Adaptive Multiple-Attribute Scenario LoRA Merge for Robust Perception in Autonomous Driving

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Autonomous driving perception models are primarily trained on clear, daytime data.
• Performance degrades significantly under rare conditions, especially combined weather and lighting (e.g., night × rainy).
• This limitation poses a challenge for robust real-world deployment.

Purpose of the Study:

• To develop a parameter-efficient fine-tuning (PEFT) framework for autonomous driving perception models.
• To enhance model performance under diverse and challenging environmental conditions, particularly joint weather-lighting scenarios.
• To enable dynamic adaptation of model components based on encountered conditions.

Main Methods:

• Proposed a PEFT framework utilizing dynamically applied, lightweight, scenario-specific Low-Rank Adaptation (LoRA) experts.
• Implemented an adaptive pipeline to select appropriate LoRA experts based on environmental cues.
• Validated the approach on a unified semantic segmentation benchmark (MUSES, BDD100K, Cityscapes) across six day/night × weather scenarios.

Main Results:

• Achieved improvements of up to 3.23 mIoU points over a strong baseline in single-attribute settings.
• In data-scarce multiple-attribute scenarios, merged LoRA experts outperformed the baseline by up to 5.99 points.
• Demonstrated effective generalization and adaptation capabilities across compounded environmental conditions.

Conclusions:

• The proposed PEFT framework with adaptive LoRA experts significantly enhances autonomous driving perception.
• The method effectively addresses performance limitations in challenging joint weather-lighting conditions.
• This approach offers a parameter-efficient solution for improving model robustness and generalization in diverse real-world scenarios.