9 Revolutionary AI Applications In Transportation

 Or get in touch with our team to discuss your project ;-)

Now, let’s begin!

Self-driving Vehicles

The concept of self-driving vehicles is nothing new. General Motors introduced it back in 1939. 

But it’s only in our current age of AI transportation that companies are able to use computer vision techniques like object detection to create intelligent systems that decode and make sense of visual data to—essentially—allow a vehicle to drive itself. 

And while a self-driving car can sound complex, the idea for building the AI behind it is actually straightforward: The algorithm is fed huge swathes of relevant data, before being trained to detect specific objects and then take the correct actions, such as braking, turning, speeding up, slowing down, and so on. 

So—

Which objects does a model need to identify? 

The likes of other vehicles on the road, road signs, traffic lights, lane markings, pedestrians and more. 

To collect and use data, autonomous vehicles use cameras and sensors. To train the model and to make it reliable, it needs to be consistently fed masses of data. 

Naturally, there still remain some challenges. 

An algorithm needs to get access to those huge swathes of relevant data, while situational conditions like bad weather and uneven terrain can also pose a problem. Other issues include poor lighting and the possibility of a self-driving car coming across an unidentified object while out on the road. 

Looking for quality data for your training model? Check out 65+ Best Free Datasets for Machine Learning.

Of course, when many of us think of self-driving cars, we automatically think of Tesla. 

Tesla—along with the likes of Uber, Waymo and Motional—has been working on automated vehicles for a number of years now, always staying one step ahead of the curve. 

Unlike some others in the AI in transportation industry, Tesla utilises the purely vision-based approach, using their camera-equipped cars to collect video and image data without using HD maps and lidars in their autonomous driving stack. 

From a technical standpoint, this is actually a more complex approach, largely due to the fact that, because the neural networks are being trained just on video data alone, the need to achieve the highest accuracy possible becomes essential. 

However, Karpathy points out: 

"Once you actually get it to work, it’s a general vision system which can be principally deployed anywhere on earth."

Tesla’s self-driving team amasses lots of data too—as much as 1.5 petabytes of data that consist of 1,000,000 ten-second videos and 6,000,000,000 objects, each of which is annotated with velocity, depth and bounding boxes. 

This isn’t to say that Tesla relies solely on manual data annotation. Rather, it improves the annotation process by combining human review with auto-annotation tools. 

Looking for the perfect auto-annotation tool? Check V7’s automated annotation to label your data 10x faster. 

AI systems are limited to self-driving cars—they’re also used in trucks, buses, and airport taxes, with the innovations having a huge impact on AI in logistics and the supply chain in general. 

Indeed, Mckinsey has predicted that self-driving trucks will reduce operating costs by some 45%. Environmental impact will also be greatly reduced. 

Excited?

Over the rest of the article, we’ll be taking a closer, in-depth look at some of the more specific computer vision and machine learning cases that are laying the foundations for autonomous driving technology.

Check out our overview of AI in Supply Chain and Logistics [20+ Practical Applications]

Traffic Detection (and Traffic Signs)

There are thousands of traffic lights in the US alone. And while you might think that stopping when a light turns to red is a simple process, the fact that each year in the US some 1,000 people are killed needlessly by vehicles running a red light means that the whole thing is a very risky, dangerous and even complex game.

It’s a game with tragic consequences, too, with over 50% of those deaths accounted for by passengers or drivers who didn’t run the red light. 

The problem is that the traffic light system itself might be perfect, but the humans behind the wheel aren’t always perfect. Mistakes happen, sometimes drivers run a red light—and accidents occur. 

The solution to this terrible problem can be found in autonomous vehicles that, alongside smart cities, can prevent those deaths. 

Indeed, automakers are putting the traffic signal issue at the front and centre of their self-driving cars capabilities. 

An AI-based system can be trained to recognise lights—green, amber and red—via computer vision models that are trained in a wide range of scenarios, such as poor light conditions, inclement weather and occlusions.

As such, a self-driving car’s cameras first spot a traffic signal, before the image is analysed, processed—and, if it turns out that the light is red, the car puts the brakes on. 

Naturally, there are issues here. When a camera is scanning what’s in front of it, it may spot other lights—such as a billboard or a streetlamp. Yes, a traffic light is different to a streetlamp in that it has three lights, but an image analyser capability still needs to be so good that it can spot the traffic signal instantly and not be fooled by other lights.

If it’s fooled, the result could be devastating. 

When annotating the data for traffic detection, then, one of two common approaches are needed: 

• Labelling light and shell individually (most common)

• Using the colour of the active lights to define each traffic light object class

You can also label pole and light individually. 

Whenever a smart city offers proximity sensing to an autonomous car, understanding its supporting structure is the only way to know the position of a free-floating object. 

This way of annotating can also be used to determine traffic lights from one another which are presented in multiple lanes, with each set of lights correlated and grouped to the lane below.

Check out A Comprehensive Guide to Convolutional Neural Networks.

This paper explores a traffic light detection and recognition approach that uses convolutional neural networks. 

Using map data and a pair of separate focal length cameras to detect traffic lights at different distances, researchers came up with a unique algorithm for light recognition which combined image classification with object detection to identify the light state classes of traffic lights. 

