PoPPL: Pedestrian Trajectory Prediction by LSTM With Automatic Route Class Clustering

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian path prediction is complex due to crowded environments and obstacles.
• Current methods primarily focus on social interactions, overlooking destination influence.
• Accurate trajectory prediction is crucial for autonomous systems and urban planning.

Purpose of the Study:

• To develop an improved pedestrian path prediction algorithm by incorporating potential destinations.
• To classify pedestrian trajectories into route classes (RCs) for pattern analysis.
• To enhance trajectory prediction accuracy in complex scenarios.

Main Methods:

• Proposed a novel algorithm, Prediction of Pedestrian Paths by LSTM (PoPPL).
• Classified pedestrian trajectories into route classes (RCs) using clustering.
• Employed a bidirectional LSTM for destination region prediction and multiple LSTM architectures for trajectory generation.
• Outputted probabilities for multiple predicted trajectories.

Main Results:

• PoPPL demonstrated superior performance compared to state-of-the-art methods on public datasets.
• Incorporating destination prediction significantly improved trajectory prediction accuracy.
• The algorithm successfully predicted destination regions and generated corresponding trajectories.

Conclusions:

• Considering potential destinations is vital for accurate pedestrian path prediction.
• The proposed PoPPL algorithm offers a robust solution for complex pedestrian movement analysis.
• This approach advances the field of trajectory prediction in dynamic environments.