Object Position Estimation from a Single Moving Camera

This paper deals with the position estimation of a road sign from a single camera attached to a vehicle. We developed, implemented, and tested two mathematical approaches based on triangulation when the object annotation in the form of a bounding box is provided. We created a synthetic dataset (a simulation of a car passing by a road sign) to test the methods in a controlled environment. Additionally, the real dataset was created by recording a car trip within a town. Results on the synthetic dataset showed that the position could be estimated within 1 m accuracy. In the case of the real dataset, we measured the accuracy to be up to 4.3 m depending on the distance from the object. We performed experiments with artificial noise on synthetic data to evaluate the impact of different types of noise. Our contribution consists of two computationally inexpensive methods for object position estimation that are easy to use as they do not require calibration of parameters.