Soft + Hardwired attention: An LSTM framework for human trajectory prediction and abnormal event detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional pedestrian flow analysis relies on hand-crafted features, limiting adaptability.
• Deep learning and recurrent neural networks show promise in learning features and sequence-to-sequence tasks.

Purpose of the Study:

• To develop a novel method for predicting future pedestrian motion using past behavior and neighborhood data.
• To leverage deep learning, specifically combined attention mechanisms, for enhanced trajectory prediction.

Main Methods:

• A novel combined attention model integrating soft and hard-wired attention to map local trajectory information to future positions.
• Application of the model to predict pedestrian trajectories based on historical movement data of individuals and their neighbors.
• Testing the model on public surveillance datasets to evaluate its navigational prediction capabilities.

Main Results:

• The proposed method outperforms current state-of-the-art approaches in predicting pedestrian trajectories.
• The combined attention model effectively handles complex scenarios with numerous neighbors.
• The architecture demonstrates direct applicability to abnormal event detection without feature engineering.

Conclusions:

• The novel combined attention model offers a powerful and adaptable solution for pedestrian trajectory prediction.
• This approach advances the field by integrating sophisticated attention mechanisms for improved real-world applicability.
• The method provides a foundation for enhanced surveillance analytics, including anomaly detection.