BIPOOLNET: An advanced UNet architecture for enhanced lane detection in autonomous vehicles.

This article focuses on BIPOOLNET, a novel encoder-decoder neural network architecture designed to improve lane detection accuracy for autonomous vehicles, particularly on curved roads where traditional AI models often underperform. BIPOOLNET integrates max pooling, average pooling, and a Feature Pyramid Enhancement Module (FPEM) to refine feature extraction, and employs Long Short-Term Memory (LSTM) networks to maintain temporal consistency in lane prediction. Evaluated on the TuSimple dataset—a collection of 6,408 high-resolution images from diverse US highway conditions—BIPOOLNET achieved a 98.45% accuracy, a 98.17% F1-score, and notably low false positive (1.84%) and false negative (1.09%) rates, outperforming several state-of-the-art lane detection methods. The architecture also demonstrates robustness across varied driving scenarios, including different weather, lighting, traffic, and road curvature conditions, while offering improved computational efficiency and lower deployment costs. These results suggest BIPOOLNET’s potential to enhance the safety and reliability of autonomous driving systems within complex urban environments.