Research On Hand Posture Estimation Based On Deep Learnin

Pose estimation is an important research content in the field of computer vision,which can be divided into human pose estimation and hand pose estimation,which are highly correlated,but due to the higher degree of freedom of human hands,hand pose estimation is often more challenging.Since the rise of deep learning,computer vision has developed rapidly,but hand posture is estimated to still have a lot of room for improvement.This paper conducts research on hand posture estimation and finds that the current hand posture estimation mainly faces two problems:(1)Hand area detection problem: the hand area occupies a small area of the entire image,which is difficult to detect;(2)Joint point detection problem in the hand area: there are many joint points in the hand,and the existing model is computationally intensive.In order to solve the first problem,this paper attempts to use the YOLOv5 object detection algorithm to predict that the hand bounding box will detect the part containing the human hand from the complex image,and the small hand size may cause problems such as missed detection and false detection,and propose to integrate the attention mechanism in the Neck part to further improve the detection accuracy of the YOLOv5 hand.For the improved YOLOv5 algorithm,experimental verification is carried out on the hand-data of the self-built hand dataset based on the public dataset,and the verification results show that compared with the original YOLOv5 algorithm,the YOLOv5 hand detection algorithm based on attention mechanism proposed in this paper increases the m AP of the hand detection task by 0.8%,and has a better detection effect on the hand on the dataset used in the hand pose estimation model.In view of the second problem,on the basis of the above hand detection,this paper aims at the problem that the position of hand joint points is more concentrated and hand feature extraction is difficult,and the high-resolution network is used as the hand feature extraction network to obtain high-resolution feature maps,and a hand pose estimation model based on high-resolution network(HRNet)is constructed to extract hand key point features as accurately as possible.Aiming at the problem of large parameter quantity in high-resolution network(HRNet),this paper proposes a highresolution hand pose estimation network G-HRNet based on the lightweight module GBlock.Firstly,this paper proposes the lightweight module G-Block,which is built based on Ghost convolution,and maintains the performance of the lightweight module through DFC attention,and uses G-Block to build a lightweight bottleneck network to replace the bottleneck network in the original HRNet,which reduces the computational complexity and parameter amount of the network model while ensuring the performance of the network model,and processes the network structure of HRNet.The results show that compared with the hand pose estimation model based on highresolution network(HRNet),the lightweight high-resolution network model G-HRNet proposed in this paper reduces the number of parameters by 66.6%,and the accuracy index PCK@0.2,AUC index and EPE index can basically maintain the original accuracy on the three datasets.The accuracy of the COCO-Whole Body dataset is slightly improved.