Urban Mobility Asisst - Deriving Typical Driver Behavior in Urban Scenarios

[Overview | Goals | People | Contact]

“It always seems impossible until it´s done.” -Nelson Mandela

Overview

Cars with some degree of Self-Driving features are not anymore any science-fiction TV show or movie, but rather a reality. Furthermore, it is a current trend with industry and academia joining efforts towards making cars completely autonomous.(BI Intelligence Estimates, 2015)

The promise of saving 1000s of lives from car accidents by relying on Self-driving Cars have motivated different efforts to realize this vision. Some of the efforts have shown some degree of success. However, after Google incidents and particularly, Tesla first fatal crash on May 7th 2016, it is clear that a solution in complex urban environments is still beyond the reach of current state of the art approaches.

Project Goals

What is missing?

Our hypothesis is that the feeling that resembles what a human would typically do in a particular situation is still absent in current research efforts. Indeed, complex urban driving situations will require a self-driving car to be able to represent, understand and reason about so many different situations that today only very well experienced human drivers can do. Thus, it seems natural to explore the idea of trying to distinguish between typical vs untypical driving behavior from human drivers towards the goal of transferring this expertise to self-driving cars. In this project, to assess the feasibility of this vision, we start with a use case of a complex situation: an intersection. To find Typical Driving Behavior, we design a model to capture a very useful idea from psychology and cognitive science “Typicality”. Typicality of objects has been discussed in cognitive science for many years [4]. The intuition behind it is simple but powerful: people think of some objects to be “better examples” of a concept than others. In computer science, typicality has been successfully modeled and implemented in a variety of complex tasks and in different ways: in entity centric query search [1], for ranking attributes of entities[2], to answer top-k queries in databases [3], among others.

In this project with funding from Volkswagen, our main research question is whether from large quantities of the behavior of drivers, despite all the individualities, typical dominant tactics for the same strategy with sufficient accuracy can be determined.

In particular, our main challenges are:

How can we formaly define Typical Driving Behavior?
How can we operationalize it?
Is this a problem beyond Deep Learning approaches typically used today in Self-Driving Cars?

People

Faculty

Prof. Dr. Wolf-Tilo Balke

PhD Students

M.Sc. José María González Pinto

References

[1] Homoceanu, S., Balke, W.T.: A chip off the old block – extracting typical attributes for entities based on family resemblance. Lect. Notes Comput. Sci. pp 493–509 (2015).
[2] Lee, T., Wang, Z., Wang, H., Hwang, S.W.: Attribute extraction and scoring: A probabilistic approach. In: Proceedings - International Conference on Data Engineering. pp. 194–205 (2013).
[3] Hua, M., Pei, J., Fu, A.W.C., Lin, X., Leung, H.F.: Top-k typicality queries and efficient query answering methods on large databases. VLDB J. 18, 809–835 (2009).
[4] Murphy, G.: The big book of concepts. (2002).

Contact

For more information about the project, please contact: M.Sc. José María González Pinto