Safety-aware transformer-enhanced unified multi-task perception for road anomaly detection, drivable area...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Developing countries like India present unique challenges for autonomous driving systems due to unstructured roads, diverse traffic, and unclear markings.
• Existing perception systems struggle with road anomalies (potholes, waterlogging) and mixed traffic, necessitating advanced solutions.

Purpose of the Study:

• To develop an efficient and safety-aware multitask perception model for autonomous driving on Indian roads.
• To jointly perform object detection, drivable area segmentation, and lane line segmentation for enhanced environmental understanding.

Main Methods:

• A lightweight YOLOv8 backbone with a C3TR Transformer module was adapted for enhanced feature extraction in unstructured environments.
• A multitask learning approach with three task-specific decoders was employed to reduce computational overhead.
• A deterministic safety-aware refinement logic was introduced to differentiate and manage critical and navigable road hazards.

Main Results:

• The model achieved a mAP@50 of 0.877 for anomaly detection, outperforming the baseline by 1.9%.
• Drivable area segmentation achieved an mIoU of 0.962, with significant practical improvements in hazard exclusion.
• Lane line detection performance was comparable, achieving a lane line IoU of 0.447.

Conclusions:

• The proposed Safety Aware Multitask Perception model effectively addresses the complexities of autonomous driving in unstructured road conditions.
• The model demonstrates robustness across various environments and provides a significant advancement in safety-aware perception for intelligent transportation systems.
• Further research can focus on improving performance in low-contrast anomaly detection and severe road degradation scenarios.