Research On Semantic Segmentation Comprehension Algorithm Of Human Activity In Complex Activity Scene

Semantic understanding of human behavior based on complex scenes is a challenging topic in recent years.Most of the semantic subdivision understanding in complex activity scenarios is carried out around human behavior activities,including: the understanding of human behavior semantic segmentation,the semantic segmentation of human identity features and the understanding and analysis of human activity trajectories.The research content of this paper is mainly aimed at the above three aspects:(1)Understanding the semantic segmentation of human behavior in complex scenarios.In a complex activity scenario,multiple different human behavior actions exist in the same scene.Since most video tags are labeled for a single behavioral activity,the necessary annotations are missing for scenes with multiple bodies.In view of the above problems,this paper proposes to use the key points of the human body to locate the active area of the human body,and then model the localized area through the convolutional neural network,and finally obtain the classification corresponding to the behavioral activities in the area.The feasibility of the method was verified in the actual video scene.(2)A subdivisional understanding of human identity semantics in complex scenarios.In a complex activity scenario,there is not only an understanding of the behavioral actions in the individual activity areas,but also a semantic understanding of the initiators of the behavioral actions(ie,the identity information in the individual activity areas needs to be understood).The human biometrics used in this paper are face information.Firstly,the position of the face is detected in the individual active area by means of multi-convolution neural network cascading.Considering that the angle of the face to the camera is different in the actual scene,It is necessary to use the affine transformation to detect the face,and then use the convolutional neural network as the feature extractor of the face feature.Finally,the face recognition function is used to perform the face recognition work to complete the human body identity semantic understanding.Task.The feasibility of the method was verified on the public dataset and in the actual video scene.(3)Understanding and analyzing the human activity trajectory in complex scenes.In complex activity scenarios,the behavioral semantics of specific behaviors are captured and understood,and in the absence of identity information in the region,the human activity trajectory needs to be tracked.In this paper,the tracking method based on nucleation-related filter is used,and multi-feature fusion and multi-track scale transformation are added to make the feature description of the area to be tracked more comprehensive.The feasibility of the improved method was verified on the public dataset and in the actual video scene.In summary,this thesis conducts human body semantic segmentation understanding for the video of actual complex activity scenes,uses a variety of feature description operators,combined with machine learning methods,and verifies the feasibility of the above research in practical applications through experiments.