Solving Sustainable Mobility
Cities around the world face increasing mobility and sustainability problems. Traffic congestion causes negative social and environmental effects, including air pollution. Solving sustainable mobility is a major challenge of our times. Superior predictive models empower the planning of smart cities and efficient road networks. Advanced traffic prediction is of great social and environmental value and allows capturing simple implicit rules underlying a complex system and modeling its behavior.
The Traffic4cast competition series reaches to the heart of this issue. It challenges competitors to understand complex traffic systems, with the goal of predicting their future states. We have collected high-resolution trajectories mapped to spatiotemporal cells with traffic volume, speed, and other features encoded as a multidimensional tensor.
Building on its success at NeurIPS 2019, the Traffic4cast competition is going into its second year. This year’s dataset is derived from an order of magnitude more data. It covers more cities adding diversity in scale, geography, culture, and economy. This will let us ask interesting questions about the similarities and differences in urban traffic, as well as explore master models trained on multiple cities. This year’s competition will also invite participants to incorporate data on air pollution, map topology, and points of interest.
The competition dataset includes static and dynamic information. The dynamic part is derived from GPS positions reported by a large fleet of probe vehicles and from live incidents feed. The static part contains road map properties and points of interest. We are collecting data in multiple cities through the course of a full year. For each city, the area is split into a grid of approximately 100m x 100m cells. The temporal probe data is grouped into 5 min time intervals. High-resolution GPS trajectories are thus mapped to spatiotemporal cells.
Each grid cell for every time bin has several dynamic and static attributes. The dynamic attributes encode traffic speed, volume, direction, and incidents. The static attributes describe road junctions and food, drink and entertainments venues, shopping, parking, etc. As they can affect the traffic patterns, they are expected to provide relevant information to traffic models. Additional features, such as air pollution, are in turn expected to be affected by traffic.
For each city, all the static and dynamic attributes over time are represented as a multidimensional tensor. The first dimension iterates through the 5 min time intervals, while the corresponding attributes are encoded as slices of the tensor.
Traffic4cast Challenge 2020
Building on insights from our 2019 contest, we challenge world class machine learners to understand traffic across the globe in a new way.
Highlights of what we learned from Traffic4cast 2019
The incredibly creative solutions from our 2019 contest revealed some surprising insights:
- Our simple movies provided sufficient information to make state-of-the-art traffic predictions.
- Additional traffic attributes (weather, road map properties) are of value for bonus challenges and improve more complex models.
- It may be possible to separate rules underlying traffic dynamics from city-specific forecasts (transfer learning).
Traffic4cast 2020 – new challenges and opportunities
The Traffic4cast 2020 competition aims to explore some of these findings in depth. Our core challenge is similar to Traffic4cast 2019, but covers a wider range of cities and encourages using complementary information. The core data includes dynamic features, encoding traffic volume, direction, and incidents, and static features, describing the road junctions and points of interest, such as food and drink, shopping, parking, transport, etc.
To solve the core challenge, algorithms must make predictions for all the cities, where city-dependent parameters can be applied. Participants can investigate differences and similarities in traffic patterns between the cities, and explore master models trained on multiple cities.
The bonus challenges invite participants to explore which traffic features have the greatest impact. The air quality data consist of daily measurement of several pollutants throughout a year in three cities. The goal of this bonus challenge is to predict the levels of pollutants and to find locations within each city and time intervals that are the most affected by traffic.
Interested in revolutionizing our understanding of mobility?
Join our forums now to help shape this year’s competition.
The top-ranked teams in the competition leaderboard will be honoured at NeurIPS 2020 conference and receive:
1st prize – 10 000 USD value and 12 month Research Fellowship at IARAI, covering stipend and expenses;
2nd prize – 5 000 USD value and 12 month Research Fellowship at IARAI, covering stipend and expenses;
3rd prize – 2 000 USD value and complimentary registration for NeurIPS 2020;
Prizes 1 – 6 – complimentary registration for NeurIPS 2020.
The top-ranked solutions will be published in the Proceedings of Machine Learning Research.