利用RGB-D感測器與統計學習模型即時估算人體計測值

A Statistical Learning Model to Estimate Anthropometric Measurements Using an RGB-D Sensor

潘映辰

指導教授 ： 王茂駿 ； 黃思皓

國立清華大學/工學院/工業工程與工程管理學系/碩士(2014年)

https://doi.org/10.6843/NTHU.2014.00283

摘要

自動量測人體計測值是許多人因研究及產品客製化中很重要的一個步驟，隨三維人體掃瞄儀器的發展，耗時且不精準的手動量測逐漸被取代。利用三維人體模型，許多半自動的演算法被應用於快速蒐集人體幾何資訊，並進一步計算人體計測值。然而，雷射掃瞄儀器價格昂貴且環境限制較多，大幅侷限其終端的應用系統發展與建置。
本研究提出了一套以RGB-D感測器以及統計學習模型為基礎的自動人體計測值估算系統，其主要包含離線參數模型學習模組(off-line parametric model learning stage)以及即時人體計測值估算模組(on-line measurement estimation stage)。在離線學習模組中，主要採用了一連串的影像處理、電腦視覺及統計型機器學習演算法，例如主成分分析、線性迴歸以及類神經網路模型，進行三維掃描人體資料庫分析。在即時估算模組方面，則是利用由RGB-D感測器擷取深度圖(depth map)，經過步進式三維點群疊合演算法(iterative 3D point cloud registration process)產生模擬之全身模型並推估人體模型變形參數。再利用離線學習模組中建立的參數模型即可獲得估算人體計測值。
實驗結果顯示，本研究利用商業型RGB-D感測器所建構的系統能夠有效的預測九項人體計測值，包含六項長度以及三項圍度。除此之外，此系統亦能接受對於不同型態的深度資訊做為系統輸入，而群集分析技術(cluster analysis)的採用亦可有效降低估測值的誤差。與先前同樣利用RGB-D感測器建構的文獻結果相比，本研究所提出的自動估算系統能更快速且準確的獲得人體計測值。

關鍵字

人體計測值 ； 群集分析 ； 參數化模型 ； 點群疊合 ； RGB-D感測裝置

並列摘要

To automatically estimate the anthropometric measurements of individuals is an important step for ergonomic researches and customized product design. With the development of 3D scanning technology, several automatic or semi-automatic estimation systems based on the scanned 3D models are proposed to substitute for the time-consuming and inaccurate manual estimation. This 3D scanning system can also be utilized to rapidly collect large amount of human geometric information for further analysis. However, the instrumental cost and the required installation space of 3D laser scanner limit the practical applications.
In this thesis, an automatic anthropometric measurement estimation system based on a low-cost RGB-D camera and a pre-learned statistical model is proposed. The proposed system includes an off-line parametric model learning stage and an on-line measurement estimation stage. A series of image processing, computer vision, and statistical machine learning algorithms, such as principal component analysis (PCA), linear regression, and artificial neural networks (ANN), are applied to analyze 3D scanned bodies in the off-line stage. In the on-line testing stage, the depth map captured by the RGB-D camera is used to synthesize the whole-body 3D model with 3D point cloud registration algorithm. The estimated surface deformation parameters can be further applied to pre-learned parametric models for anthropometric measurement estimation.
The experimental results show that the proposed system based on a commercial RGB-D camera can effectively estimate nine anthropometric measurements, including six length measurements and three girth dimensions. The results of simulation experiments and real depth data experiments also demonstrate the robustness of the proposed system with three different types of depth inputs. Finally, the cluster analysis can improve the estimation accuracy with gender classifiers and dynamic parametric models. Compared with the state-of-the-art systems based on RGB-D cameras, the proposed system can estimate anthropometric measurements accurately and efficiently.

並列關鍵字

Anthropometric Measurements ； Cluster Analysis ； Parametric Model ； Point Registration ； RGB-D Sensor

參考文獻

[1] M.-J. J. Wang, E. M.-Y. Wang, and Y.-C. Lin, “Anthropometric data book of the Chinese people in Taiwan,” The Ergonomics Society of Taiwan, 2000.

[2] ISO-7250, International Organization for Standardization, 2008.

[3] D. Protopsaltou, C. Luible, M. Arevalo, and N. Magnenat-Thalmann, “A body and garment creation method for an Internet based virtual fitting room,” Advances in Modelling, Animation and Rendering, pp. 105-122, 2002.

[4] J.-M. Lu, M.-J. J. Wang, C.-W. Chen, and J.-H. Wu, “The development of an intelligent system for customized clothing making,” Expert Systems with Applications, vol. 37, pp. 799-803, 2010.

[5] C. C. L. Wang, Y. Wang, T. K.K. Chang, and M. M. F. Yuen, “Virtual human modeling from photographs for garment industry”, Computer-Aided Design, vol. 35, no. 6, pp.577-589, 2003.