利用多觀察值型隱馬可夫模型進行人體動作辨識

Abstract

本論文的主要目的，在於對人體的動作進行辨識，以期望電腦能以較為有意義的方式，來描述動作。

本論文試著利用「隱馬可夫模型」（Hidden Markov Model）的一種變形－「多觀察值型隱馬可夫模型」（Hidden Markov Model with Multiple Feature Streams）來對於動作進行辨識。利用多觀察值型隱馬可夫模型，我們對於動作的分析，可以同時考慮到不同、獨立的特徵，而不像傳統的隱馬可夫模型僅能考慮單一特徵。此外，對於使用理想模型時無法辨識的資料，本論文也提出了三種利用「容錯可能性」的方法來解決。

在實驗的部份，本論文利用單一視角的二維影像序列作為主要的實驗資料；這一步的辨識依據主要是利用人體形狀的水平、垂直投影來進行。而除了二維的實際資料實驗外，本論文也試著模擬由人體動作的三維資訊來切割基礎動作，並以此作為辨識的依據。

The main purpose of this thesis is to recognize human motion and anticipate computers to response in a more meaningful way

In this thesis, we use a variation of “Hidden Markov Model”, which is called “Hidden Markov Model with Multiple Feature Streams” to recognize human mootions. Instead of the traditional “Hidden Markov Model” which can only consider single feature, we use “Hidden Markov Model with Multiple Feature Streams” to help us consider distinct and independent features in human motion recognition. Furthermore, for the unrecognizable data that ideal module can’t solve, we also provide three ways to solve by using “Fault-Tolerant Likeliness”.

In the experimental part of this thesis, we use 2D image sequences with single viewport as the main experimental data. We use horizontal and vertical project histograms of human figure as the main basis for reorganization. In addition to actual 2D image, we also use 3D data by simulating the primitive motions of human motions. We use these primitive motion sequences as a basis for reorganization.