Smart cars, another emerging field associated with IoT, is gaining momentum. Its progress is closely connected to the advancements made in machine learning in the last decade. If you have never heard about machine learning, you should put this book aside for a second and watch the TED talk The Rise of Artificial Intelligence through Deep Learning by Yoshua Bengio (https://www.youtube.com/watch?v=uawLjkSI7Mo). In a nutshell, machine learning makes it possible for computers to learn just like our brain does. It is another future technology that will be paired with IoT more and more to create smart, self-learning devices.
Machine learning is used in smart cars to develop many of its essential features:
- Detect the street, other cars, and people
- Understand signs and speed limits
- Identify dangerous situations and know how to solve them (for example, by applying the brakes)
The following screenshot shows a simplified view of object detection in a smart car:
Source : (https://www.pexels.com/photo/architectural-design-asphalt-buildings-city-139303/)
If you compare the visual interpretation of a human's view of a street and the digital representation of the same theme, the two differ enormously—a computer seeing through a camera just sees raw data, the amount of red, green, and blue per pixel, and nothing more. Machine learning makes this data more usable by training the computer based on input footage—for example, by supplying a large number of images depicting street views.
After many learning iterations, the computer might be able to identify a street using fresh footage. Machine learning and IoT will be good friends in the future as internet-enabled microcontrollers become smaller and more powerful.
Currently, complex machine learning models require an expensive state-of-the-art computer, but there are already experiments using the Raspberry Pi, a tiny computer that runs Linux, for simple machine learning tasks. Google and NVIDIA introduced two new development boards (so-called edge devices) in 2019, which have a similar form factor to the Raspberry Pi and are intended for machine learning prototyping: Google Coral and NVIDIA Jetson TX2.
But so-called on-device training is not the only way hardware devices can use the power of machine learning. The most common way they use machine learning today is by sending the device's data to a cloud server where the heavy analysis is done. One example of this is Google Photos. It allows you to upload your photos, in most cases taken with a smartphone, to the Google servers. The servers will analyze each and every one of them using various machine learning models.
You can already use the results in two ways, as:
- The machine learning model detects all faces in a photo and groups them together into categories. This way you can easily find all photos that contain your face or any of your friends' faces.
- The machine learning model detects objects in photos. You can then, for example, filter all of your images that contain a red car.
Apart from machine learning, to understand their surroundings, smart cars can communicate with each other. Every now and then there are reports about a mass crash on the highway. Smart cars will be able to warn each other about dangerous situations: "Attention cars behind me. There is an obstacle lying on the street. Better slow down!".
When it comes to situations like this, the amount of time it takes for another car to receive this information can make the difference between life and death—one or even multiple seconds response time is just not good enough here, the response time needs to be in the milliseconds. If the cars were using the internet for communication, it might take too long. When sending data to a satellite and spreading it from there to all of the nearby cars, there are too many things that can go wrong and prevent the warning from being delivered in time. Fortunately, there is a solution for this: using a technology called Vehicle to Vehicle (V2V). With this technology, cars can talk directly to each other by opening a network themselves (like a router). Using this, they create a mesh of connected cars without needing internet access.
