In this paper, we propose a method for estimating the position of non-line-of-sight (NLOS) obstacles for an ego vehicle by observing their reflected images on nearby reflective surfaces using a stereo vision camera. The study assumes an application scenario in which reflections are obtained from the side panel of a vehicle traveling in an adjacent lane. First, to evaluate the feasibility of this application, we measured the frequency at which the virtual image of a vehicle ahead of the preceding vehicle appears on the reflective surface of an adjacent vehicle. Next, we conducted experiments to validate the basic principle of estimating the position of the virtual (reflected) image and subsequently recovering the real-world position of the obstacle by geometrically folding the virtual image across the reflective surface. In 1/6-scale experiments, the proposed method estimated obstacle positions with an average error of 4.3 cm for targets located 1.7 m to 3.7 m ahead of the camera. These results suggest that, under equivalent real-world conditions, NLOS obstacles at distances of approximately 10 m to 22 m could potentially be estimated with an accuracy of 0.03 m to 0.26 m. Even under variations in the orientation of the reflective surface, the RMS error remained stable at approximately 45 mm, indicating that the proposed method exhibits a certain degree of robustness against changes in the relative angle between the stereo camera and the reflective surface. • A NLOS obstacle localization method using reflected images from adjacent vehicles is proposed. • Reflection frequencies were evaluated through both real-world observation and microscopic traffic simulation. • A position estimation model accounting for reflective surface angle was developed and validated. • Results suggest higher accuracy at small incident angles and denser traffic conditions. • Although tested on mirrors, the method is promising for real-world reflective surfaces such as painted vehicle panels and window glass.
Sheng et al. (Tue,) studied this question.