Sparse Trajectory Prediction.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Pedestrian trajectory prediction is vital for safe robotic decision-making.
• Existing methods often sacrifice speed for accuracy due to computational complexity.
• Real-time performance is a critical but often overlooked requirement.

Purpose of the Study:

• To develop a pedestrian trajectory prediction model that achieves both high accuracy and real-time speed.
• To address the accuracy-speed trade-off in current prediction models.
• To introduce an efficient principle leveraging sparse structures for global effects.

Main Methods:

• Proposed a Sparse Trajectory Prediction (STP) model within a transformer-style encoder-decoder framework.
• Implemented irregular interaction in the encoder to reduce computational complexity while maintaining global interaction.
• Utilized an early-sparsity strategy in the decoder to generate shared sparse motion modes for efficient multimodal trajectory prediction.

Main Results:

• Achieved state-of-the-art performance on four benchmark datasets.
• Significantly improved prediction speed by approximately 100x-150x compared to previous methods.
• Demonstrated the model's ability to maximize both prediction accuracy and speed.

Conclusions:

• The STP model effectively balances accuracy and speed for real-time pedestrian trajectory prediction.
• Leveraging sparse structures is a viable strategy for achieving global effects efficiently.
• The proposed method satisfies the demanding real-time requirements of intelligent robotic systems.