A deep learning autonomous vehicle machine vision system for road classification using LGAN-based synthetic IPM data and CNN

Despite significant advances in autonomous vehicle (AV) technology, reliable lane detection remains a critical challenge, particularly in complex urban environments. Traditional image processing techniques often fail under poor lighting, occlusions, and irregular lane conditions, while deep learning approaches suffer from data scarcity, class imbalance, and insufficiently calibrated vision systems. This study addresses these issues by developing an integrated framework to enhance road curvature classification using deep learning, synthetic data generation, and accurate camera calibration. The primary objectives are threefold: (1) to establish an optimal camera calibration methodology for accurate inverse perspective mapping (IPM); (2) to design and implement a lightweight Generative Adversarial Network (LGAN) for generating realistic synthetic IPM images; and (3) to develop a robust Convolutional Neural Network (CNN) model for classifying road types straight, left curve, and right curve. The methodology begins with precise calibration of a Point Grey camera (CM3-U3-31S4C-CS) to generate distortion free bird’s-eye view (BEV) images essential for lane feature extraction. An LGAN, enhanced with an Exponential Moving Average (EMA) strategy, is then trained to generate high-fidelity synthetic BEV images that augment limited real-world datasets and address class imbalance. Quantitative evaluations using PSNR, SSIM, FID, and Inception Score validate the realism of the generated images. The final stage involves training CNN architectures (AlexNet, MobileNetV2, EfficientNet-B0) on the combined dataset, using SGD and ADAM optimizers, to assess classification performance across various training. Experimental results demonstrate that EfficientNet-B0, when optimized with SGD and trained for 25 epochs, consistently achieves superior validation accuracy (up to 99.74%) and excellent AUC-ROC scores, outperforming baseline models in generalization and robustness. The study concludes that the combination of rigorous camera calibration, high-quality synthetic image generation via LGAN, and modern CNN architectures forms a scalable and effective solution for enhancing lane classification in AV systems.