The researchers also integrated YOLOv3 into their approach for real-time traffic light detection for better results. 

Here’s a simplified visual representation of the process: 

‍Results: The proposed approach didn’t manage to achieve 100% accuracy. Because traffic light detection and recognition requires 100% accuracy to ensure the safety of the passengers and pedestrians, improvement is needed. 

Check out how the V7 model handles traffic light detection.

Pedestrian Detection 

How cool would it be if a computer system could automatically spot and identify pedestrians in images and videos? 

Further, what if we could create a model that would allow autonomous cars to understand a pedestrian's intent so that they would know—for example—if a pedestrian intended to cross the road in real time?

Such a system would certainly help self-driving cars to swerve dangerous situations, and potentially massively reduce road accidents. 

Pedestrian detection is actually a key problem in Computer Vision and Pattern Recognition, because pedestrians can be super unpredictable in the context of road traffic. They’re so unpredictable that they pose one of the greatest threats to the success of self-driving cars. 

The key is not necessarily that a system recognises specific human features, such as beards and noses, but that it’s able to properly distinguish a human from another object, as well as understand what a pedestrian is planning to do next. Ergo, will they cross the road?

To begin the task of identifying and visualising pedestrians, computer vision systems use bounding boxes. 

Searching for the perfect bounding box tool? Check out 9 essential features for a bounding box annotation tool. 

To detect pedestrians, different types of features have been used, including motion-based features, texture-based features, shape-based features, and gradient-based features. 

Some approaches have also incorporated human pose estimation, a technique that collates information about the instant behaviour of a particular subject (in this case, a human). This is designed to relay information to the autonomous vehicle in regards to what a pedestrian is intending to do next. 

For instance, this paper takes a look at how the intentions of pedestrians can be predicted using 2D skeletal pose sequences via a deep learning network. The researchers wanted to create a model that would tell them in no uncertain terms whether or not a pedestrian is going to cross the road. 

They linked the dynamics of a human’s skeleton to an intention so as to overcome the issue of real-time discreet intention prediction in a typical traffic environment.

The result?

Experimental results show that a 94.4% accuracy was achieved by SPI-Net in pedestrian crossing prediction based on the JAAD dataset.

Here’s a simple visualisation of the process:

Of course, there are still challenges in training data to overcome, and these include varying illumination parameters in varying scenarios, the different kinds of poses and clothing that pedestrians wear, as well as changing lighting conditions. 

The latter issue is handicapped by the use of machine vision and cameras, which is why more advanced technology is needed to provide data that's' more accurate, and which can then be used to successfully identify pedestrians in all lighting conditions. 

Moreover, it’s the success rate of a machine learning algorithm that will determine how successful pedestrian detection ultimately is. 

Traffic Flow Analysis 

The flow of traffic impacts a country’s economy for the better or worse, and it also impacts road safety. Traffic congestion costs money and time, it causes stress to the drivers and passengers, and it also contributes to global warming. 

With better traffic flow, a country’s economy can grow better, and the safety of its road user’s is improved immeasurably.

With this in mind, it’s no surprise that Artificial Intelligence is now paving the way for better traffic flow analysis using machine learning and computer vision. AI can help to reduce bottlenecks and eradicate choke-points that are otherwise clogging up our roads—and our economy.

Thanks to the advancements of computer vision, drone and camera-based traffic flow tracking and estimation are now possible.

The algorithms are able to track and count freeway traffic with accuracy, as well as analyse traffic density in urban settings, such as on the freeway and at intersections. This helps towns and cities to understand what’s going on so that they can design more efficient traffic management systems, while at the same time improving road safety. 

CCTV cameras can spot dangerous events and other anomalies, as well as provide insights into peak hours, choke-points and bottlenecks. It can also quantify and track changes over a period of time so as to allow the measurement of traffic congestion. As a result, urban traffic and emissions can be greatly reduced by town planners.

The likes of Flir, Viscando and SwissTraffic have been using Artificial Intelligence for traffic flow, with Viscando using stereo vision technology to monitor and control traffic. 

Viscando’s system monitors the flow of traffic at intersections and other open areas, and it can spot and track a variety of vehicles, as well as pedestrians and bicycles, all at the same time. 

Such capabilities ensure Viscando are able to track the trajectory of road users at traffic lights, as well as identify conflict risk and calculate the gaps road users leave between cars. 

As Viscando’s CEO, Amritpal Singh says:

"It also gives the cities much more data about how the intersection is working, the length of queues and the duration of waiting time, and being able to include pedestrians and cyclists in the same optimisation scheme."

This paper puts forward a framework with numerous movements and classes for better vehicle counting. The researchers used advanced deep learning methods for vehicle detection and tracking, as well as a trajectory approach that allowed them to monitor the movements of vehicles. 

The researchers wanted to improve the process of counting traffic volume, which itself has been a complex task, based as it is on the CCTV system. The issue has always been the involvement of too many vehicle movements. If the researchers could implement distinguished regions tracking so as to monitor the different movements of vehicles, they could improve the counting process. 

The results?

The experiment results are promising, with the model achieving an accuracy for different movements between 80 and 98%, all with just a single view of the camera. 

As well as boosting traffic flow, AI complements other smart city objectives, such as sustainability.