Research Of Home Health Homomorphic Intelligent Monitoring Based On DCGAN Spatiotemporal Information Migration Compensation Mechanism

With the improvement of medical technology and living standard,global population aging phenomenon is serious,the health monitoring of the elderly has become a hot issue,especially the monitoring of falls at home has attracted much attention.The elderly are not able to move easily,their reaction ability is decreased,and they are often unable to respond to emergencies in a timely manner.Therefore,the development of reliable and effective home health monitoring system has important social value.At present,with the vigorous development of intelligent monitoring,human behavior change monitoring system based on computer vision is gradually widely used.Compared with the limitations of wearable devices and environmental devices,computer vision algorithm has the characteristics of low cost,simple maintenance and little interference.Reliable monitoring system can not only guarantee the safety of the elderly in the home environment,but also suitable for hospitals,nursing homes and other public areas,which can save a lot of human resources,material resources and financial resources,and has broad prospects for development.Although video surveillance can effectively monitor the elderly living alone,it also brings risks of personal privacy and environmental information exposure to the elderly.In this paper,the monitoring system for human behavior changes of the elderly can not only achieve multiple classification recognition of behavior changes,but also effectively avoid the exposure of privacy.Firstly,the chaotic pseudo-random Gaussian observation matrix is used to obtain the multi-layer compressed sensing coding model to process the data set,so as to achieve the visual shielding effect and backward moving target extraction,and it is proved that the video foreground frames before and after compression belongs to homomorphic information.Then,the spatio-temporal features are strengthened,the spatio-temporal information transfer compensation is carried out for the compressed video,and the appropriate features are used to represent it.Finally,a classifier is selected for multiple classification recognition.The research work of this paper consists of the following steps:(1)Based on the basic coding theory of compressed sensing,the mathematical model of multi-layer compressed sensing is obtained by combining the idea of multi-layer block;The WELL algorithm was used to generate pseudo-random numbers to improve and modify the non-negative Gaussian random observation matrix to generate chaotic pseudo-random Gaussian observation matrix.The two are combined to obtain visual concealment coding model to achieve visual shielding effect.Then,we extract the homomorphic moving target to obtain the foreground frames of the compressed state.Meanwhile,it is proved that the compressed foreground frames and the original foreground frames belong to homomorphic information.(2)In order to compensate the features of compressed foreground frames,DCGAN heuristic space-time information transfer compensation network is introduced.In this model,the compensator input the spatio-temporal feature set of compressed foreground frames,and the transporter input the spatio-temporal feature set of original foreground frames.The combination training of network and multi-classification tags is improved,and the judgment of truth and falsity is changed to the comparison of corresponding labels of four categories.The compensated foreground frames are obtained through multiple optimization and iterative training,thus strengthening the spatio-temporal features of category features and videos.Then,BWLBP-TOP feature is introduced which can strengthen the feature of moving region and weaken the environmental information and noise.(3)The traditional K-SVD algorithm combined classification error and label consistent regularization into a multi-classification recognition classifier.By constructing the confusion matrix,it can be found that in this paper,the multi-classification of human behavior change,which mainly focuses on fall detection,has a higher recognition rate,and the fall behavior can be identified more efficiently.When compared with the existing classifiers,it not only has a higher recognition rate in the recognition of multiple categories,but also avoids the problem of overfitting.Experimental results on two commonly used fall open data sets show that the accuracy of the proposed algorithm can be maintained at an ideal level,and it can effectively identify human behavior changes under visual shielding.