中文摘要：

基于可形变部位模型方法,在单目静态图像的姿态估计中获得了较好的结果,它通过描述部位表观和部位间关联来表示人体姿态.但由于在人体部位表观特征匹配过程中受到诸多因素,尤其是受到背景的干扰,检测效果不够理想.为了有效地完成部位匹配,提出基于超像素的姿态估计算法,综合考虑了图像的底层特征和中层特征,根据中层特征-超像素位置完成部位的搜索,使用底层特征-梯度直方图完成部位匹配.首先完成了图像的超像素分割,将相似度相近的相邻像素整合为一个图像块;然后以超像素为单位完成人体部位的搜索,降低背景对部位识别的影响;最后根据可形变部位模型完成人体部位匹配和识别.在数据集IP和LSP上的实验结果表明,基于超像素的姿态估计算法较好地将人体部位从背景中分割出来,提高了姿态估计的准确率.

英文摘要：

The human pose estimation method based on deformable part model obtained better results in monocular static images. How- ever, the detection rate is still not ideally accurate due to many factors especially noisy background during parts detection. The novel al- gorithm for pose estimation based on superpixels is proposed to recognize body parts considering both mid-level feature as location of superpixel and low-level feature as HOG. Firstly, superpixels is used to segment image into several blocks according to the similarity of pixels and distinguish the body part from other parts and background; Secondly, the part detection is guided by the segmentation of su- perpixels, so that lots of noise interferences are avoided. Finally, the deformable part model is used for human pose estimation. The ex- perimental results on dataset IP and LSP show that the approach with superpixels segment human body from the background exactly and increase the recognition rate.