Fast and reliable hand pose estimation has a wide application in the fields such as human-computer interaction. In order to deal with the influences to the hand pose estimation caused by the light intensity changes, self-occlusions and large pose variations, a deep network framework based on label distribution learning was proposed. In the network, the point cloud of the hand was used as the input data, which was normalized through the farthest point sampling and Oriented Bounding Box (OBB). Then, the PointNet++ was utilized to extract features from the hand point cloud data. To deal with the highly non-linear relationship between the point cloud and the hand joint points, the positions of the hand joint points were predicted by the label distribution learning network. Compared with the traditional depth map based approaches, the proposed method was able to effectively extract discriminative hand geometric features with low computation cost and high accuracy. A set of tests were conducted on the public MSRA dataset to verify the effectiveness of the proposed hand pose estimation network. Experimental results showed that the average error of the hand joints estimated by this network was 8.43 mm, the average processing time of a frame was 12.8 ms, and the error of pose estimation was reduced by 11.82% and 0.83% respectively compared with the 3D CNN and Hand PointNet.

Fast and reliable head pose estimation algorithm is the basis of many high-level face analysis tasks. In order to solve the problem of existing algorithms such as illumination changes,occlusions and large pose variations,a new deep learning framework named HPENet was proposed. Firstly,with the point cloud data used as input,the feature points were extracted from the point cloud structure by using the farthest point sampling algorithm. With feature points as centers,points within spheres with several radiuses were grouped for the further feature description. Then,the multi-layer perceptron and the maximum pooling layer were used to implement the feature extraction of the point cloud,and the predicted head pose was output by the extracted features through the fully connected layer. To verify the effectiveness of HPENet,experiments were carried out on the Biwi Kinect Head Pose dataset. Experimental results show that the errors on angles of pitch,roll and yaw produced by HPENet are 2. 3,1. 5 and 2. 4 degree respectively,and the average time cost of HPENet is 8 ms per frame. Compared with other excellent algorithms,the proposed method has a better performance in terms of both accuracy and computational complexity.

Shape registration is a common task in non-rigid 3D shape analysis. In order to solve the problems of high complexity, large computation cost and low accuracy of the traditional algorithms, a new hierarchical shape registration method was proposed. Firstly, the heat kernel signature function was defined as the scalar field for a model, and persistence-based clustering was used to extract feature points and salient regions of the model. Then, a novel tree-based shape representation was proposed, whose root node, internal nodes and leaf nodes were defined as the model, the salient regions and the corresponding feature points, respectively. Finally, a new hierarchical shape registration method was designed to make full use of the tree-based shape representation. The hierarchical shape registration algorithm was tested on the SHREC 2010 correspondence dataset and compared with the Generalized Multi-Dimensional Scaling (GMDS) and game theory algorithms. Experimental results show that the proposed hierarchical shape registration method achieves higher accuracy than GMDS and game theory under various shape transformations, including isometric transformation, holes, micro holes, scaling, local scaling, resampling, noise, shot noise and topological transformation; in addition, the computational complexity is reduced significantly.