Real-time Detection,Positioning And Motion Recognition Of Human Moving Objects Based On Deep Learning

In today's intelligent computer industry driven by artificial intelligence and big data,image processing and pattern recognition have gradually become the focus of scholars and experts.Target detection and positioning plays a very important role in video monitoring,and it is also widely used in industrial detection.With the deepening of the research,more accurate analysis of the human action in video is required.Therefore,motion recognition has gradually become an important direction in the field of computational vision research.In the field of human-computer interaction and intelligent monitoring,the foreseeable application prospect is very considerable.On the basis of carefully studying and reading a large number of references at home and abroad,this paper summarizes the advantages and disadvantages of the previous research work of scholars,and makes in-depth research on key technologies such as real-time detection,positioning and motion recognition of human moving targets for the specific scene of the stage.In target detection,several methods of target detection based on deep learning are analyzed.In view of the research scenarios in this paper,SSD(Single Shot MultiBox Detector)target detection method is adopted.Through data enhancement,feature optimization and migration learning,the detection model of this paper is optimized to improve the detection accuracy and realize real-time detection of the target.In addition,the uniqueness detection of multiple identities is realized by designing visual calibrators.In binocular visual positioning,the camera calibration adopts checkerboard calibration method,calibration is carried out by using the calibration toolbox of MATLAB camera,and internal and external parameters of camera are solved by defining the corresponding points of world coordinates and image coordinates.Finally,the target is positioned by using the constraints of polar geometry.In the experiment of detection and positioning,the detection speed of 40 fps is realized,and the average positioning error is less than 10 cm.In motion recognition,this paper presents an evaluation method of model walk show based on polynomial fitting.Firstly,the method based on Part Affinity Fields is used to detect human joint points.At the same time,in order to eliminate the difference between camera posture and body shape,the detected joint points can be calibrated by Prussian analysis.Secondly,human joint points are divided into three parts: spine,upper limb and lower limb for analysis,Polynomial fitting is carried out from the horizontal and vertical directions respectively.Then the polynomial fitting coefficient is used to reduce the dimension of PCA data.Finally,the dimensionality reduction coefficient is taken as the feature of the motion analysis and evaluation,and the SVM classifier is used to realize the classification and recognition of the model's walk-show action.The accuracy of this method is 71.9% through cross-validation,and it preliminarily realizes the professional evaluation of model walk-show.