A duplex transform heterogeneous feature fusion network for road segmentation.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning Architectures

Background:

• Road detection in autonomous driving is challenging due to boundary fuzziness, occlusion, and glare.
• Effective road segmentation requires global context and prioritized feature representation.

Purpose of the Study:

• To introduce DTRoadseg, a novel duplex Transformer-based network for robust road segmentation.
• To enhance road detection performance by addressing challenges like occlusion and glare.

Main Methods:

• DTRoadseg utilizes a duplex encoder for heterogeneous feature extraction from RGB and depth data.
• A multi-source Heterogeneous Feature Reinforcement Block (HFRB) fuses features using self-attention and feature emphasis.
• A Transformer decoder and boundary loss function optimize segmentation accuracy.

Main Results:

• DTRoadseg achieved superior performance on the KITTI road dataset compared to state-of-the-art methods.
• The network attained an average accuracy of 97.01% and a Recall of 96.35%.
• DTRoadseg processes images at a speed of 0.09 seconds per picture.

Conclusions:

• DTRoadseg effectively addresses road detection challenges through its novel network architecture and feature fusion strategy.
• The proposed method demonstrates significant improvements in accuracy, recall, and processing speed for autonomous driving applications.