Semantic scene upgrades for trajectory prediction

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Human-Motion Analysis

Background:

• Pedestrian trajectory prediction is vital for autonomous driving and robot navigation.
• Current research heavily focuses on spatial and social interactions, neglecting scene context.
• Autonomous agents need comprehensive environmental understanding, including scene dependencies.

Purpose of the Study:

• To investigate the impact of scene understanding on pedestrian trajectory prediction.
• To evaluate different encoding mechanisms for incorporating scene information.
• To determine the most effective scene representation for improving prediction accuracy.

Main Methods:

• Utilized a recurrent Variational Autoencoder (VAE) model.
• Encoded pedestrian motion history, social interactions, and semantic scene data.
• Compared trajectory prediction performance using various scene representations: fully segmented maps, semantic maps, and CNN embeddings.

Main Results:

• Fully segmented maps, providing explicit scene semantics, yielded superior performance.
• This approach outperformed other scene representation methods in trajectory prediction.
• The model successfully integrated motion, social, and scene data for enhanced predictions.

Conclusions:

• Explicit scene semantics derived from fully segmented maps are highly effective for trajectory prediction.
• Incorporating detailed scene information significantly enhances the understanding of pedestrian motion.
• This research highlights the importance of scene context in developing robust autonomous navigation systems.