Indoor Privacy-Preserving Action Recognition Via Partially Coupled Network And Super-Resolution Network

As computer vision is more and more integrated into daily life,it is very necessary for indoor monitoring systems to enhance home safety,assist human life,and deal with emergencies in a timely manner to ensure personal safety.Action recognition and analysis in indoor scenes is a very important application direction in intelligent surveillance.By analyzing indoor surveillance video information,smart devices can obtain information about what is happening in the environment and make corresponding responses and interactions.However,due to privacy considerations,many users worry that the camera will record privacy-sensitive images or videos,which may lead to privacy leakage and security risks.The indoor surveillance system also poses a huge social challenge: user privacy.Therefore,it is of great economic and social value to find a method of action recognition that takes into account both security guardianship and privacy-preserving.In order to solve the contradiction between user privacy and action recognition in intelligent surveillance applications,we propose a method of fusing super-resolution network and partially coupled network learning,which has the effect of indoor environment privacy-preserving and completes the task of action recognition at the same time.The main research content includes the following parts:(1)Privacy-preserving data processing.We first use low-resolution images with strong fuzziness,the original image to achieve the effect of privacy-preserving.Firstly we use three methods of interpolation to interpolation of the original video.Secondly we extract optical flow features from continuous video frames,and use multiple optical flow stacking to represent the video behavior.Finally,in order to improve the recognition effect of the neural network,a series of experiments are carried out on the image resolution,interpolation method and the number of optical flow stacking frame,and the parameters suitable for this method are selected.(2)Partial Coupled Convolutional Network(PCCNN)action recognition.Compared with high-resolution images,low-resolution images have less feature information.In order to improve the accuracy and robustness of the recognition model,high-resolution data is added to the training.On the one hand,it is the view of data expansion.High-resolution images are high-quality training data;on the other hand,the view of domain adaptation.There is a certain mapping relationship between high-resolution images and low-resolution images.Most of the existing network models are based on high-resolution data.so this paper proposes a 5-layer convolutional neural network suitable for low-resolution data,with 4 convolutional layers and 1 fully connected layer.Finally,in order to find the number of shared network layers between high-resolution data and low-resolution data,we design the number of shared network layers and determine the appropriate coupling parameters through experiments.(3)In order to further improve the performance of the algorithm,the above method is improved,the super-resolution network and part of the coupled network dual-network integration of privacy protection action recognition method.After low-resolution processing,the optical flow features have fewer identifiable features.We proposes to use super-resolution networks for feature reconstruction to improve the practicability of optical flow features and make it more conducive to network recognition.The comparison with other advanced methods shows that the action recognition method for indoor privacy-preserving proposed in this paper has higher recognition accuracy.