Representing Multimodal Behaviors With Mean Location for Pedestrian Trajectory Prediction

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian trajectory prediction is challenged by representing multimodal behaviors.
• Existing methods struggle with interpretability and performance due to complex latent space representations of interactions.

Purpose of the Study:

• To propose a novel Interpretable Multimodality Predictor (IMP) for enhanced pedestrian trajectory prediction.
• To improve interpretability and controllability of multimodal trajectory predictions.

Main Methods:

• Representing specific behavior modes by their mean locations.
• Modeling mean location distribution using a Gaussian Mixture Model (GMM) conditioned on sparse spatio-temporal features.
• Sampling mean locations from decoupled GMM components to encourage multimodality.

Main Results:

• The IMP provides interpretable predictions with semantic meaning for motion behaviors.
• It enables friendly visualization of multimodal pedestrian behaviors.
• Experiments show the IMP outperforms state-of-the-art methods in trajectory prediction accuracy.
• Controllable predictions can be achieved by customizing mean locations.

Conclusions:

• The proposed IMP effectively addresses the challenges of multimodal trajectory prediction.
• It offers a theoretically sound and practically effective approach for interpretable and controllable pedestrian behavior prediction.