation of pose hypotheses by image feature extraction for vehicle localization

For the realization of driving assistance and safety systems vehicles are being increasingly equipped with sensors. As these sensors contribute a lot to the cost of the whole package and at the same time consume some space, car manufacturers try to integrate applications that make use of already integrated sensors. This leads to the fact that each sensor has to fulfil several functions at once and to deliver information to different applications. When estimating very precise positioning information of a vehicle existing sensors have to be combined in an appropriate way to avoid the integration of additional sensors into the vehicle. The GPS receiver, which is coupled with the navigation assistant of the vehicle, delivers a rough positioning information, which has to be improved using already available information from other built in sensors. The approach discussed in this paper uses a model-based method to compare building models obtained from maps with video image information. We will examine, if the explorative coupling of sensors can deliver an appropriate evaluation criteria for positioning hypotheses.

Robust tracking and monocular 3D reconstruction of line-based features in urban environments

Driving assistance and safety systems are based on data processing of different sensor information. The environment of a vehicle is detected by these sensing elements for accident avoidance or reduction of the accident severity. Due to package size and cost reasons, only selective sensors are used for standard-production of a vehicle. These integrated sensors have to fulfill several tasks at once and to serve different applications. Especially the processed data of a grayscale camera is used for parking assistant or lane detection system. But known from computer vision, a three-dimensional reconstruction of feature elements is possible by analyzing image sequences. This research considers corner detection to create point clouds in space. But the dominant elements of an urban environment, more precisely buildings, are characterized by edges. With this paper we focus on the robust tracking and three-dimensional reconstruction of line-based features. In this context we present our algorithm, defining and analyzing edge tolerances, which is called Tube Principle. The accuracy of this approach is determined by comparing the three-dimensional lines with landmarks of high-precise maps.