Camera trajectory recovery for image-based city street modeling

A semi-automatic image-based approach for city street modeling was proposed, which takes two types of images as input. One is an orthogonal aerial image of the area of interest and the other is a set of street-view spherical panoramic images. This paper focuses on the accuracy enhancement of camera trajectory recovery, which is crucial in registering two types of image sources. Scale-invariant Feature Transform feature detection and matching methods were employed to identify corresponding image points between each pair of successive panoramic images. Due to the wide field-of-view of spherical panoramic images and high image recording frequency, the number of resultant matches is generally very large. Instead of directly applying RANSAC which is very time consuming, we proposed a method to preprocess those matches. We claim that the majority of incorrect or insignificant matches will be successfully removed. Several real-world experiments were conducted to demonstrate that our method is able to achieve higher accuracy at estimating camera extrinsic parameters, and would consequently lead to a more accurate camera trajectory recovery result.