TrajectoryNAS: A Neural Architecture Search for Trajectory Prediction.

Area of Science:

• Autonomous Driving Systems
• Computer Vision
• Artificial Intelligence

Background:

• Trajectory prediction is crucial for safe autonomous navigation, with LiDAR data offering superior 3D environmental perception compared to 2D cameras.
• Current LiDAR-based trajectory prediction methods face challenges with computational cost, slow inference, and accuracy limitations due to inefficient architectures.

Purpose of the Study:

• To introduce TrajectoryNAS, a novel neural architecture search (NAS) method for developing efficient and accurate LiDAR-based trajectory prediction models.
• To optimize the complete end-to-end trajectory prediction pipeline, including object detection and tracking, by addressing neural architecture design.

Main Methods:

• TrajectoryNAS employs metaheuristic algorithms to systematically search and optimize neural network architectures for trajectory prediction.
• A multi-objective energy function is introduced, balancing prediction accuracy and computational efficiency (latency).
• The method considers all stacked components of the prediction pipeline to minimize accuracy loss and overhead latency.

Main Results:

• TrajectoryNAS significantly enhances the performance of autonomous driving systems by improving trajectory prediction.
• Experimental results on the NuScenes dataset show TrajectoryNAS achieves at least 4.8% higher accuracy and 1.1x lower latency than competing methods.
• The optimized model demonstrates superior efficiency and accuracy in predicting surrounding object trajectories.

Conclusions:

• TrajectoryNAS represents a significant advancement in LiDAR-based trajectory prediction for autonomous vehicles.
• The NAS approach effectively addresses the limitations of handcrafted architectures, leading to more robust and performant systems.
• The developed model offers a promising solution for enhancing the safety and reliability of autonomous driving technology.