Artificial intelligence (AI) plays a large role in the automotive industry as it competes to develop and market vehicles that can perform tasks that previously required an experienced driver.

But convenience isn’t the only factor that drives automotive innovation; safety does, as well. According to the National Highway Traffic Safety Administration (NHTSA), from 1968–1998, the introduction of seat belts and airbags reduced the chances of death from a head-on collision by 61%.[1] Other safety adoptions to vehicles are becoming laws as well. As of 2018, the U.S. requires all new cars to feature backup cameras.

Over several decades, automakers have been implementing advanced driver-assistance systems (ADAS) functions to reduce risk to an acceptable level. These electronic systems in vehicles use advanced technologies such as AI to assist the driver in dynamic driving tasks (DDTs), i.e., all the real-time operational and tactical functions required to operate an on-road vehicle in traffic. These include lateral/longitudinal vehicle motion, monitoring the driving environment via object and event detection, recognition, classification, response preparation, object and event response execution, maneuver planning and enhancing conspicuity via lighting, sounding the horn, signaling, gesturing, etc.

The automotive industry’s innovations around visual perception have grown exponentially in recent years. What were once only image processing theories are now within the power budget and timing requirements of hardware-based neural network accelerators and graphics processing units (GPUs) to run in real-time under a vehicle’s hood.

Software has evolved and increased in complexity due to the massive computation capabilities this hardware has achieved. Several autonomous vehicle (AV) elements rely on algorithms based on machine learning (ML). This type of learning encompasses a set of tools that enable computers to learn a task using training and defined human-understandable rules. Of course, these powerful techniques do not come without cost. Their complexity and limited internal visibility require additional design effort compared to algorithms used in traditional safety-related components, including modification of validation methods.[2]

So, is the automotive industry considering guidelines for designing ML models for automotive safety assurance? In recent years, manufacturers have applied the established automotive safety engineering frameworks ISO 26262:2018 and ISO/DIS 21448:2021 in traditional model-based system development. However, these standards do not help define the artifacts for machine learning algorithms.

Newer automotive standards such as UL 4600, the Standard for Evaluation of Autonomous Products, and ISO/TR 4804:2020 introduce the terminology and some considerations of ML and AI techniques under automotive usage. Another set of standards tackles the missing point of the automotive safety standards for L3–L5 vehicles using ML complex algorithms. Unfortunately, these are in the early stages of the development process.

To learn more about the new standards and defining, specifying, developing, evaluating, deploying and monitoring ML algorithms, we invite you to join UL’s AI/ML hands-on introductory training. For more information about automotive-specific ML techniques, you can join our AI/ML for automotive safety assurance training. Both courses, coming in early 2022, will help define automotive-specific ML techniques. We encourage you to contact us to learn more.

[1] How Has Car Safety Improved Over 60 Years? https://www.visualcapitalist.com/how-has-car-safety-improved-over-60-years/

[2] ISO 4804, 6.2 Validation of (sub) systems that are based on machine learning

Leave a Reply.