Research On 3D Reconstruction Technology Of Twin Data Hand Based On Joint Information

In the real world,the hand is the primary limb with which a person interacts with the outside world;Driven by the need for human-computer interaction in the virtual world,the hand is also required to have a more vivid form of expression to be better suited for motion capture,gesture recognition,and other broader 3D scenarios.Both augmented reality and virtual reality rely on information technology support and must be realized with the arithmetic power of servers and underlying technology.How to strike a better balance between reducing computational costs and improving computational accuracy,and reducing costs and increasing efficiency has become a critical issue.At the same time,due to the physical and physiological characteristics of the hand,3D realistic reconstruction of the hand is still a challenging research topic.In this paper,we use the 3D spatial position information of the hand nodes of the target and the given original input image of the hand to generate a hand image corresponding to the target node information.The difference image between the hand image generated by the 3D spatial position information and the original hand image is predicted directly by the U-Net network,thus effectively reducing the artificial traces in the generated hand image.This process uses a traditional partial affinity field representation to describe the three-dimensional pose structure of the hand and the two-dimensional position information in pixel space.The NYU dataset is also combined to provide more informative 2D color information for 3D reconstruction of the hand.In order to realize twin data hand generation containing two-dimensional color image and depth image information,this paper use depth images as the input raw data and regress the 3D coordinates of key points of the hand to realize the 3D data hand reconstruction.Since the used dataset contains complex background information,we first determine the hand center of mass from the input image,segment the complete hand,and will keep only the data information of the hand for feature extraction.An attention mechanism is introduced in the feature extraction process to focus on the key points and provide a dense representation of the feature map to achieve end-to-end completion of the 3D reconstruction of the hand.Through comparison tests with existing advanced methods on the NYU dataset,it was demonstrated that the improved method could achieve the expected results,outperforming the existing commonly used methods,and achieving better results.