The on-board head-up display (HUD) system produces virtual images through the reflection of the windshield. Due to the aberration of the optical system, actual virtual images are not distributed in an ideal planar pattern. Existing measurement methods for HUD virtual images treat such images as planar ones perpendicular to the optical axis, which is obviously inconsistent with the actual situation. To understand the characteristics of HUD virtual images, this paper proposes a method of determining the three-dimensional (3D) coordinates of the HUD virtual image points within the eyebox range by performing ray tracing, so as to solve the position of the minimum circle of confusion. HUD virtual image shapes observed at different positions are thereby obtained. Then, an HUD virtual image measurement model based on binocular vision is further developed, and the effects of binocular baseline and camera parameters on the measurement error are analyzed. Finally, experiments are carried out to measure the 3D coordinates of actual HUD virtual images. The measured virtual image shapes are basically consistent with the simulation results, with an overall average relative error of 2.5%. The proposed method can be further applied to the 3D coordinate measurement of virtual images produced by HUD systems such as augmented-reality HUD systems.