Lane Line Segmentation Based on Comparative Analysis of Three Convolutional Neural Networks SegNet, U-Net and Attention-U-Net

In the context of the rapid development of intelligent vehicles and artificial intelligence technology, this paper independently constructs 50 small sample datasets containing different scenes and different light intensities. The dataset is divided into 8:2 training and validation sets, and three different networks, SegNet, U-Net, and Attention-U-Net, are utilized to train the above images for 300 epochs and validate them respectively, which result in the following: the mIOU value of SegNet is 89.65%, the mPrecision value is 96.84%, mRecall value is 91.98%; mIOU value of U-Net is 90.05%, mPrecision value is 95.39%, mRecall value is 93.68%; mIOU value of Attention-U-Net is 90.26%, mPrecision value is 95.98%, mRecall value is 93.39%. After comprehensive analysis, it is concluded that, under the same conditions, the Attention-U-Net network for lane line segmentation has the comprehensive performance of faster convergence speed, more accurate segmentation results, better segmentation effect on the lane lines with intricate light intensity, and strong robustness, which is very suitable for the segmentation of lane lines in intelligent vehicles and is of great significance for the development of the perception field of intelligent vehicles.