Full width home advertisement

Welcome Home

Post Page Advertisement [Top]

Deep learning allows for the prediction of traffic collisions before they occur

The modern world is a giant maze connected by layers of concrete and asphalt, which affords us the luxury of being able to navigate through it by car. Despite the fact that we have made significant strides in road-related technology, our safety measures have lagged behind. GPS allows us to fire fewer neurons as a result of map apps; cameras alert us to potentially costly scrapes and scratches; and electric autonomous cars are more fuel efficient. We continue to rely on a steady diet of traffic signals, faith, and the steel that surrounds us in order to get from point A to point B safely.

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Center for Artificial Intelligence (QCAI) developed a deep learning model that forecasts extremely high-resolution crash risk maps in order to overcome the inherent uncertainty in crashes. Crash risk maps, which are created by combining historical crash data, road maps, satellite imagery, and GPS traces, describe the expected number of crashes over a specified time period in the future, allowing for the identification of high-risk areas as well as the prediction of future accidents.


A common feature of these types of risk maps is that they are captured at much lower resolutions, typically in the hundreds of meters range, which means that critical details are glossed over as the roads become blurred. These maps, on the other hand, have a grid cell resolution of five by five meters, and the increased resolution reveals new insights: for example, the scientists discovered that a freeway road poses a greater risk than nearby residential roads, and that ramps merging into and exiting the freeway pose an even greater risk than other roads.


The research was led by Songtao He, an MIT CSAIL PhD student who is the lead author on a new paper about the research. "By capturing the underlying risk distribution that determines the probability of future crashes at all locations, we can identify safer routes, enable auto insurance companies to offer customized insurance plans based on their customers' driving trajectories, assist city planners in designing safer roads, and even predict future crashes," says He.


Automobile collisions, despite the fact they are rare, account for approximately 3% of global GDP and are the leading cause of death among children and young adults. The task of inferring maps at such a high resolution is made more difficult by the sparsity of the data. At this level, crashes occur in a sparsely distributed manner (the average annual probability of a crash in a 5 by 5 grid cell is approximately one in one thousand) and rarely repeat themselves. Previous attempts to predict crash risk have largely been "historical," with an area being considered high-risk only if a previous crash has occurred nearby.


The team's approach entails casting a wider net in order to gather critical information.. GPS trajectory patterns, which provide information about the density, speed, and direction of traffic, are combined with satellite imagery to provide information about road structures such as the number of lanes, presence of a shoulder, and presence of a large number of pedestrians. This is accomplished through the use of artificial intelligence (AI). If no crashes have been reported, then a high-risk area can be identified based on traffic patterns and topology, even if no crashes have been reported.


The model was evaluated by the scientists based on crashes and data from 2017 and 2018, as well as its predictive ability for crashes in 2019 and 2020, according to the researchers. Despite the absence of any recorded crashes or subsequent crashes, a large number of locations were identified as high-risk zones.


"Our model is able to generalize from one city to another by combining multiple clues from seemingly unrelated data sources, as shown in the example below. This represents a significant step forward in the development of general artificial intelligence because our model can predict crash maps in previously unexplored territory "Amin Sadeghi, a lead scientist at the Qatar Computing Research Institute (QCRI) and the paper's lead author, explains how the paper came to be. If there is no historical crash data, the model can infer a useful crash map that can be used for city planning and policy formulation when comparing hypothetical scenarios, according to the researchers.


The dataset covered an area of 7,500 square kilometers in the metropolitan areas of Los Angeles, New York City, Chicago, and Boston, and included data from these cities. Los Angeles was the most dangerous of the four cities studied, with the highest crash density. It was followed by New York City, Chicago, and Boston, all of which were also among the most dangerous.


"Individuals who are able to use the risk map to identify potentially hazardous road segments will be able to take proactive measures to reduce the risk associated with their travel. While some navigation apps, such as Waze and Apple Maps, include incident reporting tools, we're attempting to anticipate crashes — before they occur, in order to reduce traffic congestion "He said.

1 comment:

Bottom Ad [Post Page]