使用分層隱藏式馬可夫模型作人類動作辨識

Abstract

我們提出根據單一攝影機所擷取影像序列辨識人類上半身動作之方法。首先，將表示人類動作的時間序列影像轉換成包含人體配置的特徵向量序列。人體被定義為由身體各部位所組成的模型，各個部位以符合運動學的方式相連接。因此，從最根部的頭開始，身體的部位可根據此人體架構依序找出。找出身體各部位即可推測身體關節所在位置，關節位置之相對關係可用來估算身體的姿勢。接著我們提出分層式的隱藏式馬可夫模型將人類動作辨識的問題拆解成兩個兩個層面。第一個層面根據低階的特徵分別辨識兩隻手臂的動作。第二個層面根據兩手的互動或相對關係辨識出人類正在進行的動作。我們的方法相較於已發展的研究具有下列幾點優勢：第一點，由於大問題被拆解成小問題，訓練及辨識的程序皆在低維的觀察空間上進行，可避免模型包含過多參數。第二點，系統將每個人體動作視為手臂動作樣式的組合，以利於詮釋及擴充。第三點，由於標準的隱藏式馬可夫模型常遇到當訓練樣本過少時，會導致模型過度符合訓練樣本的現象。我們採用分層的架構及以規則估算人體動作的方法，即可成功的解決此問題。實驗結果展現我們的系統可有效的辨識六種動作，與其他隱藏式馬可夫模型比較，顯示我們系統的穩定性。

We address the problem of human action understanding of the upper human from video sequences captured by single camera. Time-sequential images expressing human actions are transformed to sequences of feature vectors containing the configuration of the human body. A human is modeled as a collection of body parts, linked in a kinematic structure. Beginning with the root part, the head, body parts are searched along the structure hierarchically. The relation of the joints, inferred from the configuration of parts, is used to estimate the human pose. A proposed layered HMM framework decomposes the human action recognition problem into two layers. The first layer models the actions of two arms individually from low-level features. The second layer models the interrelationship of two arms as an action. Our approach has some advantages over previous work. First, by decomposing the problem hierarchically, training and recognition are performed on low-dimensional observation spaces, avoiding an excess of model parameters. Second, our framework is easy to interpret and extend since each human action can be regarded as a combination pattern of arm actions. Third, the layered framework and the rule-based pose estimation method solve the problem of over-fitting with limited training data a standard HMM often faces. Experiments with a set of six types of human actions demonstrate the effectiveness of our proposed method, and the comparisons with other HMM systems show the robustness.