Join the 155,000+ IMP followers

www.auto-innovations.net
Mouser News

How Self-Driving Technologies are Redefining the Automotive Landscape

Global automotive manufacturers face many challenges. Implementing electric variants of their successful models is one, but a far more complicated goal is being a leader in autonomous vehicle technologies.

How Self-Driving Technologies are Redefining the Automotive Landscape
Self Driving Car (Source: Mouser)

Advanced driver assistance systems (ADAS) are available in many vehicles today, with blind spot detection and adaptive cruise control being the most popular, but replacing the driver takes it to a new level. Replicating the human process of negotiating a busy junction with machine learning algorithms is way beyond ADAS. Away from the technical considerations, there are more profound and emotional social concerns about autonomous driving.

It isn’t just the automotive manufacturers that want autonomous vehicles to be a success; national and regional road safety organisations wish to do it, too. Across the EU, the current road death figures are 25,000 annually, and there is a widespread belief that self-driving cars will significantly reduce fatality rates. Autonomous driving systems, equipped with multiple sensing methods, can detect potential risks far more quickly and reliably than humans. In addition, self-driving systems will have the advantage of instantly communicating with other vehicles to avoid accidents caused by misunderstandings with human drivers. Although collision risk identification is a key aspect of self-driving technologies, autonomous systems first need to learn the behaviour of human road users. There is a lot for machine learning algorithms to learn, and deep neural networks (DNNs) like data, lots of it. The more sources of information available, the better autonomous driving systems can operate safely.

Connectivity will also allow a vehicle’s road safety knowledge to be continuality updated, learning from the experiences of other autonomous vehicles. The concept of collective intelligence involves vehicles constantly streaming information to cloud-based management systems, creating a fertile knowledge base derived from the actions of individual cars that are shared with existing and future vehicles via software updates. Tesla, for example, embraces this approach and has been harvesting data from the beginning and continues to analyse the activity data to support the ongoing development of its autonomous systems.

Although it is clear there are many benefits to collecting data in this way to improve the acquired knowledge, it does bring into question privacy. How do you regulate who has the authority to collect, analyse, and distribute vehicle data? What about breaching the privacy of the vehicle occupants? Who will monitor and ensure that control mechanisms are in place? One approach is to abstract the data to a level that removes any personally identifiable information yet allows access to practical vehicle activity to analyse trends. Many aspects of data access and privacy need legal and regulatory authorities to decide on.

Extracting Intelligence from Acquired Data
Fortunately, road traffic accidents are relatively rare compared to the many hours vehicles spend on road networks. This presents challenges for data scientists tasked with creating self-driving neural networks. To be effective, these algorithms need vast amounts of data to learn how accidents occur so they can, once deployed, instantly identify and respond to a potentially fatal event unfolding as they are moving along the road. Simulation has a significant role in training machine learning systems, and researchers at Saarland University in Germany have developed a solution that doesn’t involve creating any actual accidents. They are leading the way with complex simulation-based training for neural networks by presenting possible accident scenarios and challenging events for autonomous systems to identify and resolve.

However, one of the challenges with simulation-based verification, even with access to detailed knowledge databases, is deciding how likely each scenario is. There is also the need to simulate situations resulting from a part of the system or a sensor failing. Should a sensor fail, it could cause the system to believe that a collision is about to occur when, in reality, nothing is in the way. Concurrence of data from multiple sensors may aid the neural network in self-diagnosing to alert vehicle occupants and place the vehicle into a fail-safe condition.

Recording vehicle experiences will aid this learning process so that autonomous vehicles can learn from their mistakes and those of others. Frequent data uploading so manufacturers can retrain neural networks, perhaps every night, offers a viable approach to keeping self-driving vehicles as safe as possible. As already highlighted, data privacy is a crucial consideration, protecting vehicle occupants from having their privacy violated. Abstracting data to a level where it becomes anonymised will achieve this, but it still needs clear legal guidance and agreement between vehicle owners and manufacturers.

Another aspect of developing and updating safe and reliable autonomous vehicle systems is the potential for vehicle manufacturers to collaborate rather than create proprietary systems. By sharing training scenarios, vehicle manufacturers and customers benefit from the likelihood that every possible situation has been identified. It might be that to achieve the highest level of incident identification, national road safety bodies and motor industry groups come together to set out standard training packages.

Self-driving autonomous systems rely on inputs from a myriad of sensors to navigate safely, but there needs to be a framework of higher-level protocols stipulating what to do if an accident cannot be avoided. This concept of a safety decision tree is already a controversial topic, where the safety of vehicle occupants and pedestrians are prioritised. Preparing an underlying safety framework is a complex and extremely sensitive legal and social dilemma, which we’ll investigate further in this blog series.

www.mouser.com

  Ask For More Information…

LinkedIn
Pinterest


Forgot Password?

Join the 155,000+ IMP followers