Research Of Indoor Abnormal Behavior Detection Based On RGB-D Image

In recent years,with the continuous development of human society,people's living standards are also rising,at the same time,people are increasingly concerned about the living environment and personal health and safety aspects.In order to detection the abnormal behavior in the target environment which is go against to people's personal health and property safety,more and more monitoring equipment is installed in people's living places,such as video surveillance equipment.Traditional monitoring equipment mainly through processing of RGB images to obtain information,with the birth of depth sensor,people can obtain more information through the depth sensor,including depth information and human node information,these information will be benefit of the detection of abnormal behavior.This study is based on Kinect somatosensory equipment to detect indoor smoking and fall behavior.Firstly,we study the method of determining the key frame image of the smoking behavior in the video stream,which based on using the human joints information.At the same time,in order to cut out the hand region,we use the relative position information between different joints based on the joint information.Based on the analysis of the characteristics of smoking behavior,extract the HOG feature from the captured hand image,before extract the HOG feature,we compare the different and select a gradient operators,and save the skeletal image data of the key frame image at the same time,then using some joints data to calculate four angles of upper limbs.The HOG characteristics were combined with the upper limb angle characteristics of the smoking time and the training SVM classifier was applied to the smoking behavior detection.The HOG characteristics were combined with the upper limb angle characteristics at the smoking moment,then using comprehensive characteristics to train SVM classifier,at last,the experiments show that using the comprehensive characteristics combine with SVM classifier can obtain better detection effect.In this paper,the detection of the fall behavior based on human joint information,according to the changes of the aspect ratio of the human body combined with the velocity distribution of the three joints when the behavior occurs,set the aspect ratio threshold and the velocity threshold in the experiment,test the different types of fall behavior and non-fall behavior,the experimental result show that based on the changes of human body aspect ratio and the joint velocity distribution method of human fall detection could achieve a better effect.