A deep learning approach for lane marking detection applying encode-decode instant segmentation network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Safety

Background:

• Road accidents cause significant disability, death, and financial loss.
• Advanced Driver Assistance Systems (ADAS) are crucial for preventing accidents.
• Lane Marking Detection (LMD) is a fundamental ADAS technology for maintaining lane position.

Purpose of the Study:

• To develop a robust Deep Learning (DL) methodology for accurate LMD.
• To address limitations in current LMD systems caused by environmental variations (lighting, shadows, curves, obstacles).
• To introduce the Encode-Decode Instant Segmentation Network (EDIS-Net) for reliable lane detection.

Main Methods:

• Developed the EDIS-Net, a DL framework based on the E-Net architecture.
• Utilized combined cross-entropy and discriminative losses.
• Employed binary and instant segmentation for pixel extraction and Densely-Based Spatial Clustering of Application with Noise (DBSCAN) for pixel connection.

Main Results:

• Achieved 97.39% accuracy on the Tusimple dataset.
• Attained an average accuracy of 97.07% on the CalTech dataset and 96.23% on a local dataset.
• Demonstrated superior performance compared to existing LMD approaches across multiple datasets.

Conclusions:

• The proposed EDIS-Net effectively detects lane markings in challenging environmental conditions.
• The model exhibits high accuracy and reliability, making it suitable for real-world ADAS applications.
• Validated performance across three diverse datasets confirms the robustness of the EDIS-Net framework.