How safe are the perception capabilities of autonomous cars?

One of the fundamental aspects that will need to be addressed for the introduction of automated vehicles (AV) is the standardisation of safety assurance. That is, what are the requirements that an AV must satisfy and how to verify these requirements.

Perception refers to the ability of an autonomous system to collect information and extract relevant knowledge from the environment. Perception is an essential component in automated vehicles for the system to understand the world in proximity under all working conditions and be able to navigate safely from A to B in a safe and efficient manner.

This research will look how perception systems should be specified and tested to guarantee a certain level of safety under different operational domains.

This research project aims to deliver methodologies to determine the safety of automated vehicles while operating in defined Operating Design Domains (ODD) and developing standards to enable ODD to be tested. It will place special attention on perception system requirements since the outputs will be applied to high performance Advanced Driver-Assistance Systems (ADAS) and level 4 AVs.

Participants

• IAG
• University of Sydney

Project background

One of the fundamental aspects that will need to be addressed for the introduction of AVs is the standardisation of safety assurance. That is, what are the requirements that an AV must satisfy and how to verify these requirements.

Perception refers to the ability of an autonomous system to collect information and extract relevant knowledge from the environment. Perception is an essential component in automated vehicles (AV) for them to understand the world in proximity under all working conditions and be able to navigate safely from A to B in a safe and efficient manner.

This research will look how perception systems should be specified and tested to guarantee a certain level of safety under different operational domains.

This project will address the development of algorithms and techniques to design and evaluate safety metrics that can assure safety in a particular design operation domain. The project will look at characterising perception and control uncertainty to be used with safety metrics to evaluate the level of safety and perform risk assessment of connected and automated vehicles under well-defined operation domains.

Project objectives

The project aims at developing a comprehensive understanding of how to evaluate the safety of autonomous vehicles. As such the research will include:

• A precise definition of safety
• Understanding Different Operating Design Environments (conditions where developers intend the AV system to operate): geographical, environmental condition, traffic infrastructure, etc
• Determination of safety metrics: Determination of a finite set of safety properties that can guarantee a definition of safety
• Understanding the system validation required to ascertain perception safety

December 2022: Project completed

This project has been completed, and we hope to make the final report available soon.