Drivable area recognition on unstructured roads for autonomous vehicles using an optimized bilateral neural network

Area of Science:

• Computer Vision
• Autonomous Systems
• Deep Learning

Background:

• Accurate drivable area recognition is crucial for autonomous vehicle safety, especially on unstructured roads with diverse shapes and ambiguous boundaries.
• Existing models often fail to balance real-time performance with high accuracy on complex road scenarios.

Purpose of the Study:

• To propose a lightweight bilateral semantic segmentation network with a dual attention mechanism for enhanced drivable area recognition.
• To improve the accuracy and real-time capabilities of autonomous driving systems on unstructured roads.

Main Methods:

• Developed a lightweight bilateral semantic segmentation network integrating Efficient Channel Attention (ECA) and Coordinate Attention (CA) mechanisms within the BiSeNet framework.
• Incorporated a residual network for efficiency and a global convolutional network (GCN) with a boundary refinement (BR) module to enhance segmentation accuracy.
• Utilized a dual-path structure to capture both channel and spatial information, addressing unclear boundaries and complex backgrounds.

Main Results:

• Achieved 93.89% Mean Intersection over Union (MIoU) and 97.32% Pixel Accuracy (PA) on the ORFD dataset.
• Demonstrated real-time performance with a speed of 62.49 Frames Per Second (FPS).
• Outperformed existing advanced real-time semantic segmentation models in accuracy and speed.

Conclusions:

• The proposed model offers a promising solution for real-time, high-accuracy drivable area recognition essential for autonomous vehicles.
• The integration of dual attention mechanisms and lightweight design ensures efficient and accurate performance in dynamic, unstructured environments.
• This advancement contributes to the safe and efficient operation of autonomous vehicles.