Readspeaker menu

Vulnerability Classification in Automotive Scenarios

What are the vulnerabilities of all detected obstacles ahead a vehicle?

Johannes Rühle, Erik Rodner, and Joachim Denzler

Update 10/2015: Here you can find the poster presented at the final UR:BAN event in Düsseldorf. [poster(eng)][poster(de)]


Increasing road safety means increasing the knowledge about the obstacles and objects in front of the vehicle. Our proposed vulnerability classification provides a method to distinguish how vulnerable all detected obstacles are. Therefore, we rely on image data and obstacle detections generated by a vehicle stereo camera mounted behind the wind shield, which act like the vehicle's eyes of an observant driver. The information about the vulnerabilities ahead the vehicle can directly be used to plan evasion maneuvres or accident mitigation reactions in emergency situations. So, we propose a method to minimize casualties in vehicle accidents.

Vulnerability Classification

The vulnerability classes - ranging from small (none) over medium and heavy to fatal - express the expected severity of damage from the object's perspective when assuming a collision with the driver's vehicle.

Vulnerability Classification in Automotive Scenarios

Live in action: The vulnerability classifier on the test track (Final UR:BAN presentation in Düsseldorf, October 2015)

Additional Visual Results and Examples MVA 2015

Here, we present some additional visual results that were created for the the work presented in:
Johannes Rühle and Erik Rodner and Joachim Denzler.
Beyond Thinking in Common Categories: Predicting Obstacle Vulnerability using Large Random Codebooks.
Machine Vision Applications (MVA) 2015. p.198-201 [pdf] [bib][poster(eng)][poster(de)]

For a more thorough method description and result discussion, we refer you to this publication. The following examples of the vulnerability classification results indicate the obstacle (detected by the stereo vision camera and its build-in obstacle detection unit). The color coding is accordingly to the vulnerability classes: none (blue), medium (yellow), heavy (orange), fatal (red).

Good Classification Examples Bicyclist Sequence False Cases

Demo Scenario with Wheel chair