一种基于姿态估计的视点无关的人体姿态图像检索方法

View-invariant human pose retrieval in images based on pose estimation

现有的人体姿态图像检索方法仅能检索出相近视点下的人体姿态图像。为了解决这一问题,提出了1种视点无关的人体姿态图像检索方法,首先计算出图像中人体各关节的二维坐标;然后依据各关节的二维坐标和姿态库,估计出各关节的三维坐标信息;进而将三维人体骨骼的夹角八元组作为特征表示,度量图像的相似性。由于该方法在度量图像相似性时依据的是三维人体骨骼,因此不受拍摄角度的影响。实验结果表明,该方法能够提高人体姿态图像检索在多视点运动图像库中的命中率。

Most existing human pose image retrieval methods can only retrieve the human pose image at similar point of view.In order to solve this problem,a view-invariant human pose retrieval method is proposed.First,the 2 Dcoordinates of every human joint in the image are calculated.Then,the 3 Dcoordinate information of every joint can be estimated,according to the 2 Dcoordinates and pose library.After that,the bone angles are calculated as the feature to measure the similarity.Since the measurement of image similarity is based on 3 Dhuman bones,our method can be view-invariant.The experiments show that the proposed method can improve the hit rate of human pose retrieval in multi-view action database.