Lane detection is a fundamental component of intelligent vehicle systems, particularly in assisted and autonomous driving technologies. By providing critical information about road lane markings through environmental perception, it significantly enhances driving safety. This study employs the BLD-YQ01 robot–a miniature intelligent vehicle simulation platform–to investigate lane detection methods within an intelligent transportation sandbox. The goal is to simulate and analyze real-world lane detection techniques in a scaled environment. The research includes four key steps: constructing a custom lane detection dataset, developing two semantic segmentation models based on the U-Net architecture with VGG-16 and ResNet-50 encoders, training the models, and analyzing their performance. Experimental results show that the model with a VGG-16 encoder achieves a pixel accuracy (PA) of 99.28%, mean pixel accuracy (MPA) of 95.81%, mean intersection over union (MIoU) of 91.93%, and an inference speed of 11.46 FPS. The model with a ResNet-50 encoder reaches a PA of 99.28%, MPA of 95.38%, MIoU of 91.81%, and an inference speed of 22.07 FPS.