| With the rapid development of high and new technologies such as big data and artificial intelligence,human’s social life is becoming more and more intelligent.Technologies such as intelligent robots,autonomous driving and machine translation have brought great convenience to human in the fields of medical treatment,transportation and education.However,the realization of these intelligent technologies cannot be separated from massive user data,such as the user’s face,behavior habits and location information.Due to the security holes in the system itself or malicious attacks by criminals,these data may be leaked in all links from collection to use,which increases the risk of personal privacy disclosure.At the same time,computer vision technology has made great achievements in recent years,allowing algorithms to extract other sensitive information in multimedia data at a deeper level,such as personal emotional and physiological information in facial videos,which is a greater threat to personal privacy.Therefore,it is of high social value to develop an effective and reliable visual privacy protection processing technology to ensure that intelligent applications can work efficiently while protecting people’s personal privacy.Intelligent video surveillance system is one of the main ways to collect a large number of personal visual data,which is widely used in every place of social life to ensure the normal order and security.However,in the home scene involving personal privacy,the intelligent video surveillance system should not only protect personal privacy,but also ensure good application performance,such as accurately detecting the fall behavior of the elderly or domestic violence beating behavior.In order to meet the needs of both privacy protection and intelligent application,a new privacy protection technology based on bee eye vision bionics is proposed in this thesis.It integrates the two characteristics of bee eye vision: low vision level and target perception,realizes the visual privacy protection at the low-level visual feature level and human visual level,and ensures the high availability of intelligent application at the high-level visual feature level.In addition,a set of visual privacy protection evaluation method is designed to evaluate the visual privacy protection level of bionic vision images.Finally,an association model of visual privacy protection and video behavior recognition is established by constructing an intelligent application system of home behavior recognition,and the coding range of bionic vision is further standardized.The above research work includes the following parts:(1)In view of the problems of visual privacy protection and application performance of image and video data in home intelligent video surveillance system,traditional visual privacy protection processing technology can not meet the needs of both.Considering that bees with special compound eye structure in nature have a lower level of vision than human visual system,and at the same time have the ability to perceive moving objects,this thesis combines traditional compressed sensing method and visual bionics technology to design a visual bionics coding model,namely BCBEV-CS.It combines the low visual acuity and target perception of bee eye vision,and introduces the encryption of random measurement matrix in compressed sensing,so that the encoded image or video can not only achieve the purpose of privacy protection,but also can be used for intelligent behavior recognition.(2)To evaluate the privacy protection performance of the encoded BCBEV-CS video,firstly,at the low-level visual feature level,this thesis verifies that the r PPG signal cannot be accurately extracted from the BCBEV-CS facial video through the current mainstream algorithm,so as to realize the privacy protection of physiological information.Secondly,at the human vision level,in order to test whether BCBEV-CS images achieve visual privacy protection and judge the level of visual privacy protection,this thesis proposes an image visual privacy protection level evaluation method based on dense color features and salient structural features.Through experiments on public privacy image data sets,it is proved that this method can accurately evaluate the level of privacy protection closer to the subjective perception level of human vision.(3)The BCBEV-CS model proposed in this thesis can preserve certain temporal and spatial characteristics of human skeleton movement while protecting visual privacy,so as to bring high performance to intelligent applications.However,excessive visual weakening will lose more temporal and spatial information,which does not meet the needs of practical applications.Therefore,by constructing a home intelligent behavior recognition system,this thesis established a correlation-based statistical model to characterize the quantitative relationship between the visual privacy protection level and behavior recognition accuracy of the coding model,so as to find out the optimal coding range that meets the needs of practical applications and provide guidance for practical applications. |
PDF Full Text Request