Probabilistic Traffic Motion Labeling for Multi-Modal Vehicle Route Prediction.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Accurate motion prediction for traffic participants is vital for Automated Driving Systems (ADSs).
• Current multi-modal motion prediction models lack ground truth probability scores for predicted trajectories.
• Existing datasets only provide single real trajectories, limiting ML model evaluation.

Purpose of the Study:

• To introduce a novel data-based method, Probabilistic Traffic Motion Labeling (PROMOTING), for generating probable future routes and estimating their probabilities.
• To address the lack of labeled data with probability scores for multi-modal motion prediction in urban environments.

Main Methods:

• PROMOTING utilizes a two-step clustering approach on real traffic data: first, clustering intersections by road topology, then clustering similar routes within those intersections.
• Route probabilities are estimated using a frequentist approach based on historical motion data of traffic participants.
• The method focuses specifically on urban intersection scenarios.

Main Results:

• PROMOTING successfully generates probable future routes and estimates their associated probabilities for traffic participants in urban intersections.
• Evaluation using the Lyft database demonstrates the method's appropriateness for this task.
• The approach provides a valuable labeling method for creating datasets with probabilistic future motion information.

Conclusions:

• PROMOTING offers a viable solution for generating probabilistic labels for multi-modal motion prediction datasets.
• The developed method can significantly benefit the training and evaluation of Machine Learning models for ADSs.
• The open-sourced code facilitates further research and development in traffic motion prediction.