Video Action Recognition Under Compressed Sensing Visual Privacy Protection Framework

In the research field of computer vision,video action recognition with complex semantic information has been a very challenging research topic.For video action recognition,the current mainstream deep neural networks have made significant progress.Computer and internet technologies bring convenience to people's lives,but people's privacy which is mostly in videos and images has been threatened to a certain extent because of their real-time acquisition and analysis to the information.Moreover,for some special applications(such as home environment),people also hope that the monitored video content can be kept secret,so as not to be leaked and cause unnecessary losses.However,existing video action recognition frameworks may not achieve a good balance between privacy protection and recognition accuracy.Therefore,the research on video action recognition for visual privacy protection has important significance and application value.To solve the problem of video action recognition under the visual privacy protection and the real-time problem of system operation,the original intention of this study is to hope to design an efficient and accurate video action recognition monitoring system with privacy protection advantages to achieve a balance between visual privacy protection,recognition accuracy and operating efficiency.Therefore,this paper proposes a video action recognition method under compressed sensing(CS)visual privacy protection freamework,which mainly includes the following three key steps.(1)Visual shielding compressed sensing(VSCS)coding.This paper first uses the advantages of CS theory to address the issue of privacy protection and proposes the VSCS,which performs multiple layers of CS dimension reduction on video action,which greatly reduces the amount of data and realizes visual shielding.At this time,the naked eyes can no longer recognize the video content,which achieves the effect of privacy protection.Secondly,to improve the subsequent recognition effect,through a series of experiments,a suitable measurement matrix and dimensionality reduction layers for the needs of this paper are selected.(2)VSCS video sequence action representation.To improve the efficiency of the system operation,this paper uses the speed advantage of the convolutional 3D(C3D)neural network model to efficiently characterize the video action characteristics in the compressed state.Because the input is VSCS video,this paper first preprocesses the input video,secondly changes the existing C3 D structure slightly and retrains the improved model,and finally uses the principal component analysis technology to reduce the dimensionality of the extracted feature vectors to reduce the time complexity of the system.(3)VSCS-Ad SRC video action recognition.For inputing VSCS video under visual privacy protection,to improve recognition performance,this paper proposes a new recognition algorithm Ad SRC which effectively combines the robustness advantage of sparse representation based classification(SRC)for visually degraded face images and and the idea of Ada Boost integrated classification.This algorithm makes up for the shortcomings of SRC for recognition in complex scenes.To verify the effectiveness of the proposed algorithm,this paper performed experiments on three common mainstream video action databases.The results show that the proposed method is robust in video action recognition and has the advantage of visual privacy protection.