Edge-Based Meta-ICP Algorithm for Reliable Camera Pose Estimation

Chun-Wei Chen,Ming-Der Shieh,Jonas Wang

doi:10.1109/access.2021.3090170

Chun-Wei Chen, Ming-Der Shieh + Show 1 more

Open Access

https://doi.org/10.1109/access.2021.3090170

Copy DOI

Abstract

Camera pose estimation is crucial for 3D surface reconstruction and augmented reality applications. For systems equipped with RGB-D sensors, the corresponding transformation between frames can be effectively estimated using the iterative closest point (ICP) algorithms. Edge points, which cover most of the geometric structures in a frame, are good candidates for control points in ICP. However, the depth of object contour points is hard to accurately measure using commercial RGB-D sensors. Inspired by the model-agnostic meta-learning (MAML) algorithm, this work proposes a meta-ICP algorithm to jointly estimate the optimal transformation for multiple tasks, which are constructed by sampled datapoints. To increase task sampling efficiency, an edge-based task set partition algorithm is introduced for constructing complementary task sets. Moreover, to prevent ICP from being trapped in local minima, a dynamic model adaptation scheme is adopted to disturb the trapped tasks. Experimental results reveal that the probability of unstable estimations can be effectively reduced, indicating a much narrower error distribution of repeated experiments when adopting re-sampled points. With the proposed scheme, the overall absolute trajectory error can be improved by more than 30% as compared to the related edge-based methods using frame-to-frame pose estimation.

Highlights

In augmented reality (AR) applications, the quality of virtual content registration highly relies on the understanding of the camera viewing angle and position
Depth edge points, which preserve most of the details in a scene structure, are good candidates used in geometric alignment
The dataset includes image sequences recorded from the Kinect V1 along with the corresponding ground-truth camera poses captured from a motion capture system

Summary

INTRODUCTION

In augmented reality (AR) applications, the quality of virtual content registration highly relies on the understanding of the camera viewing angle and position. Chen et al.: Edge-Based Meta-ICP Algorithm for Reliable Camera Pose Estimation minimizing the distance to the closest edge in the target image. Instead of adopting CNN-based edge detectors as presented in [15] to learn the edges, this work explores how to efficiently apply the concept of meta-learning strategy [18] to the edge-based ICP algorithm for improving the accuracy and reliability of pose estimation. Since the meta-learned model can be trained more effectively by using properly selected training tasks [27], this work presents schemes to construct comprehensive data sets that are employed to sample datapoints for training tasks. This paper presents an edge-based meta-ICP algorithm to jointly learn the transformation from multiple edge types. This work presents a novel ICP algorithm based on the meta-learning strategy.

DEPTH EDGE DETECTION

EXPERIMENTAL RESULTS

CONCLUSION