Toward Better Pedestrian Trajectory Predictions: The Role of Density and Time-to-Collision in Hybrid Deep-Learning...

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Human trajectory prediction is challenging due to complex pedestrian behavior, environmental factors, and varying crowd densities.
• Existing methods struggle to accurately predict movement in diverse and dense scenarios.

Purpose of the Study:

• To introduce a novel dataset capturing a wide range of pedestrian densities for trajectory prediction research.
• To develop and evaluate an improved methodology for human trajectory prediction, focusing on density variations and collision avoidance.

Main Methods:

• Collected a new dataset from the Festival of Lights in Lyon 2022, covering densities from 0.2 to 2.2 pedestrians/m².
• Proposed a novel two-stage processing approach for trajectory prediction.
• Utilized a collision-based error metric to evaluate prediction accuracy, analyzing its density-dependent performance.

Main Results:

• Density-based data classification significantly enhances predictive algorithm accuracy.
• The proposed two-stage approach outperforms current state-of-the-art methods.
• The collision-based error metric's effectiveness is dependent on crowd density, providing valuable insights.

Conclusions:

• Integrating crowd density considerations into predictive modeling is crucial for improving accuracy and collision avoidance.
• The developed methodology offers a robust framework for human trajectory prediction in complex, dynamic environments.
• This research advances the understanding of pedestrian behavior and prediction in dense urban settings.