Abstract. Stereo and motion analysis are potential techniques for providing information for control or assistance systems in various robotics or driver assistance applications. This paper evaluates the performance of several stereo and motion algorithms over a long synthetic sequence (100 stereo pairs). Such an evaluation of low-level computer vision algorithms is necessary, as moving platforms are being used for image analysis in a wide area of applications. In this paper algorithms are evaluated with respect to robustness by modifying the test sequence with various types of realistic noise. The novelty of this paper is comparing top performing algorithms on a long sequence of images, taken from a moving platform. 1

Abstract. Prediction errors are commonly used when analyzing the performance of a multi-camera stereo system using at least three cameras. This paper discusses this methodology for performance evaluation on long stereo sequences (in the context of vision-based driver assistance systems). Three cameras are calibrated in an ego-vehicle, and prediction error analysis is performed on recorded stereo sequences. They are evaluated using various common stereo matching algorithms, such as belief propagation, dynamic programming, semi-global matching, or graph cut. This performance evaluation is demonstrated on synthetic and real data. 1

To extract three dimensional data from a pair of images, it is essential to solve the correspondence problem. In the literature, a large number of algorithms have been implemented which differ in the token type, match constraints and search methods employed. Recently, hierarchical matching schemes have utilised multiple token types of increasing complexity. In previous reviews of stereopsis, no general framework has emerged within which to evaluate all the different contributions. This paper breaks down the correspondence problem into its general components: token type, match constraints and the method employed to encode and search match information. In common with other reported work, matching is cast as an optimisation problem, and the definition of match functionals may be separated from the method employed to search the solution space. Within this very general framework, hierarchical matching is discussed at some length including suggestions on how hierarchical constraints may be f...

Abstract. Applying computer technology, such as computer vision in driver assistance, implies that processes and data are modeled as being discretized rather than being continuous. The area of stereo vision provides various examples how concepts known in discrete mathematics (e.g., pixel adjacency graphs, belief propagation, dynamic programming, max-flow/min-cut, or digital straight lines) are applied when aiming for efficient and accurate pixel correspondence solutions. The paper reviews such developments for a reader in discrete mathematics who is interested in applied research (in particular, in vision-based driver assistance). As a second subject, the paper also discusses lane detection and tracking, which is a particular task in driver assistance; recently the Euclidean distance transform proved to be a very appropriate tool for obtaining a fairly robust solution.

Abstract. This report describes an efficient implementation of dynamic programming and belief propagation algorithms on a cell processor that may be used to speed up stereo image analysis. Dynamic programming is a method for efficiently solving optimization problems by caching subproblem solutions rather than recomputing them again. It processes image data by scanline optimization; thus it is easily implemented on a cell processor; our results show that this algorithm runs very efficiently on the cell processor. Belief propagation differs from dynamic programming by having potentially the whole image area as an area of influence for every pixel; this global optimization scheme produces improved results, but requires more run time than the dynamic programming method on a normal PC. Besides the tests on synthetic data, we use real-world image sequences captured by a test vehicle (HAKA1); they are typically degraded by various types of noise, changes in lighting, differing exposures, and so on. We use two methods to process the original images: Sobel edge detection and residual image analysis. Our results show that a cell processor also reduces running time for these processes. Sobel and residual images can improve the stereo matching result compared to the use of original real-world images, however, due to the used block structure and cell architecture limitations, the accuracy is also degraded slightly. 1

The paper discusses various options for testing correspondence algorithms in stereo or motion analysis, designed or considered for vision-based driver assistance. The main focus is on testing on video sequences of real-world data. The authors suggest the classification of recorded video data into situations, defined by a co-occurrence of some events in recorded traffic scenes. About 100 to 200 frames (or 4 to 8 seconds of recording) are considered to be a basic sequence, to be identified with one particular situation. Future testing is expected to be on data that is reporting on hours of driving; multiple hours long video data may be segmented into basic sequences, and classified into situations. The paper prepares for this expected development. The paper explains the use of currently already available EISATS test data, aiming for such “wide-angle ” tests. The paper uses three different evaluation approaches for demonstrating (by means of EISATS examples) ideas, difficulties, and possible ways in this future field of extensive performance tests in vision-based driver assistance. The paper also contributes with proposals of evaluation techniques, especially in the case of lack of ground truth. The study shows that the complexity of real-world data often does not support an identification of absolute rankings of correspondence techniques; this is true already for a small set of selected situations. It is suggested that correspondence techniques need to be adaptively chosen in real time using fast situation classifiers (e.g., based on a few feature distributions).