Application Of Machine Learning In Intelligent Home Monitoring System

For a long time,family security protection has been a hot issue of public concern.Video surveillance is an important means to ensure family security and the demand continues to grow.The traditional family monitoring system is mostly completed by the staff of the central monitoring station to analyze and respond to every abnormal event.Obviously,this traditional mode costs a lot of manpower.With the development of machine learning algorithm,intelligent monitoring system has become a new trend.In this paper,image preprocessing based on video surveillance,home intrusion detection and fall detection are studied in depth.A variety of commonly used image preprocessing methods are studied in this paper.Finally,Laplacian filter and discrete cosine transform are used in the experiment.The texture features of the image are highlighted by Laplacian filtering.Then the low-frequency coefficients of the image are extracted by discrete cosine transform,and then the image is restored.Finally,the effect of highlighting texture,weakening noise and reducing image dimension is achieved.For home intrusion detection,a face recognition algorithm based on weighted fusion of HOG-LBP features is proposed after image preprocessing.Firstly,the image of the person entering the room is captured by the front door camera,and the image is preprocessed.Then,using the complementary advantages of HOG and LBP features.face features are extracted by HOG operator and LBP operator respectively,and the final features of face images are obtained by weighted fusion.Finally,SVM classifier is used for recognition.Experimental results on multiple face datasets show that the proposed algorithm has better recognition performance and robustness than most traditional algorithms in indoor scenes.For the detection of falls,a foreground projection coincidence rate algorithm based on homoallergic transformation is proposed in this paper.When a person walks normally and falls down,the relationship between the motion prospects formed in the two videos is different.This paper describes this feature by means of allergic transformation.Projecting the human motion foreground in one video through the ground homography matrix into another video will coincide with the motion foreground of another video.In the experiment,it is found that the magnitude of coincidence rate represents the degree of human being’s proximity to the ground,while the change rate of coincidence rate represents the intensity of movement.Based on these two characteristics,this paper preliminarily judges the fall event.In order to further eliminate squatting and other interference,the ellipse fitting method is used to extract the external ellipse of human body to analyze the attitude information,and ultimately to determine whether the event is a fall.Tests on several falls data sets show that the method can detect falls effectively and has good real-time performance.