基於結構光的RGB-D攝影機之深度修正

Abstract

過去機器因為缺乏深度資訊無法識別物體為平面或立體。深度攝影機的問世使機器擁有準確估測物體距離的能力，可以完成過去做不到的事情，例如：可以藉由深度資訊重建三維的人臉來達到更準確的人臉辨識。但是使用深度資訊重建3D模型時，如果深度攝影機深度值之估測不夠準確，會導致建造的3D模型失真，因此為提高3D模型重建準確度，許多研究提出深度值修正方法以提高準確度。但目前已有的深度值修正方法尚有改進空間，例如：Karan提出的線性模型太過簡略，Herrera的方法在非線性最佳化時沒有包含IR攝影機內在參數的約束，因此本篇論文對Karan以及Herrera的方法進行修改，欲更進一步降低深度攝影機的量測誤差。實驗結果顯示現有的方法經過我們的修改後，Karan的方法估測誤差減少22%，Herrera的方法減少26%。

In the past, machines lacked depth information and could not accurately identify the object as a plane or a solid. The advent of the depth camera allows the machine to accurately estimate the distance of the object, and can do things that were not possible in the past. For example, people can reconstruct 3D face by using depth information to achieve more accurate face recognition. However, when using depth information to reconstruct the 3D models, if the estimation of depth camera’s depth value is not accurate enough, the 3D model of the construction will be distorted. Therefore, in order to improve the accuracy of 3D model reconstruction, many studies have proposed the depth correction methods to improve the accuracy. However, there is still room for improvement in the existing depth correction methods. For example, the linear model proposed by Karan is too brief. Herrera method does not contain IR camera intrinsic parameter constraints in nonlinear optimization. This paper modified the Karan and Herrera methods to further reduce the depth camera measurement error. The experimental results show that the measurement errors of the modified methods have been reduced, modified Karan method reduces the estimated error by 22%, and modified Herrera method reduces it by 26%.