Atten-LTC-Enhanced MoE Model for Agent Trajectory Prediction in Autonomous Driving

Area of Science:

• Autonomous Driving Systems
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving relies on accurate prediction of agent trajectories.
• Understanding diverse and unpredictable behaviors of vehicles and pedestrians is crucial.
• Current methods face challenges in handling complex spatio-temporal interactions.

Purpose of the Study:

• To address the complexities of Single-Agent Trajectory Prediction (SATP) and Multi-Agent Trajectory Prediction (MATP).
• To propose an innovative and extensible model for trajectory prediction in autonomous driving.
• To enhance computational efficiency and prediction accuracy.

Main Methods:

• Development of the Atten-LTC-MoE model, integrating attention mechanisms, Liquid Time-Constant (LTC) networks, and Mixture of Experts (MoE).
• Utilizing spatio-temporal features, agent data fusion, and vectorization techniques.
• Leveraging lane and agent vectorization for improved data representation.

Main Results:

• The Atten-LTC-MoE model demonstrated superior performance on the Argoverse and Interaction datasets.
• Achieved state-of-the-art results in minADE₆ and minFDE₆ metrics.
• Showcased significant improvements in agent trajectory prediction accuracy and computational performance.

Conclusions:

• The proposed Atten-LTC-MoE model offers a general and extensible solution for SATP and MATP.
• The model effectively captures complex spatio-temporal dynamics for reliable trajectory prediction.
• This advancement contributes to the practicality and safety of autonomous driving technology.