An Investigation of Vehicle Behavior Prediction Using a Vector Power Representation to Encode Spatial Positions of...

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Predicting traffic participant behavior is crucial for safe navigation by both human drivers and automated vehicles.
• Existing methods often struggle to capture the complex interactions between multiple vehicles in dynamic environments.

Purpose of the Study:

• To develop a robust vector representation for encoding spatial information of multiple traffic participants.
• To investigate the effectiveness of this representation in predicting future vehicle positions using deep learning models.
• To create a system that can adapt its prediction strategy based on real-time driving conditions.

Main Methods:

• Utilized Vector Symbolic Architectures (VSAs) with a convolutive power encoding for a fixed-dimension vector representation of multiple objects.
• Integrated this vector representation as input for a Long Short-Term Memory (LSTM) network for sequence-to-sequence prediction of vehicle positions.
• Developed an online-learning mixture-of-experts prototype to dynamically select the best predictor based on driving situations.

Main Results:

• The proposed VSA-based vector representation effectively captures the relations and mutual influence between multiple traffic participants.
• The fixed-dimension encoding remains independent of the number of surrounding vehicles.
• The approach demonstrated competitive performance compared to benchmark LSTM models and simpler prediction methods in extensive evaluations.

Conclusions:

• The VSA-based structured vector representation offers a promising approach for multi-agent behavior prediction in automotive settings.
• The ability to combine symbolic processing with neural network learning provides flexibility and power.
• The adaptive mixture-of-experts system enhances prediction accuracy by selecting the optimal model for specific driving scenarios.