Blog post

AI in Transportation: A Pathway to Safe and Scaled Implementations

Scalable AI in transportation — are we there yet, and what does it take to accelerate the journey?

Updated on October 21, 2021

10 mins read

It may sound odd, but I was once called an AI in transportation skeptic, despite being a rather vocal advocate for this technology. But let’s be honest: it does feel that there’s more AI aspiration than action in our industry. Yes, self-driving vehicle pilots are being held all over the world. But does it mean I get to ride a self-driving shuttle to work next year? Probably not.

When it comes to self-driving vehicles, we have a clear ranking system of Level 1 to Level 5 autonomy. But what about other transportation systems and software solutions? How do you define “intelligent” for them, and what does it take to apply an AI use case in transportation? Let’s strategize.

Artificial intelligence in transportation: why it’s a promising but complex relationship

As we’ve written before, AI in urban mobility, logistics, and fleet management is highly anticipated and already deployed to an extent.

For instance, 65% of logistics company leaders name AI as an important technology for the next three to five years. Also, as of 2018, one in four public transport managers already use AI for real-time operations management and customer analytics.

Moreover, many governments are doling out funds and support for an array of AI transportation projects. Canada has an ambitious ACATS Program, offering up to $2.9 million in grant and contribution funding to companies to Advance Connectivity and Automation in the Transportation System.

Singapore has a national AI strategy promoting the creation of “intelligent freight planning” by 2030.

Intelligent freight planning
AI in Transportation: A Pathway to Safe and Scaled Implementations

Source: Smart Nation Singapore — National Artificial Intelligence Summary

Successful AI-based transportation pilots have been done across countries, too — mainly demonstrating pilot runs of self-driving electric pods, AI-regulated traffic light scheduling, and smart road infrastructure.

Yet despite successful reports for years, the number of successful pilots drastically outweighs the number of commercial AI transportation solutions.

Why is that? Because implementing and scaling AI deployments is a difficult balancing act of benefits vs concerns and tradeoffs.
AI in Transportation: A Pathway to Safe and Scaled Implementations

How to safely infuse AI in transportation: a roadmap

AI has the potential to solve pressing transportation problems. But as usual, execution and delivery is the real bane.

Collectively, the transportation industry is actively exploring how machine learning, intelligent automation, neural networks, and deep learning can be used to make transportation safer, greener, cheaper, and more efficient.

But given the complexity of the domain itself and the algorithm, there’s still no one-size-fits-all path to success. Yet there are a few proven steps that could help refine your product development vector.

Develop a data management strategy

Road sensors, connected car dashboards, floating cellular data, location-based services — the transportation industry has overwhelming access to data sources and big data use cases.

Discover viable big data use cases in transportation

Read more

Far fewer players have mature DataOps. Or even a formalized long-term data management and governance process — a cadence for processing, securing, standardizing, and operationalizing incoming insights.
AI in Transportation: A Pathway to Safe and Scaled Implementations

Yet AI systems are fickle when it comes to the quantity and quality of data. Scalable transportation algorithms require better, faster, and cheaper data processing. And in this case, you can’t choose just two options — it’s all or nothing.

When assessing a specific AI use case, think if your company can:

  • Obtain sufficient, variable, and labeled data sets for model training and validation.
  • Establish an ethical, anonymized, representative, and secure data collection process for passenger and vehicle data.
  • Enable low-latency data processing between all connected systems — IoT and edge devices, cloud data repositories, cloud-based or on-premises analytics systems.
  • Apply data fusion techniques to consolidate, pre-process, and operationalize records obtained from different sources and in different formats.

The intricacies of collecting and processing data for algorithms is one of the core reasons why AI transportation use cases are cautious small-scale pilots rather than city-wide deployments.

Getting sufficient training data is arguably one of the biggest issues. Transportation companies solve it in different ways. Tesla, for example, built a patented system for sourcing self-driving training data from its customers’ vehicles.

IEEE recently published a classification study that ranked different types of transportation data sources by accuracy, reliability, and cost of use.

Comparison of traffic data sources
AI in Transportation: A Pathway to Safe and Scaled Implementations

Source: IEEE — Data Sources for Urban Traffic Prediction: A Review on Classification, Comparison and Technologies

And even when you have obtained all the data you need, matters can still go awry if that data is skewed in some way. Take it from Volvo. When the company tested its AI-powered Large Animal Detection system in Australia several years ago, it was defeated by a kangaroo. Why? Because their system was designed in Sweden and originally trained and tested on elk, deer, and caribou — the native species — and wasn’t quite ready for a marsupial suddenly interfering in a road test.

