Abstract: In this work, we propose a probabilistic method for detecting and classifying urban road intersections from a moving vehicle. In order to evaluate our approach we demonstrate its capabilities in intersection topology classification with respect to the 7 intersection patterns shown in the figure below. We manually annotated the following KITTI sequences to create the intersection topology ground truth, associating each frame of each sequence to the corresponding topology. We also introduced a flag (referred to as the "crossing" flag), which indicates whether the vehicle is inside an intersection.

0	1	2	3	4	5	6

Road surface dataset
The road surface training dataset is based on the work of the Torr Vision Group, which can be found at the following link: Urban 3D Semantic Modelling Using Stereo Vision.

We labelled again their dataset using the following 3 classes only:

• road
• sidewalk / pavement
• other, i.e., everything else

Moreover, we additionally labelled 10 images from the original KITTI sequences. The new images contain only intersection areas. Thus, the resulting dataset consists of the original 70 images (25 from the training set + 45 from their test set) plus 10 images.

The resulting dataset can be downloaded from the following  iralab-intersection-training.zip (740 downloads )

We tried also to match the original TORR framenames with the corresponding KITTI dataset. The matching is included in the aforementioned link and in the following table:

The following are the 10 new images we inserted in our training set.

0000000106.png
0000000187.png
0000000368.png
0000002595.png
0000002854.png
0000003381.png
0000003514.png
0000003937.png
0000004121.png
0000004407.png