When 6D-Vision detects a moving object, the question arises whether it is a pedestrian, for whom we might not only brake but also consider a swerve maneuver in the future.
Fur this purpose, we use so-called classifiers that are trained with many pedestrian examples. This technique is applied successfully e.g. in traffic sign recognition of modern vehicles. However, traffic signs vary moderately in size and appearance whereas pedestrian recognition has many challenges:
every pedestrian wears different clothes, hence appears different to a classifier
depending on motion and view direction the appearance changes
in urban traffic often pedestrians are partially occluded by other obstacles, which makes it only partially visible to the system.
6D-Vision can support this process. Members of the Daimler research group showed that considering the stereo information improves the performance of the state-of-the-art classifers by a factor of 5 (!) . Intuitively this can be explained looking at the two images above. While the silhouette in the grayvalue image on the left is difficult to see, one can easily spot the silhouette in the distance image on the right (in false color-coding).
The motion information delivered by Dense6D can even boost this performance. Currently, we are working on the answer to the question: "Will the pedestrian cross - or will he stop?" , hence we try to predict pedestrian behavior. A publication of us with this topic was recently recognized with a best paper award .