Real-time estimation of the 3D transformation between images with large viewpoint differences in cluttered environments

TitleReal-time estimation of the 3D transformation between images with large viewpoint differences in cluttered environments
Publication TypeConference Paper
Year of Publication2017
Authorsvan de Wouw, DWJM, Pieck, MAR, Dubbelman, G, de With, PHN
Conference NameElectronic Imaging 2017
Date Published[In press]
PublisherIS&T
Conference LocationSan Francisco
Abstract

Accurate image registration of images acquired from a moving vehicle, is a challenging task for images with viewpoint differences in cluttered environments, especially in the case of large lateral displacements between camera poses. This work focuses on estimating an accurate 3D transformation in real-time, which is used to register such images. The main challenges are significant image appearance differences, which originate from lateral displacements, inconsistencies in our 3D model and achieving real-time execution. To this end, we propose a feature-based method using a single synthesized view, which copes with image
appearance differences. The 3D transformation is estimated using an EPnP refinement to minimize the influence of inconsistencies in the 3D model. Additionally, we exploit an estimate of the lateral displacement together with GPS/IMU information to construct the synthesized view and dynamically adapt the transformation thresholds. We demonstrate in our experiments that our method achieves over 95% transformation accuracy for lateral displacements up to 350 cm, while still achieving 85% accuracy at displacements of 530 cm. Additionally, with a running time of 100 milliseconds, we achieve real-time execution as a result of efficiency optimizations and GPU implementations of time critical components.

AttachFile: