Road accidents: artificial intelligence from MIT to predict them

Predicting a car accident just before it happens: it sounds like science fiction but it is reality. Approximately 1.35 million people die from road accidents every year. Between 20 and 50 million, however, suffer serious but not fatal damage. Road accidents also have a significant economic impact: their cost is around 3% of the entire world gross domestic product (GDP). For this, at the Computer Science and Artificial Intelligence Laboratory (CSAIL) of the Massachusetts Institute of Technology (MIT) they have developed a model based on artificial intelligence capable of predicting accidents in a certain area.

The idea of ​​the researchers

The work, conducted in collaboration with the Qatar Center for Artificial Intelligence , has just been presented at the International Conference on Computer Vision 2021. The researchers' goal was to obtain an artificial intelligence system to accurately predict how many road accidents occur. occur on a certain road.

"Whether by acquiring the distribution of risk, which determines the likelihood of future accidents in all places, even without historical data, we can find safer routes, enable auto insurance companies to provide customized insurance plans based on customer driving routes. , help engineers design safer roads, and even predict future accidents. "

Songtao He, graduate student of the CSAIL

Says Songtao He, a CSAIL graduate student and first author of the study.

The artificial intelligence model developed by the MIT researchers, together with the QCAI, uses images and GPS trajectories from satellites and past accident data to train the network to predict traffic accidents. Credit: MIT / QCAI.
The model developed by the MIT researchers, together with the QCAI, uses images and GPS trajectories from satellites and past accident data to train the network. Credit: MIT / QCAI.

Deep learning as an artificial intelligence model for predicting road accidents

In reality, there are already many examples of models of this type, as the authors of the research themselves underline. This system, however, unlike the previous ones, is based on a deep learning algorithm that uses five different pieces of information as input . Deep learning literally means "deep learning". It represents that field of machine learning that uses learning techniques based on the creation of complex artificial neural networks. For their model, MIT scientists used satellite RGB images of some US cities and a database. The database contained both the GPS trajectories traveled by motorists and the accidents that occurred in cities between 2017 and 2018 .

The cities involved in the study

The cities considered in the study are Los Angeles , New York , Chicago and Boston . For each of these cities they have created a risk map based on artificial intelligence algorithms, preventing the presence of road accidents. The grid that creates the map has cells that are only 10 square meters large . This means that the forecasting system has a very high accuracy, which previous models had not been able to achieve. Such a result is achievable only if we resort to models based on complex architectures, such as the one developed by MIT researchers.

Neural network architecture developed by researchers to train artificial intelligence to predict traffic accidents. Credit: MIT / QCAI.
Artificial neural network architecture developed by researchers. Credit: MIT / QCAI.

A neural network to develop the artificial intelligence algorithm that predicts traffic accidents

As can be seen from the architecture in the figure, the incoming data are first processed to make them comparable to each other, coming from multiple sources. The implemented neural network is of the CNN type (ie, a convolution neural network) with 6 layers. Inside there is an encoder that reduces the size of the input and increases the width of the channel from 4 (three RGB channels plus a map) to 32. Another two-layer encoder brings the size of the GPS function input from 13 to 30. After that, all the feature maps obtained after each remaining block are analyzed and used to make the final map.

Heterogeneous road data to train artificial intelligence to predict accidents

Unlike other models used to predict traffic accidents, the one of the MIT researchers does not give much weight to past history. Usually, algorithms of this type were based on binary classifiers capable of assigning a certain probability to the occurrence of an accident, based exclusively on the number of accidents that occurred in that same area. The artificial intelligence algorithm behind this system, on the other hand, uses a complex set of information to predict the possibility of road accidents. GPS trajectories provide traffic information, while satellite images describe road structures: in this way high-risk areas are identified. In doing so, even though no accidents have ever been recorded in an area, it is identified as high risk based on its traffic patterns and topology .

Artificial intelligence-based neural network to predict road accidents.
Thanks to deep learning techniques it is possible to create very complex artificial neural networks.

Predicting road accidents will be a prerogative of artificial intelligence systems

This result is a giant step for systems based on artificial intelligence, in predicting areas with a high risk of road accidents. Systems of this type could be used to improve urban planning, minimizing the likelihood of accidents in a given area. The model dataset currently covers approximately 7,500 square km of the cities of LA, NYC, Chicago and Boston. However, MIT researchers argue, it can also be extended with other types of input and used for other cities.

The article Road accidents: from MIT an artificial intelligence to predict them comes from Tech CuE | Close-up Engineering .