基于频率筛分的无监督人体姿态特征提取与识别研究

Unsupervised Frequency Sifting for Human Gesture Detection and Recognition

基于视频序列的人体行为分析需要检测和判别人体姿态,已有人体姿态检测与判别方法往往达不到实用性要求。从两个方面探讨应用BEMD(bidimensional empirical mode decomposition)算法提升特征分离度与判别性,以进行人体姿态检测和判别:BEMD分解源图像得到的多层固有模态图BIMF具有判别特征,可形成具有强边缘的对比度高的区域,其中包括人体轮廓区域;从低分辨率尺度BIMF图像到高分辨率尺度BIMF图像递归计算,建立基于BEMD的多尺度树(BEMD multiscale-trees tructured)模型,快速提取目标区域并获取人体形状轮廓特征。实验证明,利用该方法进行人体姿态轮廓特征提取,并建立人体姿态的简化模型,可快速检测并判别人体姿态,以达到实时识别。

Human behavior analysis based on video sequences needs to detect and determine human posture. The BEMD （bidimensional empirical mode decomposition） algorithm is used to enhance the feature separation and discriminance. The multi-layer intrinsic modal graph BIMF obtained by decomposing the original image through BEMD has discriminant features, which can form a region of high contrast with a strong edge, including a contour region of the human body. Through recursive computation from low resolution scale BIMF image to high resolution scale BIMF image, BEMD multiscale-trees tructured model is established, the target region is extracted quickly and the contour features of human body shape are obtained. Experiments show that this meth- od can be used to extract the contour feature of human posture and establish a simplified model of human pos- ture, which can quickly detect and determine human posture to achieve real-time recognition.