Pedestrian Trajectory Prediction for Real-Time Autonomous Systems via Context-Augmented Transformer Networks

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Pedestrian trajectory forecasting is crucial for autonomous vehicle (AV) safety in complex urban environments.
• Recurrent Neural Networks (RNNs) are commonly used but struggle with long sequences and real-time performance.
• Integrating contextual information improves prediction but can hinder speed.

Purpose of the Study:

• To develop an efficient and accurate framework for multi-pedestrian trajectory prediction in shared urban traffic.
• To leverage transformer networks for improved performance over traditional RNNs.
• To achieve real-time prediction capabilities without sacrificing accuracy.

Main Methods:

• A novel framework based on transformer networks.
• Fusion of multiple sensor modalities: past positions, agent interactions, and scene semantics.
• Evaluation on three real-life pedestrian datasets.

Main Results:

• Outperformed baseline approaches in both short-term and long-term prediction horizons.
• Achieved superior accuracy with lower Average Displacement Error (ADE) and Root Mean Squared Error (RMSE) compared to state-of-the-art (SOTA).
• Demonstrated real-time performance, providing 40 predictions per second (0.025s per prediction).

Conclusions:

• Transformer networks offer a more efficient and effective solution for pedestrian trajectory prediction than RNNs.
• Sensor fusion combined with transformer architecture enables robust and fast multi-pedestrian forecasting.
• The proposed framework significantly advances the capabilities of autonomous vehicles in urban environments.