以加速後最低增量編碼長度分類器分析人類步態資訊之行為辨識

Abstract

在本篇論文中，我們提出了一個利用人類步態資訊以及最低增量編碼長度分析人類行為的方法。作為特徵，步態能量圖 (Gait Energy Image) 被轉換為向量，我們利用最低增量編碼長度分析這些步態資訊。除此之外，我們也引入對同一動作的多角度分析，以多數決的方式更進一步增進整體分析的表現。實驗結果顯示這個方法在人類行為辨識應用上的準確率平均約高達百分之九十五，除此之外，藉由圖形運算器 (GPU) 技術的協助，我們可以在很短的時間之內完成前述人類行為的分析與辨識。換言之，本文所提出的方法可以作為一個實用的模組，被整合到偵測不尋常動作的監視器系統之中。

In this thesis, we present a novel approach based on the gait energy image (GEI) and the minimum incremental coding length (MICL) for classifying human actions. GEIs are transformed into vectors as input features, and MICL is employed to be the classifier. Moreover, the strategy of majority voting is applied to the MICL classification results to improve the overall system performance. Experimental results show that the proposed approach can achieve approximately 95% of precision, and the classification task can be accomplished in a very short time with the aid of GPU. In other words, the proposed approach can be integrated as a useful component for detecting abnormal events in video surveillance applications.