1. Insights
  2. AI Data
  3. Article
  • Share on Facebook
  • Share via email

Does data hold the key to autonomous vehicle safety?

Posted September 22, 2022
model cars driving in the same lane

Artificial intelligence (AI) proponents around the globe believe that autonomous vehicles, enabled with smart sensors and intelligent algorithms, will significantly improve road safety in the upcoming years. On the other end of the spectrum, opponents feel that autonomous vehicles pose a greater risk to pedestrian and driver safety, requiring additional protective measures. To quell these fears, in July 2022, The Council of the European Union mandated advanced safety systems in all automated road vehicles, including cars, trucks, buses and vans.

Despite criticisms, automotive innovators aren’t hitting the brake pedals on innovation. According to a report by ResearchAndMarkets.com, the global autonomous vehicle market will grow to over $2.16 trillion by 2030. Self-driving cars and robo-taxis from front-runners like Tesla, Waymo and Cruise are already hitting the roads. As autonomous vehicles continue to grow in popularity, AI technologies — powered by high-quality labeled data — will play a key role in ensuring superior safety standards for autonomous vehicles.

The case for autonomous vehicles improving road safety

Despite considerable safety enhancements made over the past decade, recent reports published by the National Highway Traffic Safety Administration (NHTSA) in the U.S. estimated that there were over 40,000 traffic fatalities in the previous year, a 10.5% increase from 2020. Most of these accidents were attributed to human error and caused by distractions, negligence or split-second misjudgments. Thus leading self-driving car experts to believe enhancing the role of technology will save many lives.

Autonomous vehicles offer distinct levels of autonomy, ranging from Level 0 to Level 5 (L5), where each level reduces the amount of human intervention in driving. Today, many AI-enabled driver assistance technologies like adaptive cruise control, automatic emergency braking and lane departure or forward collision warning systems — considered Level 2 — offer built-in support for drivers to reduce accidents and the resulting injuries or fatalities caused by impact. They are common additions to new vehicles sold in the United States and Europe.

A study conducted using the RAND Model of Automated Vehicle Safety concluded that highly automated cars that are 10% safer than human-driven cars offer a better probability of reducing fatalities and saving more lives than waiting for systems to be 70% or 90% better. “We find that, in the short term, more lives are cumulatively saved under a more permissive policy (Improve10) than stricter policies requiring greater safety advancements (Improve75 or Improve90) in nearly all conditions, and those savings can be significant — hundreds of thousands of lives,” the study outlines.

The sheer potential of autonomous vehicles continues to push car manufacturers to pursue L5 autonomy, where autonomous vehicles drive and make critical decisions on the road without human intervention.

High-quality annotated data will fuel the autonomous vehicles of tomorrow

Autonomous vehicles, functioning via intricate multi-neural network systems, require large volumes of diverse and accurately-labeled datasets to execute complex actions in response to real-world stimuli. Supervised learning is one of the main machine learning (ML) techniques employed to train autonomous systems.

Primarily, computer vision models enhance visual senses for AI-powered mobility systems. Data collected from various sensors such as cameras, lidars and radars are labeled accurately to help autonomous vehicles develop machine vision and perception. These algorithms are trained using the annotated data to perceive its surroundings and detect other vehicles and pedestrians sharing the road.

Cars with bounding boxes to show image annotation

Building high-quality 3D flash lidar datasets for autonomous vehicles

Learn how TELUS Digital helped a leading automotive lidar systems provider create high-quality flash-lidar 3D datasets with our fully-managed data annotation solutions.

Download the case study

The standard data inputs for computer vision models include image, video, point clouds and geo data. Data annotation procedures, such as image classification, object localization, detection and tracking, image or point-cloud segmentation, HD mapping and more, help train and familiarize models with the real world. From simple bounding boxes or polygons, to more complex 3D semantic segmentation, the type of labels used for the data varies from one use case to another. Case in point — a lane detection machine learning model will benefit from accurate lane markings made using polylines, and a pedestrian detection ML algorithm will produce better results using polygons.

Increasingly, innovators are also experimenting with speech recognition and natural language processing applications to enable voice-assisted features in autonomous vehicles to refine the overall driving experience. Text and audio data annotation such as transcription, translation, text categorization, linguistic annotation, classification and more provide inputs for these models. The key to unlocking all of these possibilities lies with high-quality and accurately labeled data.

The road ahead for autonomous vehicles

There’s still a long road ahead before self-driving cars are fully capable of making judgments without human intervention. Better autonomous vehicle perception requires high-quality data spanning several different real-world scenarios. However, sourcing training datasets that accurately represent a large volume of edge cases and safety-critical scenes can be challenging. Moreover, the dynamically evolving nature of our world and human unpredictability on roads calls for advanced and sophisticated mobility systems trained on accurately annotated data.

Generating comprehensive datasets for fail-safe models demands labor-intensive processes and sophisticated data solutions to collect and label data, test and validate quality, and ensure data diversity and volume of edge cases. TELUS Digital is a leading AI Data Solutions provider with years of experience building high-quality datasets for autonomous driving use cases. Talk with our experts and leverage our expertise to accelerate your autonomous vehicle projects.


Check out our solutions

Test and improve your machine learning models via our global AI Community of 1 million+ annotators and linguists.

Learn more