MIT Uses Deep Learning to Predict Car Crashes

Car accidents are a leading cause of children and young adults, with an average of 6 million collisions happening in the US every year alone. More than 90 people die in a car accident, on average every day. These are frightening statistics that, among other things, have given rise to tremendous investments in safety measures for car transportation-- from the seatbelts to self-driving cars, and much in between. But we may have a new and powerful weapon to combat the dangers of driving: using machine learning to predict where car accidents are likely to happen.

Scientists from MIT’s Computer Science and Artificial Intelligence Laboratory and the Qatar Center for Artificial Intelligence announced this week that they had trained a deep learning model on historical crash data, roadmaps, satellite imagery, and GPS traces to create high-resolution crash risk maps. Their work is a significant achievement in the application of deep learning and may help mitigate the huge financial cost, bodily injury, and loss of life that is still such a part of modern car travel.

One of the fascinating developments here is the extent to which the model developed was able to work so well in novel environments. In other words, the model was able to learn enough about traffic patterns, roadways, and historical data to draw conclusions about how the interplay of all of these data inputs contributed to the likelihood of a crash, even when it considered completely new maps or maps where there was no historical crash data. Despite the prevalence of crashes, the likelihood of a particular block having had a car crash on it in the last several years is actually quite low, so if the model could only make predictions about crashes in areas where there had been crashes before, that would be quite limiting. 

Instead, the additional data points- GPS data, satellite images, etc.- proved to be sufficiently instructive. In fact, the scientists were able to use data from 2017 and 2018 to accurately predict areas of great risk in 2019 and 2020, even when the areas identified as high risk in 2019 and 2020 had experienced no incidents in 2017 or 2018. This would seem to speak to the significance of gathering appropriate data inputs that may be contributory to a targeted output, even when the extent of those contributions is not clear and/or the outcomes haven’t yet been observed in the full range of possible contexts. That is certainly the promise of deep learning specifically, and machine learning more generally, and this application of it would suggest great real-world implications for realizing that promise.

Models such as these could power an entire generation of safety products and measures. For instance, public safety officials could use it to better staff high risk areas in the near term and/or produce additional signage, and public works departments could use it to make modifications that reduce the danger of certain roadways in the medium term. Indeed, the scientists have already been able to draw early conclusions about the types of roads and scenarios that seem most risky, based upon the model’s risk assessment, and these insights could drive future infrastructure planning. Commercially, companies could use such data to help consumers avoid risky areas, or at least be more alert when traversing roads that are considered more problematic. 

Transportation and safety are two areas where machine learning can potentially have great impact, as demonstrated by this research. Even incremental improvements in safety can mean saving thousands of people from death or injury, as well as millions of dollars in damages. The authors of this research will share more of their findings this week at the 2021 International Conference on Computer Vision.