以二維形狀特徵為基礎的人體姿勢辨識系統

Abstract

本論文中建立一個以影像形狀為基礎的人體姿勢辨識系統並加以分析與實現。此一系統利用混合高斯機率模型建構背景模型，對目前影像做背景濾除以取得前景，並結合肯尼邊緣偵測法與加速的梯度向量流動態輪廓偵測法得到人體姿勢的輪廓。分析目前主要形狀特徵描述子應用在人體姿勢辨識的適應性，比較傅立葉描述子與曲率比例空間描述子的特性以選定適當的描述子。此一姿勢辨識系統建構在一個以2D形狀特徵為基礎的3D姿勢特徵資料庫，解決同一姿勢因不同視角所導致的特徵差異而建立過多的姿勢類別。並同時建構一個針對人體姿勢行為的機率建模機制以輔助此辨識系統，以預防不同姿勢間不合理的轉換並提高其辨識率。實驗結果分別呈現對特徵的選取、機率建模機制、姿勢類別多寡的辨識率與序列行為的姿態辨識結果並加以分析與討論。

In this thesis, a human posture recognition system based on shape-based descriptors is proposed and implemented. The foreground image of human is acquired by the background model built by Gaussian Mixture Model method. We applied the Canny Edge Detection and Speedy GVF Snake to obtain the contour of the foreground as the input of the human posture recognition system. The suitability of the major contour descriptors, Fourier Descriptors and Curvature Scale Space Descriptors, applied to human posture issues are discussed. The 3D postures databases are constructed by 2D contour-based features to avoid the posture classes increase with the different points of view for the same posture. In order to prevent the unreasonable transitions between postures and improve the recognition rate, we integrate a probability map based on human behaviors to the recognition system. Lastly, the results of Experiments for different features, a probability map, different numbers of posture classes, and the recognition of the behavior sequence are listed and discussed.