Introduction

The first step to testing scenarios generation is to construct the road networks where all the traffic participants (e.g. EGO vehicle, NPC vehicles, pedestrians etc.) move.

In city-driving scenarios, the two most common elements of the road networks are the roads and junctions (including roundabouts).

The dynamic driving tasks of the EGO vehicle usually involves starting from any road side, driving through various junctions and reaching the destination at another road side. Such a route can be decomposed into a sequence of unit routes, each including a start point on one road, one junction to drive through and an end point on another road (i.e. the other side of the junciton).

Once we have the definition of unit routes, we can then extract all the possible routes given a city-driving map and treat each route as a testing scenario.

But here comes the issue of redundancy.

For example, for the San Francisco map provided by LG SVL, we noticed that over 90% of the routes extracted have similar roads and junction structures. This leads to large amount of duplicated test cases, thus waste of testing efforts.

To solve this, a way to classify routes need to be proposed to quantitatively group test cases and identify representives test cases from each group.