A new type of artificial intelligence system has been developed to allow autonomous vehicles (AVs) to better navigate in severe weather.
Researchers from Oxford University’s Department of Computer Science collaborated on the project with a team at Bogazici University in Istanbul, Turkey.What they found was post In the peer-reviewed scientific journal Nature Machine Intelligence.
The abstract articulates the problem the researchers hope to solve: “Because sensors are flawed in adverse environmental and weather conditions, being able to know their precise location on a map is a challenging prerequisite for safe and reliable AVs, which is critical for Their widespread use poses a huge hurdle.”
Essentially, in extreme weather, such as heavy rain, sleet or snow, there is an increased risk that a self-driving car will detect itself in the wrong position, which could affect the sensors. In some cases, this could have dangerous consequences, for example, if a self-driving car detects that it is in the wrong lane or in the wrong place at an intersection before making a turn, it may not be able to stop in time.
To address this problem, researchers developed a novel, self-supervised deep learning model Used for so-called “ego motion estimation”. This is a key component of the autonomous driving algorithm stack that estimates where the car is moving relative to obstacles observed from the vehicle.
The model takes into account information from sensors such as cameras, lidar, and radar—all affected differently by different weather (such as low light or precipitation)— So that each person’s benefits can be used in different conditions.
A collection of publicly available AV datasets was used to generate algorithms that recreate the geometry of the scene and calculate the car’s position based on the new data. Tests under various conditions including fog, snow and rain demonstrate the robustness of the model.
The team believes its research marks a major breakthrough, and the abstract sums it up as follows:We anticipate that our work will bring self-driving cars one step closer to safe and reliable 24/7 autonomous driving. “
Professor Andrew Markham, from the University of Oxford’s Department of Computer Science, who co-supervised the study, added: “Estimating the precise location of an AV is a key milestone in enabling reliable autonomous driving in challenging conditions.”
“This research effectively leverages the complementary aspects of different sensors to help AVs navigate difficult everyday scenarios,” he said.