To develop a prospective motion correction (MC) method for phase contrast (PC) MRI of penetrating arteries (PAs) in centrum semiovale at 7 T and to evaluate its performance using automatic PA segmentation. Head motion was monitored and corrected during the scan based on fat navigator images. Two convolutional neural networks (CNN) were developed to automatically segment PAs and exclude surface vessels. Real-life scans with MC and without MC (NoMC) were performed to evaluate the MC performance. Motion score was calculated from the ranges of translational and rotational motion parameters. MC versus NoMC pairs with similar motion scores during MC and NoMC scans were compared. Data corrupted by motion were reacquired to further improve PA visualization. PA counts (NPA ) and PC and magnitude contrasts (MgC) relative to neighboring tissue were significantly correlated with motion score and were higher in MC than NoMC images at motion scores above 0.5-0.8 mm. Data reacquisition further increased PC but had no significant effect on NPA and MgC. CNNs had higher sensitivity and Dice similarity coefficient for detecting PAs than a threshold-based method. Prospective MC can improve the count and contrast of segmented PAs in the presence of severe motion. CNN-based PA segmentation has improved performance in delineating PAs than the threshold-based method.