An End-to-End Trajectory Prediction Method for Unmanned Ground Vehicles via Multimodal Fusion

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Artificial Intelligence

Background:

• Unmanned Ground Vehicles (UGVs) require advanced intelligence for applications in smart cities, disaster rescue, and infrastructure inspection.
• Existing multimodal fusion architectures face challenges in dynamic feature alignment and efficient decision-planning for closed-loop autonomous navigation.

Purpose of the Study:

• To develop a high-precision, low-latency closed-loop autonomous navigation framework for UGVs.
• To enhance the collaborative optimization of multimodal fusion end-to-end architectures.

Main Methods:

• A Multi-Head Distillation Attention-based Trajectory Prediction (MDA-TP) method is proposed, integrating a BEVFormer-based multimodal fusion perception model.
• A two-stage progressive knowledge distillation framework is employed for efficient learning.
• Dynamic alignment of heterogeneous features and parallel decision-planning are utilized.

Main Results:

• The MDA-TP method achieved an Average Displacement Error (ADE) of 0.88 m and a Final Displacement Error (FDE) of 1.32 m on the NuScenes dataset, representing significant reductions.
• A collision rate of 15.2% was recorded with low latency (46 ms) and a manageable parameter count (42.6 M).
• Real-world validation demonstrated Euclidean deviation within 0.5 m, improved speed prediction (51.5% MSE reduction), and enhanced route completion (97.26%).

Conclusions:

• The developed framework offers a practical solution for autonomous driving in complex environments.
• The proposed MDA-TP method effectively optimizes multimodal fusion for enhanced UGV perception and navigation.
• The study provides strong evidence for the efficacy of attention distillation and knowledge transfer in improving autonomous navigation performance.