This AI Research Introduces FollowNet: A Comprehensive Benchmark Dataset for Car-Following Behavior Modeling

Following another vehicle is the most common and basic driving activity. Following other cars safely lessens collisions and makes traffic flow more predictable. When drivers follow other vehicles on the road, the appropriate car-following model represents this behavior mathematically or computationally.

The availability of real-world driving data and developments in machine learning have largely contributed to the boom of data-driven car-following models during the past decade. Models that rely on data to follow a vehicle include neural networks, recurrent neural networks, and reinforcement learning. Several limitations exist, though, in the current body of research, as follows:

• To begin, car-following models are not yet well evaluated because of the absence of standard data formats. Despite the availability of public driving datasets like NGSIM and HighD, it is difficult to compare newly suggested models’ performance with existing ones due to the lack of standard data formats and evaluation criteria for car-following models. 
• Secondly, limited datasets in current studies make it impossible to accurately portray car-following behavior in mixed traffic flows. Modeling car-following behavior with small datasets that don’t consider autonomous vehicles has been the main emphasis of prior research, which comes at a time when both human-driven and autonomous vehicles are sharing the road.

To solve these problems and create a standard dataset, a new study by the Hong Kong University of Science and Technology, Guangdong Provincial Key Lab of Integrated Communication, Tongji University, and the University of Washington released a benchmark known as FollowNet. They used consistent criteria to extract car-following events from five publicly available datasets to establish the benchmark. The researchers executed and evaluated five baseline car-following models within the benchmark, encompassing conventional and data-driven methodologies. They set the first standard for such behavior using uniform data formats to facilitate the creation of car-following models. It might be difficult to handle diverse data structures and frameworks from different datasets, but their standardized car-following benchmark considers that.

Two conventional and three data-driven car-following models—GHR, IDM, NN, LSTM, and DDPG—are trained and evaluated using the benchmark. Five popular public driving datasets—HgihD53, Next Generation Simulation (NGSIM)54, Safety Pilot Model Deployment (SPMD)55, Waymo56, and Lyf57—comprise car-following events that comprise the proposed benchmark. The researchers look at several datasets for patterns of car-following behavior and basic statistical information. The results show the use of consistent metrics to assess the baseline models’ performances. In particular, Waymo and Lyf datasets show that car-following occurrences occur in mixed-traffic situations. The researchers did not include events with more than 90% static duration.

Collisions are still possible, even when data-driven models achieve lower MSE of spacing than classical models. The development of car-following models with zero collision rates and fewer spacing errors is desirable. It would be beneficial to include collision avoidance capabilities to make data-driven models more practical and safe for use in the real world. All cars are assumed to exhibit consistent and similar behavior patterns in the proposed benchmark. Realistically, though, driving habits can differ significantly depending on the driver, the vehicle, and the traffic conditions. As a result, creating adaptable algorithms and representative datasets that cover a range of driving styles, behaviors, and traffic situations is essential for including driving heterogeneity in car-following models.

The researchers suggest that future datasets must incorporate additional features to improve the performance and realism of car-following models even further. For instance, a more complete picture of the road environment may be achieved by adding traffic signals and road conditions data. The algorithms may also account for complicated relationships and provide better predictions if they integrate data about nearby vehicles and their activities. Future datasets will be able to better reflect real-world driving scenarios with the use of these extra data sources, which will allow for the creation of car-following algorithms that are both robust and effective.

Check out the Paper and Github. All credit for this research goes to the researchers of this project. Also, don’t forget to join our 33k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..