Research On Human Pose Estimation Algorithm Based On Improved High Resolution Network

Technological innovation has driven the development of social life forward,and artificial intelligence has made human production convenient and fast.The emergence of computer vision technology has accelerated the development of artificial intelligence,and as one of its important branches,the human pose estimation technology has been widely used in various fields such as medical,transportation and military.In the intelligent development of military equipment,human pose estimation algorithms provide certain technical support for battlefield situational awareness,and real-time precision strikes.The research of the human pose estimation algorithm is of great significance for military operation command and national defense security,but due to the complex military environment,there are many influences such as background interference,multi-scale interference and key point occlusion,which lead to the human pose estimation algorithm in military environment still has problems such as low detection efficiency.Therefore,in this paper,based on the high-resolution network HRNet,the research of human pose estimation algorithm based on an improved high-resolution network is carried out,aiming to improve the accuracy and rate of human pose estimation algorithm and realize the efficient intelligence of military equipment.The main research content of this paper is as follows.(1)To address the problem of insufficient human key point feature extraction in complex multi-scale scenes,this paper proposes a human key point feature extraction network incorporating a dual-attention mechanism.Firstly,in order to improve the feature acquisition capability of human pose estimation algorithm in multi-scale scenes,the pyramidal slice attention module EPSA is introduced here to improve the acquisition capability of the network for multi-scale information,and the EPSA module is introduced during downsampling to avoid the problem of information loss during the downsampling process of the network.Secondly,in order to enhance the ability of the human pose estimation network to obtain the location information of key points in complex scenes,the network is combined with the location attention mechanism CA to enhance the extraction ability of location features and ensure the adequate extraction of the location information of key points on the human body.The experiments show that the network constructed in this chapter can fully acquire the multi-scale feature information in the image and can effectively extract the location information of human key points from the complex background,which effectively enhances the human joint point feature extraction ability of the pose estimation network.(2)To address the problem of poor detection of occluded human key points,this paper proposes a heat map-based human key point regression optimization algorithm.Firstly,to address the problem that more feature loss during upsampling will lead to the lack of features of occluded key points,the lightweight upsampling module CARAFE is used to improve the original nearest neighbor interpolation upsampling method to reduce the information loss during upsampling.Secondly,to address the problem that insufficient feature information fusion exists during key point heat map regression,which will lead to insufficient prediction information of occluded key points,an improved feature fusion module is designed to achieve sufficient fusion of feature information at different resolutions to provide more reference information for key point location regression.Finally,to address the problem that the key point heat map output has multiple peaks leading to the lack of accurate regression location of obscured key points,the heat map output is optimized by combining Gaussian filtering to make the final output of accurate key point location.The results show that this chapter network effectively improves the detection effect of occlusion key points,improves the detection accuracy of the model,and achieves the precise location of occlusion key points in complex multi-scale scenes with better detection accuracy and stronger model generalization ability.(3)To address the problem of complex network structure,large number of model parameters and large size,which affect the efficient detection of human pose estimation network,a lightweight human pose estimation algorithm is proposed in this paper.Firstly,a lightweight human posture estimation network based on the Ghost convolution module is constructed,and the number of parameters and operations of the network are significantly reduced by combining the Ghost convolution module with the conventional convolution.Secondly,The model pruning algorithm mainly prunes the redundant links between the neurons of the convolutional network,and prunes the redundant links and unimportant neurons according to their weights to reduce unnecessary model operations,compress the model size and improve the detection efficiency on the basis of ensuring the detection accuracy.The effectiveness of the human pose estimation algorithm based on the improved high-resolution network proposed in this paper is finally verified through experiments.