Decouple Ego-View Motions for Predicting Pedestrian Trajectory and Intention

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Pedestrian trajectory prediction is crucial for autonomous driving safety in urban settings.
• Egocentric-view algorithms simplify 3D reconstruction but struggle with coupled vehicle-pedestrian motion analysis.
• Existing models inadequately address the complexities of predicting pedestrian intent from an egocentric viewpoint.

Purpose of the Study:

• To develop a novel egocentric pedestrian trajectory prediction approach.
• To improve the understanding of coupled vehicle and pedestrian motion from an egocentric perspective.
• To enhance the accuracy and flexibility of pedestrian intention and action prediction.

Main Methods:

• A two-tower model architecture with multi-modal inputs.
• One tower processes ego-vehicle actions and initial pedestrian position; the other processes prior pedestrian trajectory and visual features.
• An action-aware loss function decomposes trajectory predictions into ego-vehicle and pedestrian motion components.

Main Results:

• The proposed model outperforms existing algorithms in ego-view pedestrian trajectory prediction accuracy.
• The action-aware loss function effectively decomposes motion components, even with combined ego-view data.
• Experiments on three benchmark datasets validate the model's superior performance.

Conclusions:

• The novel two-tower approach with an action-aware loss function significantly advances egocentric pedestrian trajectory prediction.
• This method provides a more flexible and accurate estimation of pedestrian actions and intentions.
• The findings contribute to safer and more reliable autonomous driving systems.