基於動態時間扭曲之人體姿勢辨識

Abstract

辨識技術在電腦視覺的領域中，是非常重要的一項課題。在生活中處處可見，如指紋辨識、眼睛辨識、姿勢辨識等應用。人體姿勢由開始到完成可以視為連續的靜態姿勢組成，利用一般攝影機進行辨識會因為其高複雜度的資訊而有所限制。近年來由於體感控制加入姿勢辨識中，原理為利用更具特徵性的骨架資訊和深度資訊，使用者只需要站在攝影機前面就可以達到控制滑鼠和鍵盤的效果。 因此本研究使用Kinect作為輸入端，讓使用者能夠自行輸入動作後經由系統進行辨識，使用者不需侷限於特定的姿勢，利用Kinect所提供的人體關節點座標 值，經過正規化後，可得到以人體肩膀中央為原點的新座標系統。利用動態時間扭曲(Dynamic Time Warping)演算法和最近鄰居法(1-Nearest Neighbor)對使用者的姿勢進行辨識，辨識時計算目前姿勢和訓練樣本的歐幾里得距離，挑選差距最小的為辨識結果，並設計增量(Incremental)演算法讓使用者不需額外訓練就可以維持理想的辨識率。 本系統的固定姿勢在概括樣本(General Sample)平均辨識率可達86.02%，辨識五種自訂姿勢時也可達到75.60%的辨識率，使用增量法比未使用提升了近3%的辨識率，証明此法可行性。Recognition technology is a very important issue in the area of computer vision. The applications include fingerprint recognition, iris recognition and gestures recognition in our lives. Body gestures are composed of many continuously static poses that are segmented from the action, it’s difficult to work because high complexity information by using common webcam. In recent years, somatosensory system had become popular in gestures recognition, user could control the object in the screen like using mouse or keyboard. This paper establishes a human body gestures recognition system, which can recognized the gestures from user defined. Using Kinect to generated 3D coordinates value of skeleton joints, and using translation method to normalized coordinates, after that, we get a new original point by Shoulder Center point. We use Dynamic Time Warping (DTW) algorithm and 1-Nearest Neighbor to compare and classify body gestures. Calculating the Euclidean Distance between training data and testing data. The minimum distance is the result. Finally, we design an Incremental method to keep recognition rate without any extra training by user. In our system, the average recognition rate of static gestures in general sample is 86.02%, five gestures defined by user is 75.60%, and increasing almost 3% after using incremental method.