Uncertainty weighted multi task learning for robust traffic scene semantic understanding.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Perception systems in autonomous vehicles struggle with degraded sensor data due to adverse weather, occlusion, and asynchronous sampling.
• Robust semantic understanding of traffic scenes is crucial for safe navigation and decision-making.

Purpose of the Study:

• To develop a novel framework, uncertainty-weighted multi-task learning (UW-MTL), for robust perception in challenging traffic scenarios.
• To improve performance on key tasks like 3D object detection, BEV semantic segmentation, and trajectory prediction.

Main Methods:

• Differentiable multi-source spatiotemporal alignment to fuse data from camera, LiDAR, radar, and IMU into a Bird's-Eye View (BEV) sequence.
• A hybrid backbone combining a Mixture of Experts Transformer and a spatiotemporal graph neural network for balanced global and local feature learning.
• Evidential prediction heads to explicitly output task confidence and uncertainty, enabling stable joint optimization via soft-temperature weighting and gradient conflict resolution.

Main Results:

• UW-MTL consistently outperforms existing methods like BEVFusion and UniAD on the nuScenes benchmark.
• Significant performance gains are observed in 3D object detection, BEV semantic segmentation, and trajectory prediction.
• The framework shows pronounced improvements in challenging conditions, including long-range detection, heavy occlusion, and low-visibility scenarios.

Conclusions:

• The proposed UW-MTL framework offers a robust solution for semantic understanding of traffic scenes under adverse conditions.
• Explicitly modeling uncertainty improves the reliability and performance of multi-task learning in autonomous driving perception.
• UW-MTL demonstrates superior performance, particularly in scenarios where traditional methods falter.