The benefits of a large database are twofold. First, it provides a wealth of training data for statistical pattern matching techniques. These «learn» the VRU appearance from examples; which is important, since good prior, explicit models are hard to define. Second, it allows to evaluate system performance on a truly large dataset, so that the results can be considered representative of the true physical traffic situation.

Establishing “Ground Truth”

In order to train statistical pattern matching methods, or to measure system performance during testing, we need to know the “true” position and spatial extent of the VRU in the images of the database, i.e. the “ground truth”. To establish this, a semiautomatic labeling-tool called VisiCurve has been developed by DaimlerChrysler, which assists the user in outlining the VRU object contours in images and in establishing temporal correspondence across the images of a sequence.

The main aim of VisiCurve is to avoid the need for a user having to specify pixel-by-pixel the entire object contour. This “manual” contour labeling is not only very time consuming but also a very tedious. Computer vision techniques for segmentation can in principle support the labeling process, since they are designed to find object boundaries automatically. They usually utilize a certain degree of prior knowledge on object appearance and lock on low level image features. The algorithms typically involve an object model, whose parameters can be restricted by the prior knowledge to certain ranges of valid object configurations.
In order to get a head start, simultaneous with the development of VisiCurve, DaimlerChrysler, had the VRU object contours in IR dataset labeled “manually” (1039 images, 58 sequences).