Action Recognition Algorithm Based On Human Pose Estimation

Human motion and action recognition is one of the most popular problem of computer vision.It is the main research content in the fields of anti-terrorism warning,intelligent video surveillance and camp security.Aiming at the problems of human body occlusion,camera shooting angle and small scale human body in the image,in this paper,we propose methods to locate the human body position by pedestrian detection firstly,then extract the human body skeleton feature information,and finally use the human motion classifier to analyze the human behavior.Aiming at the problem of small-scale pedestrian and mutual occlusion,the paper proposes a pedestrian detection algorithm based on end-to-end deep learning network.Innovatively proposed a circular convolutional network model,which fully integrates the shallow layers information on the human body and the deep semantic information through multiple upsampling and deconvolution,improves the occlusion of the single-stage pedestrian detector and the robustness of small-scale pedestrians.Leading levels have been achieved in single-stage pedestrian detection algorithms on public datasets.Aiming at the problem of confusing multi-person key point estimation and multi-scale human body parts in the image,this paper proposes a multi-scale cascaded human body pose estimation algorithm.First,a single-stage pedestrian detector is used to locate the human candidate box,and then the cascaded network model is used to estimate the position of the human key point.The method utilizes multiple network cascades,the first step is to estimate the rough position of the human key points,and then the cascaded network model refines the position,thereby improving the accuracy of the algorithm for locating the small key points such as the human ear and nose.In the human action recognition,this paper uses the single-stage pedestrian detector based on deep network and the multi-scale cascaded pose estimation method to extract the human skeleton feature information and train the human action classifier to determine the human behavior.This method can simultaneously analyze multiple human actions and can be used as the basis for multi-person action analysis.