Research On Human Posture Recognition And Abnormal Behavior Understanding Methods

With the increasing improvement of information technology and large-scale use of computer vision technology,intelligent recognition has become a research focus.This paper deals with video data processing for human posture recognition and abnormal behavior understanding by modern and intelligent means.Human posture recognition is the accurate identification of the skeleton information.The skeleton information carries human posture from video sequence images,which is the characteristic information for human behavior description.There are some problems in the actual video data,such as: low accuracy in detecting human nodes due to the vulnerability of video images to illumination,observation angle,occlusion and scale changes.The diversity,randomness and unpredictability of behavior categories lead to low accuracy of behavior recognition.The understanding of abnormal behaviors in specific environments is less studied.The proposed algorithm in this paper is applied to practical scenarios using a hierarchical strategy to determine the state of human targets in customized contexts and to understand the abnormal behavior of objects in extreme danger,providing some theoretical and technical support for security monitoring and key research areas in specific regions.Meanwhile,Key research elements are as follows.In order to address the problem of low detection accuracy due to difficult nodal point recognition caused by target instances affected by multi-scale changes and occlusions,an improved top-down framework human pose recognition algorithm(GOS-HRNet)is proposed.Firstly,Octave convolution is used to reconstruct the high-resolution network to enrich the multiscale spatial information of feature representation.Secondly,second-order attention is applied to focus on different weights and channel scale spatial information,which improve the joint point detection accuracy.Furthermore,a gradient equalization mechanism is adopted to improve the balance of the network’s attention to recognizing difficult points and to strengthening the robustness of the model algorithm to image scale sensitivity.The results show that the algorithm performance improves the joint point detection accuracy on the COCO2017 dataset.In terms of the problem of fixed spatial templates for graph construction and lack of temporal contextual information in traditional graph convolution,an improved algorithm for graph convolutional behavior recognition is proposed.In the spatial dimension,adaptive graph convolution is used to aggregate neighbor joint features within different distances.It can increase the flexibility of the global graph in the model and the generality of various graph structures.Its ability to reduce the dependence of interlayer graph features on specific graph structures.In the temporal dimension,a combined scale structure is used to enable the model to capture changes in the joint feature vectors at the temporal level.Experiments have shown that the proposed algorithm is effective in improving the saliency representation of spatial and temporal features of sequences on publicly available datasets,which,in turn enhances the behavioral recognition accuracy.When normal and abnormal behaviors are studied,there is a lack of clearly definition.A method for understanding abnormal behavior based on a hierarchical strategy is proposed to explain how to understand abnormal human behavior in specific scenes and create a dataset of abnormal behavior.To further illustrate the feasibility of the method,the movement of limbs is studied at the overall level and local level based on the correlation between scene and behavior.Firstly,the defining layer is used to define abnormal behavior within the constraints of the scene.Secondly,the local layer identifies the category of abnormal behavior with a behavioral recognition model.Additionally,the holistic layer focused on the motion state information of the target(e.g.,trajectory).Finally,the analysis results of different layers are integrated to achieve abnormal behavior understanding.Experiments show that this method is feasible and exhibits good accuracy.In this paper,the network framework is optimized,the spatial representation information learnt by the model is enriched,and the recognition accuracy of the algorithm is improved to achieve human pose recognition and abnormal behavior understanding in videos.