Abnormal Behavior Detection Based On Computer Infrared Vision

With the current social aging process of deepening,the phenomenon of injury and sickness among the elderly is more serious because of the fall.In recent years,"Violent abuse" of the elderly,children,and people living alone is a frequent occurrence.It deepens the social concern about such abnormal behaviors.Automatic detection and timely warning of abnormal behaviors have become a new focus in the field of behavior recognition.Abnormal behavior detection based on computer vision,it can realize the advantages of not directly contacting the monitored and having little influence on the life of the monitored.Indoor home scenes often appear under the dark environment.At the same time,the requirement of 24-hour detection for abnormal behavior is proposed.Therefore,this paper proposes to build an indoor abnormal behavior recognition model based on infrared band to solve this series of problems.This research topic aims at the abnormal behavior under the indoor home scene.It based on computer infrared vision,with the depth model and realize the detection of abnormal behaviors in indoor home scenes under the infrared band.The main content of this paper is as follows:(1)An abnormal behavior detection model based on computer infrared vision is established.It deal with the problem that system can't detect human body in the condition of no light in the visible band and the foreground of contour extraction is covered or lost When the visible band has light condition.This model realize the abnormal behavior of 24-hour detection.(2)The infrared band anomaly behavior dataset of indoor home scene is established.The subject uses long-wave infrared camera to acquire data.The normal behaviors of "walk","carry","crouch ","sweep" and "shake hands" are collected.And abnormal behavior under the home scene of "fall","kick" and "push" are collected.It solves the lack of indoor home scene abnormal behavior dataset and dataset construction for this particular application scenario is implemented.It provides the foundation for the training of depth model.(3)A new expression of overall spatial-temporal information of the action is established,the Global Silhouettes Difference Stack Image(GSDSI).And a deeper convolutional neural network is also used,with cumulative time series processing capability.A new global spatial-temporal information learning framework GSDSICNN structure is obtained.And then it merges with the "two-stream" structure and forms a new "three-stream" structure to realize the recognition and detection of abnormal behaviors in indoor scenes.This paper is based on the requirement of abnormal behavior detection in indoor home scene and constructs the corresponding experimental dataset of infrared band.The proposed GSDSI-CNN model and the fused "three-stream" structure are used for training and testing of abnormal behavior detection.Experimental results show that the proposed GSDSI-CNN model realize the total behavior recognition accuracy of 96.25% and detection accuracy of abnormal behaviors is 100% on the dataset of indoor home abnormal behavior.The fused "three-stream" structure realize the total behavior recognition accuracy of 98.75% and detection accuracy of abnormal behaviors is 100%.The results verify the validity of the proposed model...