Research And Implementation Of Human Pose Estimation Algorithm Based On Monocular View

In recent years,three-dimensional human pose estimation technology with monocular view as input has been widely used in tasks such as pedestrian recognition,criminal behavior detection,and smart home.However,the existing methods are not ideal for solving the problems of large pose and occlusion and the ambiguity of three-dimensional depth information.Graph convolution uses a graph model composed of key points of the human body as the input of the network to improve the above-mentioned problems in terms of spatial relations and semantic information.However,since the topological structure of the human body is a connected graph,it is prone to oversmoothing when training in deep graph convolutional networks.This paper proposes a low-pass filter graph convolution network based on the Inception module,which improves the above problems through two methods:the network structure and the convolution kernel.In addition,this paper proposes a human body fall detection algorithm based on the center of gravity based on the human body pose estimation algorithm.The acceleration threshold of the center of gravity and the design of the fall detection network are used to filter actions that are easily misjudged by the model as falling,such as squatting and jumping,and effectively improve the accuracy.The main work of this paper is as follows:1.Use the Inception network design idea and the multi-path high-order graph convolution method to fuse nodes at different distances to obtain a multi-scale feature representation.And use the low-pass filter graph convolution kernel function that meets the task of human pose estimation.During the training process,this method not only fuses the output information of the initial convolution layer,but also combines the properties of the adjacency matrix low-pass filter to perform specific Feature fusion.2.This paper proposes a fall detection algorithm based on human pose estimation.The relationship between time and center of gravity offset is used to calculate the acceleration of the center of gravity to determine the tendency of the human body to fall.If there is a tendency to fall,the current image is input into the fall detection network for fall judgment.In order to make the fall detection network model more robust,this paper collects fall images and creates a fall dataset.3.According to the method proposed in this paper,a fall detection system based on human pose estimation algorithm is designed and implemented.The elderly can be monitored in real time,and the guardian will be notified in time if a fall occurs.