KI Data Tooling

Survey on Anomaly Detection Methods published at CVPR workshop

A new publication in KI Data Tooling sheds light on current methods to detect rare scenes and scenarios in data or in online applications for autonomous driving. The work Anomaly Detection in Autonomous Driving: A Surveyby Bogdoll et al. was accepted at the Workshop on Autonomous Driving at CVPR and presented in New Orleans in June.

While the perception of autonomous vehicles performs well under closed-set conditions, they still struggle to handle the unexpected. This long tail of unknown or rare scenes and scenarios – also known as corner cases or anomalies - is one of the core challenges with respect to the scale-up of autonomous driving fleets.

One of the work packages in KI Data Tooling tackles this topic: Its aim is to gain a detailed look at corner cases, their representations and ways to detect them. This publication works directly towards this goal of an assessment of different data sources and formats with respect to their suitability for identification of corner cases.

To deal with such situations or create more diverse datasets, they need to be detected first. However, with the work Corner Cases for Visual Perception in Automated Driving: Some Guidance on Detection Approaches by Breitenstein et al.,so far only a single, domain-specific overview of corner case detection methods for camera data was available. As a multi-sensor setup is typical in the domain of autonomous driving, the motivation was born to create a more general overview for all typical sensor modalities as well as abstract representations.

The new survey provides an extensive overview of anomaly detection techniques based on camera, lidar, radar, multimodal, and abstract object level data. It furthermore provides a systematization including detection approach, corner case level, ability for an online application, and further attributes. The publication outlines the state-of-the-art and points out current research gaps.

With this work as a baseline with regard to the state of the art, the project can now focus on actual methods to detect corner cases. For other researchers who are new to this field, this survey serves as an ideal entry point. For experienced researchers, the paper can help discover research gaps to develop new and creative corner case detection methods.