Multi-human Pose Estimation Based On Deep Learning

In the complex environment such as disaster and battlefield,it is necessary to automatically identify the movement and analyze the state of the injured.Therefore,it is of great significance to detect the joints and limbs of the human body in the complex image,i.e.human posture estimation.At present,the technology of multi-human body posture estimation based on deep learning in outdoor complex environment has developed rapidly,but there are still many problems in the application of search and rescue,such as the poor rescue environment,the complex background can not get clear and high-quality image of people;people's posture is limited by the environment,the body of the same limb presents different appearance;many people are interlaced,the limbs are blocked Overlapping results in the decrease of body integrity;the rescue environment belongs to a strange environment,and there is no prior knowledge about human body posture.Each limb of a single or multiple human body may exist in any area of the image,which results in the need of large-area search for image content during model detection.The main purpose of this paper is to build a multi person human pose estimation model based on deep learning.Based on the research status of human pose estimation at home and abroad,the related theoretical basis is studied.The human pose feature extraction network of complex environment image is built to extract its features;secondly,the joint points are estimated and clustered by joint point detection network;secondly,the joint points are matched by component affinity field(PAFS)to form limb features,and the matched limbs and joint points are stored in different sets;finally,the Hungarian algorithm and human body are used The body frame is constructed into a complete human posture model.Based on the commonly used data sets of human pose estimation,LSP and MS coco data sets are selected to train and verify the model,and PCP and PCK standards are used to evaluate the accuracy of human pose estimation.The experimental results show that the multi-human pose estimation model based on deep learning has certain detection ability for multi-human limbs and joints in outdoor complex environment.At the same time,in order to detect the stability of the model,in three scenes with low visibility,such as war,earthquake and smoke,100 images of more than one person are selected for experiments,and the average accuracy of human body estimation is 0.83.Limb estimation with obvious features,such as trunk,has good results(up to 0.84).Limb ends without obvious features,such as forearm and hand,will be tested according to the environment The accuracy of the change estimation is decreased.