Dense Channel Spatial Semantic Guidance Attention U-Net for robust road lane detection in autonomous vehicles

Detection of road lanes is an essential part of autonomous vehicles, as it allows the cars to see the lanes and drive accordingly. However, it becomes complicated because of poor lane markings, various lighting, and weather conditions. This study proposes a Vision Framework for Road Lane Detection to Enable Robust Perception and Safe Navigation in Autonomous Vehicle Systems (VRLD-SNAVS). Within this framework, Dense Channel Spatial Semantic Guidance Attention U-Net (DCSSGA-UNet) is used for lane segmentation, and the Multiplayer Battle Game Optimizer (MBGO) is used for hyperparameter optimization. For consistency, the complete proposed pipeline is referred to as the DCSSGA-UNet-MBGO framework throughout the paper. In recent years, vision-based navigation driven by deep learning (DL) has emerged as a promising and widely used approach to enhance the accuracy, reliability, and robustness of perception and navigation systems for autonomous vehicles. DL-based vision systems have significantly enhanced the ability of autonomous vehicles to detect lanes, recognize obstacles, and understand complex driving scenes, thereby supporting timely and reliable perception for autonomous driving under diverse operating conditions. The proposed framework has been validated in a desktop GPU-based environment and serves as a foundation for future real-time embedded deployment. Parameter count, floating-point operations and inference latency were not measured, so the computational efficiency of the framework has still to be established. The input images are taken from the Udacity dataset and the Cracks and Potholes Dataset to enable pixel-level semantic segmentation and thus distinguish lane regions from non-lane background regions within the images. After this, these images undergo pre-processing such as resizing and pixel normalization. The proposed model performs semantic segmentation at the pixel level to classify each pixel in the image as belonging to the lane region or the non-lane (background) region. This binary segmentation method enables us to extract lane markings precisely and suppress the scene elements that are irrelevant to the lane detection task, thus improving the performance of the lane detection task under diverse road and environmental conditions. For additional performance improvement of this algorithm, the Multiplayer Battle Game Optimizer (MBGO) method is implemented to perform hyperparameter tuning of the proposed DCSSGA-UNet network. From the numerical results achieved on the selected datasets, the suggested DCSSGA-UNet-MBGO framework achieves competitive performance relative to the values reported for existing methods, with an average Dice Coefficient of 98.8% and an average Intersection Over Union (IoU) of 97.4%, in addition to a minimal Dice Loss of 1.25.