Scholarlink

Test-drives of Automated Driving Systems (ADS) generate a rich pool of data that can be used to analyze and improve the ADS software stack. In this context, disengagement events, i.e., situations where the safety driver takes over control of the ADS are of specific interest. While it is easy to automatically identify when a disengagement happens, it is a non-trivial, and therefore manually performed task to classify disengagements with regards to its cause. The goal of our study was to replace the current manual classification process with a more efficient, scalable and reliable automatic approach. To this end, using supervised learning, we developed and tested a set of eight CNN-based binary classifiers, one for each label type. The evaluation indicated a high performance for six of the eight label types. A follow-up SHAP analysis gave insights into the reasons for the good performance of the classifiers.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated Classification of ADS Disengagements Using Convolutional Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Elisabet Hein,
  • Ali Ihsan Güllü,
  • Faiz Ali Shah,
  • Dietmar Pfahl

摘要

Test-drives of Automated Driving Systems (ADS) generate a rich pool of data that can be used to analyze and improve the ADS software stack. In this context, disengagement events, i.e., situations where the safety driver takes over control of the ADS are of specific interest. While it is easy to automatically identify when a disengagement happens, it is a non-trivial, and therefore manually performed task to classify disengagements with regards to its cause. The goal of our study was to replace the current manual classification process with a more efficient, scalable and reliable automatic approach. To this end, using supervised learning, we developed and tested a set of eight CNN-based binary classifiers, one for each label type. The evaluation indicated a high performance for six of the eight label types. A follow-up SHAP analysis gave insights into the reasons for the good performance of the classifiers.