Institut für Mess- und Regelungstechnik (MRT)

Probabilistic Models for Urban Intersection Understanding

 

Motivation

By today, modern driver assistance systems (DAS) are already supporting the driver in many tasks, such as lane keeping, collision avoidance, cruise control and blind spot assistance. However, the complexity of urban scenes still poses many challenges to a fully automatic understanding of the scene, which is essential to future safety systems and autonomous driving.

Scene understanding is an active research area in computer vision. The general task consists in classifying natural or urban scenes or inferring geometrical relations from images or video streams. The main focus of this work is on extracting topological and topographical information from stereo video streams.

The following figure depicts the left camera image of a typical urban scene, captured by the stereo rig mounted on top of our autonomous vehicle 'Annieway', which took part in the Darpa Urban Challenge 2007:

 

 

Interpreting the geometry of the scene is a hard task for several reasons: On one hand, buildings and vehicles are occluding the direct view onto the road. On the other hand, lane markings are often missing and the slanted installation angle of the cameras leads to high uncertainty in areas distant from the observer.

 

Stereo Matching

To infer the road and intersection topography, we therefore propose to combine multiple features in a probabilistic framework. High-resolution disparity maps are extracted for road plane estimation and for building a temporally local map:

 

 

 

Object Flow

We also compensate for the camera's movement by means of egomotion estimation and extract the 3-dimensional flow of dynamical objects on the road:

 

 

Probabilistic Model

The goal of this project is to infer the intersection geometry as exemplified in the following illustration, which depicts a T-type intersection:

 

To tackle the problem, several probabilistic models are developed. For the task of learning the models and for inferring the most likely crossroad configuration we compare deterministic with monte carlo methods.