Social-Aware Pedestrian Trajectory Prediction via States Refinement LSTM

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian trajectory prediction is crucial for autonomous systems.
• Modeling social interactions is a key challenge due to their complexity.
• Existing LSTM-based methods suffer from delayed observation of neighbor intentions.

Purpose of the Study:

• To propose a novel data-driven states refinement LSTM network (SR-LSTM).
• To enable real-time utilization of neighbor intentions through a message passing framework.
• To enhance pedestrian trajectory prediction by modeling social interactions more effectively.

Main Methods:

• Developed a states refinement LSTM network (SR-LSTM) with a message passing framework.
• Implemented a self-updating mechanism for joint state refinement of all participants.
• Designed a social-aware information selection module with element-wise motion gate and pedestrian-wise attention.
• Utilized spatial-edge LSTMs to model pedestrian interactions as a graph.

Main Results:

• The SR-LSTM model demonstrated superior performance in pedestrian trajectory prediction.
• Effectiveness validated on four widely used datasets: ETH, UCY, PWPD, and NYGC.
• The proposed methods significantly improved the modeling of social interactions.

Conclusions:

• The SR-LSTM effectively addresses the limitations of previous methods by incorporating current neighbor intentions.
• The joint state refinement and social-aware information selection enhance prediction accuracy.
• The model offers a promising approach for complex social behavior modeling in trajectory prediction.