Driving on snow-covered roads is a major challenge for many self-driving cars. When the road markings are no longer visible and when snowy weather obscures the vision, the car has a hard time calculating the route.
Researchers at Michigan Technological University in the United States have trained an AI to better handle snowy conditions. The researchers have developed a technology they call Sensor Fusion where they used a variety of different types of sensors to train the AI in all types of snowy conditions.
The researchers have developed a method for clearing the data before handing it over to the AI. There are many different types of snow weathers, and by first categorizing the type of snow present in the data, the AI can make better decisions.
“AI is like a chef — if you have good ingredients, there will be an excellent meal. Give the AI learning network dirty sensor data and you’ll get a bad result.”
says Jeremy Bos, one of the researchers behind the study, in a press release.
Taking multiple data in account
The AI takes data from cameras, radar, lidar, and infrared and, based on that creates an image of what the road actually looks like and what needs to be done. One problem that the AI solves is that the sensors do not always agree on what they see. For example, it might be the case that one sensor signals that there is no obstacle at all, and another that there is and it’s a car, and the third that there is a deer on the road. The AI can take all this data, calculate the probability of each case, and determine what the worst possible outcome is. It can is then arrive at a solution that balance all the important parameters.
The researchers need to develop the AI more, before it can steer a car through a snowstorm. But by focusing entirely on snowy conditions, they hope to be able to complement other research in the field that often uses data from roads in summer conditions to train their AI.
Read the full study here.
