Hand pose estimation is one of the most attractive research areas for image processing. Among the human body parts, hands are particularly important for human–machine interactions. The advent of commercial depth cameras along with the rapid growth of deep learning has made great progress in all image processing fields, especially in hand pose estimation. In this study, using depth data, we introduce two hybrid deep neural networks to estimate 3D hand poses with fewer computations and higher accuracy compared with their counterparts. Due to the fact that the dimensions of data are reduced while passing through successive layers of networks, which causes data to be lost, we use the concept of residual network to compensate this phenomenon. By incorporating data from several views, the estimated poses are more robust in the occlusions. Evaluation results show the superiority of the proposed networks in terms of accuracy and implementation complexity.

手部姿态估计是图像处理中最具吸引力的研究领域之一。在人体部位中，手对于人机交互尤为重要。商用深度相机的出现以及深度学习的快速发展在所有图像处理领域都取得了长足的进步，尤其是在手部姿态估计方面。在这项研究中，我们使用深度数据，引入了两个混合深度神经网络来估计 3D 手部姿势，与它们的同行相比，计算量更少，准确度更高。由于数据在通过连续的网络层时会降低维度，导致数据丢失，我们使用残差网络的概念来补偿这种现象。通过结合来自多个视图的数据，估计的姿态在遮挡中更加稳健。