Occlusion-Free Road Segmentation Leveraging Semantics for Autonomous Vehicles.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning for Scene Understanding

Background:

• Deep convolutional neural networks (CNNs) are state-of-the-art for vision-based road detection.
• Challenges remain in monocular vision for autonomous driving, particularly in segmenting occluded road areas due to dynamic scenes.
• Accurate road segmentation requires understanding scene geometry and semantics.

Purpose of the Study:

• To develop a robust method for detecting occluded road areas in autonomous driving scenarios.
• To propose a lightweight and efficient neural network for occlusion-free road segmentation.
• To create a specialized dataset for training and evaluating road segmentation models under occlusion.

Main Methods:

• Introduction of the KITTI-occlusion-free road segmentation (KITTI-OFRS) dataset, derived from the KITTI dataset.
• Proposal of the occlusion-free road segmentation network (OFRSNet), a fully convolutional neural network.
• Integration of a global context module within the network's architecture for enhanced feature representation.
• Implementation of a spatially-weighted cross-entropy loss function to improve segmentation accuracy.

Main Results:

• OFRSNet effectively predicts occluded road portions by leveraging surrounding visible road layout and foreground objects.
• The global context module significantly enhances network performance.
• The spatially-weighted cross-entropy loss demonstrably increases task accuracy.
• Extensive experiments validate the approach's effectiveness across diverse datasets.

Conclusions:

• The proposed OFRSNet achieves superior performance compared to baseline models in occlusion-free road segmentation.
• The network offers an improved balance between accuracy and runtime, suitable for real-time applications in autonomous vehicles.
• The method advances the capability of autonomous systems to navigate complex and dynamic driving environments.