以去浮點數實現GMM之研究

Abstract

本論文的主要研究內容，為將GMM(Gaussian Mixture Modeling)演算法中的參數及其運算過程中的變數去浮點數化，此為演算法硬體化之前或應用在嵌入式系統中所必須做。在PC上實作不同資料格式的GMM演算法，並將其參數去浮點數化後，比較其去浮點數化前後的影像處理速度、記憶體使用的大小及前景物件辨識的準確度。提出一個參數的資料格式，可以維持高準確度，並有效降低記憶體的使用量。將GMM演算法中的參數去浮點數化後，除了可以提高日後硬體化實作開發的速度，並可大量降低記憶體的使用量。

The goal of this thesis is investigation of Fixed-Point implementation of GMM (Gaussian Mixture Modeling). It must be done before hardware implementation or porting to embedded system. We implement it for several data types of GMM in PC and Fixed-Point parameters of GMM then compare its speed of image processing, memory size and recognition precision of foreground objects with the result after Fixed-Point. We provide a data type of parameters that can keep high precision and down size memory. After Fixed-Point implementation of GMM, it can save developing time and a lot memory when implementing the hardware of GMM or porting GMM to embedded system.