Hand Pose Estimation Based On Deep Learning

Hand posture estimation is an important requirement for various intelligent applications for activity recognition.Research on it has been carried out in the field of computer vision for decades.Due to the development of deep learning and the emergence of low-cost depth cameras,the research on hand pose estimation has attracted more and more researchers' interest.With the development of deep neural networks and the establishment of large gesture datasets,three-dimensional hand pose estimation method based on neural networks have become increasingly prominent.Some works use the depth image as input,but hand posture information cannot be fully utilized,which affects the estimation accuracy.Converting the depth image into three-dimensional voxels will increase the amount of unnecessary calculations and may loss some details of the hand.Here,we directly uses the point cloud data as input to estimate the hand pose,which can make full use of the depth image information,and achieves a good performance on the open data set.The main work of this article is as follows:First,the method of acquiring depth information and the imaging principle of the depth camera are studied,and a method of converting a depth image into a point cloud is deduced,which is applied to the point cloud conversion processing of the depth image data set.The point cloud transformed from depth image was visualized.Second,the characteristics and principles of common filters for point cloud data are analyzed and studied.After the depth image is transformed into a point cloud,the method flow of retaining image information with fewer points and more is given.Finally,the hand in each depth image of the ICVL dataset is completely represented by 1024 points.Third,the related knowledge of deep learning is researched.Based on the Point Net model,a deep learning network using point clouds as input for hand pose estimation is designed,and the feasibility and accuracy of the method in this paper are verified on challenging public data sets.