Keypoint based automatic image orientation and skew investigation on tie points

Abstract

Purpose

The relative orientation (RO) is an important step on photogrammetric processes of stereoscopic images. The relationship between the stereoscopic images is constructed by tie (conjugate) points. Many automatic tie point selection methods have been introduced by photogrammetry community so far. The scale invariant feature transform (SIFT) and speeded‐up robust features (SURF) are frequently used for automatic tie point selection from stereoscopic images. However, any research has been performed related to RO errors (y‐parallaxes) on SIFT and SURF extracted tie points. The purpose of this paper is to compute errors on tie points and investigate their distributions on the model area in terms of size.

Design/methodology/approach

The experimental studies were performed on an historical building as it enables more tie points for investigation. While a couple of the stereoscopic images include rich details, the other has poor details. The image orientation and tie point selection accuracy were evaluated by root mean square and y‐parallaxes, respectively. The relationship between y‐parallaxes of tie points and their distances from centre of the images were investigated.

Findings

SIFT and SURF have a large number of tie points according to manual method. The y‐parallaxes on tie points have uniform distribution for two methods. There are relations between the precision of the SIFT and SURF keypoints and their distances from the centre of the image. Moreover, the accuracy of the RO and size of the y‐parallaxes on tie points depend on matching accuracy of the keypoints. Furthermore, although there are a few tie points that have large y‐parallax especially by the SURF, RO could be performed with high accuracy thanks to numerous tie points.

Originality/value

Stereoscopic images of close‐range photogrammetry have different scale and rotations, unlike aerial photogrammetry. Manual selection of tie points is time consuming and tedious. Furthermore, if the measurement surface has no implicit entities, enough tie points from the images cannot be selected by manually. However, tie point selection can be performed by SIFT and SURF automatically, even if there are scale, noise and rotation between the images.

Citation

Altuntas, C. (2013), "Keypoint based automatic image orientation and skew investigation on tie points", Kybernetes, Vol. 42 No. 3, pp. 506-520. https://doi.org/10.1108/03684921311323725