Even a small improvement in our ability to predict when and where the next incident will take place could make a major difference for traffic safety. The importance of incident prediction will be magnified when autonomous vehicles will become the norm.
Traffic flow can be affected by a multitude of events, ranging from minor local disturbances (an object on the road) to major accidents blocking the entire traffic flow for hours. Some are the result of random external causes, but many are influenced by known factors, such as speed, vehicle density or weather conditions. The influence is complex and highly non-linear and it is probably impossible to precisely predict incidents for a specific time and place.
We focus on the interpretation of patters in the underlying influencing factors and on detecting simultaneous conditions that could build up and lead to incidents. We use road, weather and infrastructure sensors to feed Deep Learning (DL) Neural Networks and learns patterns that in the past have led to incidents in a certain time window in the future. The method provides a significant improvement to our ability to spot the next incident in time-space compared to pure statistical methods.
The frequency of incidents is usually linked to traffic features, the most obvious one being the speed-intensity profile, which has predictable time patterns. Intersections and other road features contribute to the creation of hot spots. This time-space concentration of incidents provide the most basic form of incident prediction.
Of the many factors potentially contributing to incidents, some have a known relationship with incidents. On average and at large scales these relationships can be measured. However, the relationship between incidents and condition patterns for a specific road section at a specific time are very hard to ping down. To make this possible one needs to identify very distinct signal patterns within high-resolution time and space data so that conditions can be associated to specific incidents.
An additional challenge is the small number of incidents that can be used for training an algorithm. While road incidents cause thousands of deaths and injuries every year, their density in time and space is - fortunately - low. This makes machine learning at a local level very challenging.
In this project we use data from road traffic sensors and weather sensors from the Netherlands. The traffic sensors measure speed and flow at regular locations on the road network, every minute for each lane. The incident data is based on incident logs of local authorities and captures events and their description on the entire road system for multiple years. The DL Neural Network is trained for specific sections of the highways of various lengths (e.g. 10km or 20km). The models can be tuned to predict for various future time windows, such as 15 minutes, 30 minutes or longer in the future.
Compared to pure statistical methods, the DL Neural Network can uplift prediction accuracy of about 40% on a training set of ± 300 incidents over 2 years of data. We trained different network to operate for a wide range of time windows in the future (from 15-30 minutes to 2 hours). The methods are robust and can be developed for a wide range of operational situations. We tested about 50 models, broadly clustered into five classes of (hyper)parameters for Fully Connected (FC), Convolutional (CNV) and Recurrent Neural Networks (RNN) designs. Generally speaking, FC can be applied across all situations, while CNV and RNN provide superior outcomes on specific local settings. The image below compares the outcome of the neural network predictions with an actual incident.
© CS RESEARCH | Additional information: email@example.com