臺灣博碩士論文加值系統

在此論文中，我們提出了一個僅使用簡單的模型來合成人物動作的系統。本系統首先學習某個人一連串的動作，由這段動作中取得一些必要的資訊，然後再把這些資訊套用到欲合成動作的人身上，藉此合成出欲合成的動作。本系統藉由強健式點配對(Robust point matching)演算法來建立樣本姿勢間決定外型重要的控制點(如關節點等)的對映關係。另外，藉由簡易的外型比對及強健式點配對演算法來偵測欲合成動作人物之控制點。當控制點決定後，我們應用輻射基底函數網路（Radial basis function networks）來合成出此未知人物之新動作。實驗結果證明本論文所提出方法的可行性。

In this thesis, a system for analyzing and synthesizing human postures has been proposed. The proposed system first learns a series of sample posture images of a people and extracts some necessary information from this series of images. Then, based on the learned information, the system tries to synthesize new postures of another people similar to the learned postures by using one single posture image of the people.
In order to establish the control-point correspondences between the sample posture and the input posture, the proposed system first applies the robust point matching algorithm (RPM) to establish the control-point correspondences between the consecutive sample postures beforehand because the consecutive sample postures are similar and the RPM often works well on this kind of point patterns. Then, the sample posture most similar to the input posture is retrieval by a two-stage procedure. The first stage uses three proposed features to find some sample postures similar to the input posture in a flash. Then, the second stage uses the RPM to perform detailed matching to find the most similar one, and also obtains the control-point correspondences between the input posture and the most similar sample posture. Hence, the control-point correspondences from the input posture to any one of the sample postures can be established via the correspondences among the sample postures. Once the control-point correspondences between the input posture and the sample posture to be synthesized have been established, the radial basis function networks are then applied to synthesize the new posture of the people similar to the sample posture. The proposed system was tested against a subject and experimental results show the feasibility of the proposed approach.

目錄

第一章 簡介 1
1.1 研究動機與目的 1
1.2 相關論文研究 1
1.3 研究方法,流程 4
1.4 論文架構 10
第二章 肢體動作學習 11
2.1 學習肢體連續動作 11
2.2 肢體特徵擷取 12
2.3 建立動作與動作間骨幹點對映關係 16
2.3.1 RPM簡介 16
2.3.2 骨幹點配對結果 19
2.4 控制點選取 20
第三章 肢體動作分析與合成 22
3.1 檢索最相似肢體姿勢 22
3.2 控制點對映 24
3.3 應用輻射基底函數網路合成肢體動作 24
3.3.1 輻射基底函數網路介紹 25
3.3.2 以輻射基底函數網路合成姿勢 26
第四章 實驗結果 28
第五章 結論及未來方向 55
參考文獻 56

[1] Linda G. Shapiro and George C. Stockman, Computer Vision, Prentice Hall, Upper Saddle River, New Jersey, 2001.
[2] J. K. Aggarwal and Q. Cai, “Human motion analysis: a review”, Computer Vision and Image Understanding, vol.73 no.3, pp.428-440, March 1999.
[3] Steven Gold, Anand Rangarajan, Chien-Ping Lu, Suguna Pappu, Eric Mjolsness, “New algorithms for 2D and 3D point matching: pose estimation and correspondence”, Pattern Recognition, vol.31, no.8, pp.1019-1031, 1998.
[4] Suguna Pappu, Steven Gold and Anand Rangarajan, “A framework for non-rigid matching and correspondence”, Advances in Neural Information Processing Systems 8, pp.795-801, 1996.
[5] Rangarajan, A., Mjolsness, E., Pappu, S., Davachi, L., Goldmanrakic, P.S., Duncan, J.S.. “A robust point matching algorithm for autoradiograph alignment”, Visualization in Biomedical Computing, 1131, pp.277-286, 1996.
[6] Steven Gold, Anand Rangarajan, “A Graduated Assignment Algorithm for Graph Matching”, IEEE Transactions on pattern analysis and machine intelligence, vol.18 no.4, pp.377-388, April 1996.