以關鍵姿態及運動軌跡的權重向量進行人類動作辨識

Abstract

本論文提出了一個室內人類動作辨識系統。一般來說，動作是由一組有順序的關鍵姿態來組成。本論文的方法不同處在於，我們沒有考慮關鍵姿態彼此之間的順序性，只考慮關鍵姿態的「組成」，而關鍵姿態的順序性改以分析目標物的重心移動方向取代。 首先，藉由圖形匹配，將輸入的人類剪影和資料庫內事先儲存的各種動作的關鍵姿態作比對，找出一些最相像的關鍵姿態。本方法的基本想法是，當人類看到一張姿態影像時，腦內會自動聯想該影像中的人可能正在進行的動作。也就是說，每個人類的姿態影像都會對一個或多個動作有相關性。藉由這個概念，我們將每個關鍵姿態都對每個動作賦予權重分數，形成「權重向量」。不同時間點被匹配到的關鍵姿態的權重向量分數相加起，就是關鍵姿態對各個動作的權重分數。由於此權重分數並沒順序姓，我們另外分析了重心的位移情況以及其垂直方向，用來辨識一些由相同關鍵姿態組成但順序不同的動作。 本方法的特色是我們只考量動作由「哪些」關鍵姿態組成(也就是說不考慮關鍵姿態彼此的順序性)，再搭配分析重心的移動軌跡來進行人類動作辨識。此方法的好處是可以對圖形匹配的錯誤結果具有較佳的強韌性，並且能更完整的考量到靜態姿態的描述。此外，在圖形匹配的程序中，我們新嘗試了一種已用於步伐分析[41]但尚未被拿來應用在圖形辨識的特徵，此特徵能保留姿態外觀的輪廓特性。我們以17個室內常見的動作為辨識目標，並且測試了五個受測者以及四種視角，實驗結果證實本方法的辨識率可高達89.23%。此外，本方法具有可擴充性，可自由的增加想判斷的動作。

In this thesis, an indoor human actions recognition system is proposed. Generally, actions are composed by sequences of key postures. In our approach, the order of key postures is not considered, that is, only the compositions of actions are considered. The center point trajectory of the target is analyzed to substitute for the order of key postures. By using pattern matching process, the input human silhouette is matched with the key postures pre-stored in the database, and some key postures are matched in each frame. The idea of our approach is that people associate the possible actions when seeing a key posture. In other words, each key posture is related to one or more actions. Taking this idea, every key posture in database has weights for all actions, called weighting vector. The weighting vectors of matched key postures in recent frames give action scores for every action. Because this method has no property of the order of key postures, the center point trajectory is analyzed to distinguish actions with same key postures but have different orders. The feature of our approach is that the order of key postures is not utilized. The composition of key postures and center point trajectory are used to recognize human actions. This method is robust against the error result of pattern matching, and the stationary temporal situations are also considered. Besides, in pattern matching process, a feature called “distance vector” [41] which was applied in gait recognition is modified and utilized as the pattern feature. This feature keeps the characteristic of human exterior contours well. Seventeen common human actions and static postures are recognized, and 5 subjects are tested in 4 viewpoints. The recognition rate of our approach is 89.23%, and the experiment result shows that our approach has potential to recognize more actions.