利用曲率滤波-经验模式分解预处理检测并提取人体目标特征, 以降低图像分解运算复杂度, 同步增强边缘和纹理特征, 提高特征区分性。表现在: (1) 在第一层经验模态分解中, 以曲率滤波曲面映射原图像的连续平滑曲面, 形成包络曲面及均值面, 首层分解图像纹理特征显著, 以下各层凸显边缘与结构特征; (2) 从低分辨率到高分辨率尺度图像匹配出相邻层强边缘区域, 易于人体目标轮廓候选区域的提取; (3) 以首层分解图纹理特征筛分背景, 在相邻层中匹配前景特征区域, 形成人体姿态特征的轨迹图, 易于判别人体姿态及行为。在人体行为识别实验中, 采用曲率滤波-经验模式分解预处理提取的轮廓特征与人体行为典型数据库ground truth对比, 在UIUC示例数据的轮廓提取的精度和召回率都达到90%以上。对人体姿态及行为做识别处理, 验证了预处理方法的有效性。

Curvature filtering-empirical mode decomposition preprocessing was used to detect and extract human target features. It reduced the computational complexity of image decomposition, enhanced edge and texture features simultaneously, and improved feature differentiation. Its performance was in the following areas: (1) In the first layer of empirical mode decomposition, the continuous smooth surface of the original image was mapped by curvature filtering plane to form the envelope surface and the mean surface. The first layer extracted texture features, and the following layers highlighted structure features. (2) Matching regions had sharp edges from adjacent layers which varied from low-resolution to high-resolution, easy to extract the target contour candidate regions. (3) Decomposing the texture features in the first floor to screen the background, matching the foreground feature areas in the adjacent layers to form the trajectory map of the human body posture, it made easy judging the human body posture and behavior. When applying it to human behavior recognition experiments, and comparing it with the ground truth of the human behavior database, the accuracy of the contour extraction and the recall rate of the UIUC sample data were all over 90%. And experiments verify the preprocessing method is effective in human pose recognition and behavior recognition.