Fast-CenLaneNet: A Lightweight Instance Segmentation-Based Network for Real-Time Lane Detection

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning Architectures

Background:

• Lane detection is crucial for autonomous driving, demanding high accuracy and real-time performance.
• Instance segmentation methods offer adaptability for varying lane counts but suffer from high computational costs.
• Existing methods struggle with deployment on embedded platforms due to large network requirements.

Purpose of the Study:

• To introduce Fast-CenLaneNet, a lightweight architecture for efficient and accurate lane detection.
• To address the computational challenges of instance segmentation-based lane detection.
• To improve the real-time performance of autonomous driving systems.

Main Methods:

• Designed a lightweight backbone to minimize parameters and computational load.
• Incorporated a learnable spatial similarity attention module for enhanced feature discriminability.
• Developed multi-branch output heads utilizing Ghost convolutions for efficient feature refinement.

Main Results:

• Fast-CenLaneNet achieves a 96.40% accuracy on the TuSimple benchmark with 4.7M parameters and 9.9 GFLOPs.
• Demonstrated a significant reduction in parameters (89.1%) and a 107.5% increase in inference speed compared to CenLaneNet.
• Maintained a competitive accuracy-efficiency trade-off on TuSimple and CULane benchmarks.

Conclusions:

• Fast-CenLaneNet provides an effective solution for real-time lane detection in autonomous driving.
• The proposed architecture balances high accuracy with significantly reduced computational overhead.
• Enables the deployment of advanced lane detection capabilities on resource-constrained embedded platforms.