The takeaway: Effective pattern recognition in the transportation industry requires varied, representative data. Obtaining it, however, may not always be immediately feasible or cost-effective.

But you are not alone in your battle for better transportation data. Private companies and legislators worldwide are working hard to provide access to better public data sources, standardize data formatting, and open data API integrations.

For example, the US Department of Transportation has several cool initiatives underway:

  • Safety Data Initiative — Aimed at integrating available data on car crashes and highway designs with anonymous data from GPS devices to develop better speed regulation practices
  • Waze Pilot — Tests the feasibility of using crowdsourced data from the Waze app to predict crash risks and other types of road accidents
  • Computer vision tools applied to naturalistic driving data — The feds are building a set of computer vision tools for collecting and analyzing naturalistic driving data so that everyone can get access to better datasets

Understand the advantages and limitations of different AI methods

When we talk about artificial intelligence in transportation, people with a technical background often think of algorithmic methods, whereas most other people imagine use cases.

AI as a term often lacks substance. Is Google autocomplete AI? Yeah, it is as of late. Is it the same AI that powers Alphabet’s Waymo self-driving vehicles? Nope; the methods are different.

Artificial intelligence is a collection of supervised, unsupervised, and reinforcement learning methods for developing (self-)learning systems:
AI in Transportation: A Pathway to Safe and Scaled Implementations

The common machine learning techniques in the transportation industry are:

  • Artificial neural networks (ANN)
  • Genetic algorithms (GA)
  • Simulated annealing (SA)
  • Fuzzy logic model (FLM)
  • Ant colony optimizer (ACO)

The use cases of neural networks in transportation engineering alone are plentiful — transportation demand forecasting, predictive maintenance of transport infrastructure, driver behavior monitoring, and more.

Transportation software development

Securely integrate intelligence into your core transportation, logistics, and supply chain products

Learn more

That’s a lot of choices, right? But every option comes with some inherent limitations. Let’s take ant colony optimizer algorithms. These were found to be effective for developing better routing of public transport and improving passenger pick-ups on popular ride-sharing platforms.

But scaling “intelligence” in transportation is hard. Researchers have found that ACO algorithm speed drops significantly when the number of iterations (e.g. the number of optimized routes) grows.

The use of neural networks in transportation is also ubiquitous. For instance, Uber uses such algorithms to provide accurate ETA predictions for journeys, and many customer analytics solutions are powered by artificial neural networks. Yet working out the optimal combination of weights, hidden layers, and training parameters for a neural network to perform efficiently is a tedious trial-and-error process. Oftentimes, the solution ends up performing well on a small dataset, but the performance drifts and flops in the pre-production stage, forcing your ML transportation division back to square one to try something else.

The bottom line: Scaled machine learning in transportation is hard to implement since no single method provides a universally plausible result. Thus, the development timeline is long and assumes some tradeoffs along the route to fast, cheap, and high-quality performance.

Focus on model explainability

One of the major constraints of AI in the transportation industry is that algorithms may be forced to make life or death decisions in dire circumstances (e.g. killing one passenger to save ten pedestrians). That’s why the question of model explainability and AI ethics is rather acute.

Before we allow our cars to make ethical decisions, we need to have a global conversation to express our preferences to the companies that will design moral algorithms, and to the policymakers that will regulate them.

The Moral Machine experiment, findings from Nature

The current concerns are mostly related to black box algorithms such as neural networks and reinforcement learning techniques, where the algorithms are given full rein in determining the optimal course of action.

In some cases, the algorithm may decide that imposing lower speeds in an area would be good for preventing casualties, but at the same time would lead to longer cruising for parking and subsequent traffic bottlenecks. Likewise, having an algorithm decide on the moral dilemma of letting an ambulance pass to save one patient versus keeping the road blocked for the fire crew to pass isn’t something most of us would feel comfortable with.

What adds to the complexity is that humans are controversial creatures too. A major study around the ethics of autonomous vehicles found that people in surveys said they wanted an autonomous vehicle to protect pedestrians over passengers. Yet most respondents also said they wouldn’t purchase a vehicle that was programmed this way!

One of the likely paths to ethical and sustainable AI adoption in transportation could be explainable artificial intelligence (XAI) — a methodology aimed at building descriptive AI models where developers can explain and moderate the model’s accuracy, fairness, and outcomes. In other words, XAI conveys the rationale behind the AI decision-making.

