Authors:
Philip Elspas
1
;
Yannick Klose
1
;
Simon Isele
1
;
Johannes Bach
1
and
Eric Sax
2
Affiliations:
1
Dr. Ing. h.c. F. Porsche AG, Weissach, Germany
;
2
FZI Research Center for Information Technology, Karlsruhe, Germany
Keyword(s):
Scenario Detection, Data-Driven Development, Time Series Segmentation, Fully Convolutional Networks, Driving Scenarios.
Abstract:
Leveraging measurement data for Advanced Driver Assistant Systems and Automated Driving Systems requires reliable meta information about covered driving scenarios. With domain expertise, rule-based detectors can be a scalable way to detect scenarios in large amounts of recorded data. However, rules might struggle with noisy data, large number of variations or corner cases and might miss valuable scenarios of interest. Finding missing scenarios manually is challenging and hardly scalable. Therefore we suggest to complement rule-based scenario detection with a data-driven approach. In this work rule-based detections are used as labels to train Fully Convolutional Networks (FCN) in a weakly supervised setup. Experiments show, that FCNs generalize well and identify additional scenarios of interest. The main contribution of this paper is twofold: First, the scenario detection is formulated as a time series segmentation problem and the capability to learn a meaningful scenario detection is
demonstrated. Secondly, we show how the disagreement between the rule-based method and the learned detection method can be analyzed to find wrong or missing detections. We conclude, that the FCNs provide a scalable way to assess the quality of a rule based scenario detection without the need of large amounts of ground truth infromation.
(More)