Effective lane detection to enable accident-free autonomous vehicle driving is highly essential and inevitable in the present scenario. Though conventional methods are efficient under known road conditions, their capabilities fail under complex scenarios, such as curvy lanes, occlusions, worn-out markings, and poor illumination during low light. The proposed research focuses on lane detection using Convolutional Neural Network (CNN) tailored for diverse road conditions to address the above challenges. The focus transitioned to engage VGG16, DenseNet, and U-Net architectures on the KITTI Road dataset, followed by U-Net and ENet architectures on the TuSimple dataset for semantic detection of lane markings. Further, the Indian Driving Dataset (IDD), which illustrates highly unstructured and diverse road environments was used to assess the model robustness under real-world conditions. Upon implementation, it is observed that the proposed U-Net+EfficientNet-B3 encoder model demonstrated higher adaptability to varying road structures, lighting variations, and occlusions, achieving an Intersection over Union (IoU) of 90.08%, indicating substantial improvement in lane detection performance under complex scenarios. The results highlight the potential of deep learning-based methods in enhancing autonomous vehicle safety and reliability across diverse driving environments.
 

Self Driving Vehicles,Convolutional Neural Networks (CNN).,Semantic Segmentation,U-Net,EfficientNet-B3,Road Safety