Ways to manifest and convey the reasoning behind AI decisions
AI in Transportation: A Pathway to Safe and Scaled Implementations

Source: Accenture — Understanding Machine: Explainable AI

Many other big data analytics models are actually explainable since they are powered by programmatic instructions and statistical methods for modeling different outcomes. Data science and machine learning models belonging to the “explainable” family include:

  • Linear models
  • Decision trees
  • Generalized additive models
  • Rulesets

However, because such systems are not self-learning, they cannot propose novel (and perhaps better) solutions that an ML/DL algorithm might devise. However, the cost of poor decisions in the transportation industry can be too high. Thus, tradeoffs in the levels of black box thinking may be required to ensure that your solution is ready for the market.

Collaborate with other ecosystem partners

The transportation industry has many private and public players. Plus regulators overseeing their actions. Oftentimes, market participants choose a strategy of competition over community. And this decision leads to a host of issues:

  • Low interoperability between different transportation systems
  • Lack of data format, quality, and integration standardization
  • Disparate and hard-to-integrate systems
  • Poor access to underlying infrastructure (e.g. road sensors, 4G/5G connectivity)

But without systematic, rigorous, and joint planning efforts, scaled implementations of new technologies are hardly possible.

System interoperability is not being advanced because the market evolution is not clear. If market norms could be agreed upon, then questions such as ‘who is liable if the service fails?’ will become clearer and investment decisions potentially easier to make. In addition, Future of Transport (FoT) is advancing new collaborative approaches to working with data as both a commodity and as an infrastructure.

The UK Department for Transportation

Interoperability between different software systems for the transportation industry — from city-wide traffic management software to individual fleet management systems — is crucial for ensuring that all vehicles and devices can effectively exchange data. Such dynamic data can then be used for optimizing transport system operations, leading to greater safety, predictability, and efficiency of the mobility experience for all end users.

Learn the role of AI in adapting cities to citizens’ needs in transportation, traffic management, and delivery services

Read more

Specifically, to realize the benefits of AI in transportation, more joint effort should be aimed at standardizing data quality so that most participants agree to specifications on:

  • Data structure, content, and exchange protocols
  • Definitions and classification of data types
  • Organizational practices for creating, exchanging, and using data

The good news is that many regulators already have working or approved frameworks in place. For example:

  • ISO/TC 204 was developed for regulating the system and infrastructure aspects of intelligent transport systems (ITS).
  • TC 241 specifies regulations for road traffic safety and management standards.
  • ISO/IEC 27001 is an IT systems security standard.
  • INSPIRE-MMTIS is an EU directive aimed at opening access to static and dynamic data categories by Member States to improve multimodal travel information services.
  • Auto-CITS is an EU body promoting unified standards for autonomous driving and V2X technology adoption.

Apart from achieving compliance, a greater degree of interoperability also makes transportation industry players more competitive. By pursuing a platform business model, you can tap into new pockets of growth without launching new features. Instead, you can gain extra revenue through API monetization, new sales channels (e.g. in-car commerce or digital freight matching), or launching “as a service” offerings (e.g. a mobility as a service solution).

As you assess different examples of transportation technology, think from the partnership perspective: will investing in this use case benefit just my business, or can it open up interesting new partnerships and cross-integrations?

Treading ahead with AI in transportation, safely and confidently

AI has a future in the transportation sector. The technology’s potential has already been shown in many amazing pilots, use cases, and market-ready solutions such as Rapid Flow real-time adaptive traffic light controls, TuSimple’s autonomous trucking, and Sensible 4 self-driving buses.

However, to see more scaled deployments of AI-powered transport solutions, more groundwork should be laid today, from tech architecture standardization and operating safety protocols to fine-tuning ML/DL methods and realigning business models. AI is a complex technology to implement. So how do you plan to turn it into an ROI-backed investment rather than an ongoing expense?


Contact Intellias for a personal consultation on incorporating AI into your transportation software solution.

Your subscription is confirmed.
Thank you for being with us.

4.9 Thank you for your vote. 37454 bd7ac300b7

Tell us about your project

I give consent to the processing of my personal data given in the contact form above under the terms and conditions of Intellias Privacy Policy. I want to receive commercial communications and marketing information from Intellias by electronic means of communication (including telephone and e-mail).
* I give consent to the processing of my personal data given in the contact form above under the terms and conditions of Intellias Privacy Policy.

Awards and recognition

top-employer
clutch
Inc-5000
GO100_Generic
dou
gsa
clutch-design
iso-27001 (1)
iso-2001-2015 (1)

Thank you for your message.
We will get back to you shortly.

Thank you for your message.
We will get back to you shortly.