A Direct Method for Robust Model-Based 3D Object Tracking from a Monocular RGB Image
This paper proposes a novel method for robust 3D object tracking from a monocular RGB image when an object model is available. The proposed method is based on direct image alignment between consecutive frames over a 3D target object. Unlike conventional d
- PDF / 4,147,702 Bytes
- 12 Pages / 439.37 x 666.142 pts Page_size
- 118 Downloads / 298 Views
Abstract. This paper proposes a novel method for robust 3D object tracking from a monocular RGB image when an object model is available. The proposed method is based on direct image alignment between consecutive frames over a 3D target object. Unlike conventional direct methods that only rely on image intensity, we newly model intensity variations using the surface normal of the object under the Lambertian assumption. From the prediction about image intensity in this model, we also employ a constrained objective function, which significantly alleviates degradation of the tracking performance. In experiments, we evaluate our method using datasets that consist of test sequences under challenging conditions, and demonstrate its benefits compared to other methods. Keywords: Pose estimation · Object tracking image alignment · Motion model
1
·
Model-based
·
Direct
Introduction
3D tracking (or 6D pose estimation) of target objects is a crucial issue in computer vision, robotics, and augmented reality. Over the last decade, numerous methods have been proposed and successfully demonstrated for 3D object tracking. Despite that, achieving accurate, robust, and fast tracking is still challenging in everyday environments where there exists a large range of 3D objects under various backgrounds, illuminations, occlusions, and motions. In early methods, feature points have prominently been used to handle pose estimation problems of 2D/3D objects [27,30], but such feature-based methods require that the objects have sufficient texture on their surfaces. For poorly textured 3D objects, strong edges have been popular and are still promising in many industrial applications [7,11]. However, they are often troublesome against heavy background clutter due to the nature of edge property. As recent RGBD cameras enable to obtain more dense information about 3D scenes including objects, Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-49409-8 48) contains supplementary material, which is available to authorized users. c Springer International Publishing Switzerland 2016 G. Hua and H. J´ egou (Eds.): ECCV 2016 Workshops, Part III, LNCS 9915, pp. 551–562, 2016. DOI: 10.1007/978-3-319-49409-8 48
552
B.-K. Seo and H. Wuest
Fig. 1. Tracking results using the proposed method under challenging conditions (green lines visualize object models projected on images with estimated poses). (Color figure online)
RGBD-based methods have been boosted to tackle challenging pose estimation problems [15,18,29]. Nevertheless, RGBD cameras have several issues need to be considered: depth information is quite noisy and only available within limited ranges with material difficulties (such as specular and transparent materials). Moreover, they are not commonly supported yet in real application domains, compared to RGB cameras. On the other hand, direct methods have been attractive because they allow that rich information in an image can be contributed to pose estimation, instead of being limited by local features [1,
Data Loading...