Research On Dynamic Multi-person Localization Method Based On Monocular Vision

In recent years,the trend of population aging in China is increasing,and the problem of empty nest elderly care is becoming more and more prominent.The service level of the traditional way of employing family caregivers is uneven,and the human cost is also rising.With the rapid development of artificial intelligence technology,such as robot,machine learning,computer vision and so on,the related methods and technologies of home care robot have entered a new development stage.In view of the premise of the normal work of the home care robot is to obtain the location of the care object,so as to further implement the following work such as state evaluation,motion planning,navigation and obstacle avoidance,so it is necessary to study the multi person dynamic real-time localization method suitable for the complex scene of unstructured family.In this paper,the method of dynamic multi-person localization based on monocular vision is studied.Through the multiple regression human body localization model,the nonlinear mapping relationship between human body from 2D image space to 3D localization space is established,and the multi-target tracking method is combined to realize the function of vision multi-person localization.The main contributions of this paper are as follows:Aiming at the practical application scenarios of multi-person localization,and in order to improve the effect of human body localization model,the construction method of human body localization data sample set and the construction method of human body localization features are proposed.According to the time stamp,the real value label of 3D location of human body and the skeleton feature of 2D image of human body are synchronized and matched,the uniformly distributed sample set of body localization is obtained.In addition,through the feature construction of human 2D image skeleton features,the human body localization features closely related to the modeling target are obtained,thus enhancing the interpretability of the human body localization model.Aiming at the problem that the camera has calibration error and monocular vision camera can't get the depth information of human body directly,a vision human body localization model based on multiple regression is proposed.The nonlinear mapping from 2D space to 3D space of human body is integrated into the multiple regression prediction model of human body localization to realize human body localization.This method is based on monocular vision camera,with small calculation and no need to calibrate the internal and external parameters of the camera.Aiming at the problems such as object occlusion,object interaction and object disappearance in the unstructured complex scene of family care,especially in the multi-person scene,a motion model based on 3D location data of human body and an appearance model based on color feature and HOG feature are proposed.By fusing the human body location data into the state variables of the motion model,and using the color feature and the HOG feature to describe the appearance state of the human body,it can better reflect the real motion state of the accompanying object,and reduce the wrong matching phenomenon caused by the occlusion of the object,so as to enhance the robustness of the object matching.The research results of this paper are verified by experiments and applications in the developed visual multi person localization platform system,which shows the feasibility and effectiveness of the above research.