| The purpose of human pose estimation is to detect the key nodes of human body in the image.It is the basis of behavior recognition and other technologies.At present,people’s exploration of the world is no longer satisfied with land,but to the sea.When human pose estimation technology is applied to divers,it will bring new application scenarios and challenges.Therefore,aiming at the subject of multi diver pose estimation,this paper proposes a multi diver pose estimation method based on deep learning with the top-down technical route.The specific research contents are as follows:Firstly,in order to verify the proposed multi diver pose estimation algorithm,the corresponding multi diver pose estimation dataset is constructed.The samples in the multi diver pose estimation dataset can be divided into three parts: the images in the public dataset of human pose estimation,the diver images captured and collected in the field,and the paired underwater images for underwater image enhancement.Among them,more than 6000 diver images were taken and collected on the spot,including single diver and multi divers.Secondly,aiming at the problem that the low contrast of underwater images will cause the blurring of divers’ features and make it difficult to estimate divers’ pose,based on the conditional Generative Adversarial Networks FUn IE-GAN,this paper integrates the attention model to model the weight relationship between feature map channels,and adds a high-resolution multi-scale feature extraction module,A real-time underwater image enhancement algorithm is proposed to improve the quality of diver image.High-quality diver image can improve the accuracy of pose estimation.The whole network includes discriminator and generator.During training,the generator and discriminator are trained alternately.The effect of the algorithm is verified on EUVP dataset.Finally,in order to improve the real-time performance of the pose estimation algorithm,based on Simple Pose,the idea of Shuffle Netv2 is integrated,and the hybrid domain attention mechanism is introduced to propose a new lightweight pose estimation backbone network.While ensuring the accuracy of pose estimation,the real-time performance is improved.The training method of the whole convolutional neural network is based on knowledge distillation,which makes the lightweight network learn the generalization ability of large-scale network and further improve its pose estimation accuracy.The algorithm is verified on multi diver pose estimation dataset and MSCOCO-2017 val.Finally,experiments show that the proposed algorithm can effectively ensure the real-time performance while ensuring the accuracy of pose estimation. |
