基於物件偵測之多平面物體姿態估測系統

Abstract

隨著終端裝置技術的快速發展，擴增實境（AR）和混合實境（MR）正受到越來越多的關注。估測平面物體六自由度（6-DoF）姿態是這些應用的關鍵之一。多年來，人們進行了大量研究以估測單一已知平面目標的姿態。然而，在實際應用中可能會遇到多個平面目標，並且根據具體應用可能會替換為不同的目標。我們的目標是設計一個通用的平面物體姿態估測系統，能夠在最小的訓練要求下高效地估測多個任意平面物體的姿態。 在這篇論文中，我們提出了一個基於物件偵測和直接姿態估測算法的多平面物體姿態估測系統，稱為DetDPE。我們只使用少量合成影像數據對現成的物件偵測器進行微調。物件偵測器識別用於姿態估測的平面物體，並顯著降低估測算法的複雜度。我們的DetDPE系統展現了高效性，同時保持了原始直接姿態估測算法的高精準度。此外，我們將物件偵測方法與基於特徵的姿態估測算法結合。結果顯示這種方法確實可以提升基於特徵的算法的性能。因此，我們可以將此物件偵測方法視為一個可應用於各種姿態估測算法的框架。

With the rapid growth of technologies on edge devices, Augmented Reality (AR) and Mixed Reality (MR) are gaining more and more attention. Estimating the six degrees of freedom (6-DoF) planar object pose is one of the keys to these applications. Throughout the years, numerous research studies have been conducted to estimate the pose of a single known planar target. However, in real-world applications, the presence of multiple planar targets may be encountered, and they might be replaced with different ones depending on the specific application. Our objective is to design a general planar object pose estimation system capable of estimating the poses of multiple arbitrary planar objects efficiently with minimal training requirements. In this thesis, we propose a multiple planar object pose estimation system, DetDPE, based on object detection and a direct-based pose estimation algorithm. We fine-tune an off-the-shelf object detector with only a modest amount of synthetic image data. The object detector identifies the planar objects for pose estimation and significantly reduces the complexity of the estimation algorithm. Our DetDPE system demonstrates efficiency while maintaining the high accuracy of the original direct-based algorithm. Furthermore, we integrate the object detection approach with a feature-based pose estimation algorithm. The results show that this approach can indeed enhance the performance of feature-based algorithms. Therefore, we can regard the object detection approach as a framework applicable to various types of pose estimation algorithms.