Research On Person Abnormal Behavior Detection Based On Enviroment Feature And Pose Feature

In recent years,with the continuous improvement of computer performance and the continuous improvement of computer vision technology,smart video monitoring technology has been widely used.Compared with eye monitoring,smart video monitoring does not produce visual fatigue.It can monitor multiple cameras at the same time and keep working 24 hours uninterruptedly.It has the incomparable advantages of eye monitoring.The purpose of using Intelligent Monitoring is to identify dangerous situations in video.Traditional Intelligent Monitoring technology can only recognize a specific target and its type in a specific scene,and does not have the ability to understand the intent of the target behavior or the nature of the event in video.In order to obtain a wider range of application scenarios for smart video monitoring,anomaly behavior detection technology has become the focus and difficulty of its research.Abnormal behavior detection technology can generally be divided into two tasks.Feature extraction is the focus of research on anomalous behavior recognition technology.The quality of extracted features directly determines the performance of the algorithm.For anomalous behavior recognition technology,the extracted features are generally divided into motion features and environmental features,because human behavior is a continuous process,the extracted features should be time series.Motion features,including light flow and pose,can describe the features of human behavior and are important basis for recognizing human behavior.Environment features are used to describe the characteristics of the environment in which the subject of action is located.Extracting environment features is also an indispensable and important link for understanding human behavior intentions.The focus of the feature classification task is on the selection of classification models.This paper mainly focuses on two main tasks of anomaly behavior detection.First,the feature extraction task,a single action feature can not describe the environment in which the action is located,while the environmental feature relies too much on the background and appearance characteristics,and lacks the modeling of the movement itself,so this paper will extract the fusion feature of the action and environment.This paper first carries out the research on the attitude detection algorithm,realizes the multi-person real-time attitude detection algorithm in bottom-up mode,then generates input data based on the attitude information,implements an anchor-free target detection method using the method of predicting the human target center and area,completes target segmentation based on the result of target detection,and takes the pose as the basis.The state information is used as input,and the motion feature extraction is completed by using a LSTM model with time positioning capability.Uses inception module and Res Net structure to improve the structure of C3 D neural network,so that it can extract the environmental features better.After completing the feature extraction task,this paper will extract the environmental features.Fusion with action features and use case learning model to complete the feature classification task.Finally,this paper compares the proposed algorithm with several advanced algorithms in many aspects,and designs simulation experiments in real environment to prove the performance advantages of this algorithm.