Even a little crummy weather might pose big problems for self-driving systems.

New research from Michigan State University suggests light rain and drizzle can confound the algorithms that autonomous systems use to detect pedestrians, bicyclists and other road users.

The findings raise the prospect that until these algorithms can better handle a variety of weather conditions, self-driving vehicles may be limited to Sun Belt states, or fleets of vehicles might need to be grounded when weather conditions are subpar.

“When we run these algorithms, we see very noticeable, tangible degradation in detection,” said Hayder Radha, an MSU professor of electrical and computer engineering who oversaw the study. “Even low-intensity rain can really create some serious problems, and as you increase the intensity, the performance of what we consider state-of-the-art mechanisms can almost become paralyzed.”

The researchers are finalizing their report, but Radha previewed the findings in a conversation with Automotive News.

Although radar and lidar are often used to detect obstacles, Radha said the research focuses on measuring the competence of computer vision systems because cameras are most often the primary sensor automakers and tech companies use to classify pedestrians and other road users.

But the problem is not the cameras, Radha stressed; it’s the algorithms distilling information from them.

“Once you throw in a few drops of rain, they get confused,” he said. “It’s like putting eyedrops in your eye and expecting to see right away.”

Cold reality

Researchers evaluated various parameters for the study, including the size of the raindrops, the number of raindrops per square inch and the effect of wind velocity. Using a scale that ran from clear weather to a blinding rainstorm, they found algorithms failed to detect as many as 20 percent of objects when the rain intensity was 10 percent of the worst-case scenario. When rain intensity increased to 30 percent, as many as 40 percent of objects could no longer be detected.

Information from radar and lidar may mitigate some of the problems, especially as self-driving engineers develop ways to use returns from those sensors to classify objects. But the most improvement will come, Radha says, when self-driving tech is routinely tested beyond sunny locations in Arizona, Florida and Texas.

The ongoing research is one of Michigan State’s efforts to study self-driving technology and mobility. The school created an interdisciplinary program called CANVAS — Connected and Autonomous Networked Vehicles for Active Safety — part of which examines the situational-awareness capabilities of self-driving systems, including their competence in detecting and handling adverse weather.

Broadly, Michigan State has strived to increase its focus on autonomous systems and advanced transportation.

That has been an ambition throughout the state’s higher-education ranks. Initially, the University of Michigan was a primary focus thanks to its Mcity closed-course autonomous-vehicle test track. But more recently, the state has sought to ensure all its schools are keeping mobility in mind. Last year, Washtenaw Community College in Ann Arbor launched an Advanced Transportation Center to help train field technicians to build and repair intelligent transportation systems.

At Michigan State, researchers have affixed cameras to traffic lights around the university’s home in East Lansing, allowing them to gather data on pedestrian, bicyclist and vehicle movements and better predict their movements, a crucial capability for self-driving systems. Elsewhere on campus, university officials have reclaimed Spartan Village, an area filled with dilapidated apartments, and used it as a public course to test autonomous technology and gather data.

Seasonal effect

That testing has yielded another weather-related finding. Depending on the environment, high-resolution maps that autonomous systems use to determine their location may need seasonal updates. As the leaves fell from towering trees in Spartan Village, researchers realized their maps were outdated.

“You can imagine in environments where there are a lot of leaves on trees or on shrubs close to the road, they are an essential part of the map,” Radha said. “So summer and winter are completely different. When they fall down in winter, you have nothing to work with. So that tells you that for this technology to be robust, it needs to be developed in different conditions than you see only in Arizona and Silicon Valley.”

Cold weather can also cause problems for autonomous vehicles. Last winter, he found temperatures of 10 degrees Fahrenheit and lower increased the “noise,” or amount of poor-quality or irrelevant returns from lidar sensors. He says manufacturers told him they could not guarantee the sensors would operate in extreme cold, and that they’re working to expand the range of the sensors.

Whether it’s cold weather, foliage or light rain, Radha says the shortcomings underscore the fact that weather diminishes the capabilities of self-driving technology for the foreseeable future.

“Frankly, at first we thought we’d look at this for a couple of years and then be done,” he said. “But I’d say now that challenging weather conditions are going to be a problem for many years to come.”