Lineal perspective estimation on monocular images

Abstract

Depth estimation from monocular images can be retrieved from the perspective distortion. One major e ect of this distortion is that a set of parallel lines in the real world converges into a single point in the image plane. The estimation of the coordinates of the vanishing point can be retrieved directly by di erent ways, like Hough Transform and First derivative approaches. Many of them work on speci c real scene characteristics and often lead to spurious vanishing points. Technology and computational advances suggest that some re nements to these simple techniques or a combination of them could lead to more con dent vanishing point detection than modelling and developing a new complicated ones. In this paper we study the behaviour of two classical approaches, introduce them some improvements and propose a new combinational technique to estimate the location of the vanishing point in an image. The solutions will be described and compared, also through the discussion of the results obtained from their application to real images.