利用稀疏编码结合深度学习的人体姿态估计

Human pose estimation using sparse coding and deep learning

针对部分遮挡及模糊场景中的人体动作图像姿态估计精确不高的问题,提出了一种基于稀疏编码和深度学习的人体姿态估计方法。首先,对动作图像进行稀疏编码,用较少的已知基函数重构测试样本。然后,在训练算法中利用空间关系独立地分类像素,进一步通过像素分类将图像分割为部分区域,将前景区域分割加入到人体部位估计中。最后,利用空间关系训练分类器实现人体部位估计的实时处理。在CDC4CV数据集上的验证结果表明,提出的方法能够有效提取出高分类性能,在许多情况中精度超过了85%。

Aiming at the problem that pose estimation of human motion images in partially occluded and fuzzy scenes is not accurate,a human pose estimation method based on sparse coding and deep learning is proposed.Firstly,the motion images are sparse coded,and the test samples are reconstructed with fewer known basis functions.Then,in the training algorithm,spatial relations are used to independently classify pixels,and the image is further segmented into part regions by pixel classification,and the foreground region segmentation is added into the estimation of human body parts.Finally,the spatial relation training classifier is used to realize the real-time processing of human body parts estimation.Verification results on CDC4 CV dataset show that the proposed method can effectively extract high classification performance,and the accuracy exceeds 85%in many cases.