Computational Verification Methods for Automotive Safety Systems
This thesis considers computational methods for analysis and verification of the class of automotive safety systems which support the driver by monitoring the vehicle and its surroundings, identifying hazardous situations and actively intervening to prevent or mitigate consequences of accidents. Verification of these systems poses a major challenge, since system decisions are based on remote sensing of the surrounding environment and incorrect decisions are only rarely accepted by the driver. Thus, the system must make correct decisions, in a wide variety of traffic scenarios. There are two main contributions of this thesis. First, theoretical analysis and verification methods are presented which investigate in what scenarios, and for what sensor errors, the absence of incorrect system decisions may be guaranteed. Furthermore, methods are proposed for analyzing the frequency of incorrect decisions, including the sensitivity to sensor errors, using experimental data. The second major contribution is a novel computational framework for determining the errors of mobile computer vision systems, which is one of the most widely used sensor technologies in automotive safety systems. Augmented photo-realistic images, generated by rendering virtual objects onto a real image background, are used as input to the computer vision system to be tested. Since the objects are virtual, ground truth is readily available and varying the image content by adding different virtual objects is straightforward, making the proposed framework flexible and efficient. The framework is used for both performance evaluation and for training object classifiers.