Attention-Enhanced Pedestrian Trajectory Prediction via Compressed Point Cloud Representation

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Raw point cloud data presents challenges in storage and spatial representation for multi-pedestrian trajectory prediction.
• Existing methods often struggle with high data overhead and geometric accuracy.

Purpose of the Study:

• To propose a compressed point cloud-based and attention-enhanced trajectory prediction method (CPCAE).
• To address storage overhead and improve spatial geometric representation in multi-pedestrian trajectory prediction.

Main Methods:

• A lossy compression module is designed, enhancing the Depoco framework with multi-feature extraction and coordinate decomposition.
• Spatial features are extracted from video frames using a 2D CNN; pedestrian interactions are captured via a Transformer encoder.
• Spatial and modal attention mechanisms are employed to balance feature contributions and identify key regions.

Main Results:

• The proposed compression module optimizes quality and spatial representation for point cloud data.
• The CPCAE framework demonstrates competitive trajectory prediction performance using compressed point cloud representations.
• Experimental results validate the effectiveness of the approach in both compression and prediction tasks.

Conclusions:

• Compressed point cloud representations are viable for high-performance trajectory prediction.
• The CPCAE method offers an efficient solution for multi-pedestrian trajectory prediction with reduced storage requirements.
• Attention mechanisms effectively integrate multi-modal features for enhanced prediction accuracy.