基于关键帧轮廓特征提取的人体动作识别方法

为了在人体动作识别中获得更加准确的前景分割和防止关键信息的几何丢失,提出一种利用关键帧提取关键姿势特征的人体动作识别方法.由于背景建模和差分获得的前景不准确,利用基于纹理的灰度共生矩阵提取动作轮廓,并对原图像帧进行分割;然后计算人体Blob的能量,选取最大信息内容的帧作为关键帧,关键帧的获取使得特征提取对时间的变化具有一定鲁棒性;在特征分类识别阶段,为了提高分类的准确性,提出使用支持向量机-K最近邻(support vector machine-k nearest neighbor,SVM-KNN)混合分类器完成分类.在Weizmann,KTH,Ballet和TUM 4个公开数据集上实验验证了该方法的有效性.相比于局部特征方法、全局特征方法和关键点方法等,该方法获得了更高的识别率.此外,实验结果表明,该方法在KTH和Weizmann数据集上的早期识别效果优于Ballet数据集.

Human activity recognition based on contour feature extraction on key-frame

In order to acquire more accurate of foreground segmentation and prevent the loss of critical geometry information in human action recognition, a human action recognition method based on extracting key-gesture features by key-frame is proposed. Concerning that foreground obtained from background modeling and background differential is not accurate, the action contour is extracted by using texture-based gray level co-occurrence matrix with segmentation on original image frame. Then, body energy Blob is calculated, and frame of maximum information content is selected as key-frame. Key-frame makes feature extraction robust to the change of time. The last is the stage of feature classification. support vector machine-K nearest neighbor (SVM-KNN) hybrid classifier is used so as to improve the classification accuracy. The effectiveness of the proposed method has been verified by experiments on the four public data sets Weizmann, KTH, Ballet and TUM. The recognition accuracy of the proposed method is higher than local feature method, global feature method, key-point method and etc. In addition, the experimental results show that early identification of data sets KTH and Weizmann is better than that of Ballet data set.