在多種生物辨識技術中，步態特徵能夠在中距離進行辨識，適用於街道監控系統，在常見的步態辨識方法中經常使用背景建模方法，此方法遇到複雜的動態背景，會使得前景提取困難，由於人體姿勢特徵不易受到動態背景的影響，因此本研究基於人體姿勢特徵，設計一個步態辨識系統，其中包含人體姿勢偵測、行人追蹤、步行週期偵測、建立步態特徵和步態辨識。人體姿勢偵測使用OpenPose，行人追蹤使用行人重識別方法，步行週期偵測使用人體腳踝偵測方法，步態辨識本研究結合人體姿勢與光流特徵，並提出人體骨架遮罩方法來去除部分的光流背景雜訊，在遮罩部分保留了骨架輪廓特徵，光流特徵則保留運動方向，最後使用CNN步態辨識方法與CASIA Dataset B步態資料庫進行訓練，步態辨識率為85.5，驗證此方法是有效的。;Among various types of biometrics, gait analysis is used for middle-distance gait recognition and is suitable for street surveillance system. A common gait recognition method is background subtraction. In this method, foreground extraction is difficult if the background is complex and dynamic. Because the features of human poses are not easily influenced by dynamic backgrounds, this study designed a gait recognition system based on human pose features. The system included functions of human pose detection, pedestrian tracking, and walking cycle detection and enabled the establishment of gait features and gait recognition. Human pose detection was performed using OpenPose; pedestrian tracking was conducted through person re-identification; and walking cycles were detected using the data of human ankles. For gait recognition, this study combined human poses and optical flow features, proposing a mask method for human skeletons to remove a part of the noises in the optical flow background. Features of skeleton outlines were retained from masks, and movement directions were retained using optical flow features. Finally, this study adopted convolutional neural network gait recognition method and the Dataset B of Chinese Academy of Sciences Gait Database for data training. The obtained gait recognition rate was 85.5%, confirming that the proposed method was effective.
