Study On Person Re-identification Algorithm For Intelligent Video Surveillance

Person re-identification refers to the retrieval of other pictures containing the person under the cross-monitoring equipment for a given person picture,which can be used in railway stations,hospitals,Banks and other places with large flow of people or key security places to identify fugitives and search for abducted children.This is another research with important practical value after face recognition and license plate recognition,which is of great significance for building smart cities and intelligent transportation.In person re-identification,it is very important to extract the scale-invariant features in time and space.Compared with ordinary target recognition,the following three factors increase the difficulty of extracting the effective characteristics of persons : due to the difference in shooting direction and shooting time,the person attitude changes;due to the change of light intensity at different times,the color of person clothing will also change;as the types or specifications of security cameras may be different,the resolution of person pictures taken will be inconsistent.In order to solve the above problems,the main research contents of this paper are as follows:A strategy of combining person attributes with identity learning is proposed for the change of person attitude and appearance color.Firstly,quadruplet image pairs were input into the pre-training model Res Net50 to extract person attributes and identity characteristics information,and then the feature information by minimizing status Softmax loss and attribute Quadruplet loss adjusting network model parameters,with persons identity and attributes of network training,finally the trained model used in person re-identification task.Experimental results show that this method can improve the accuracy of person re-identification.Considering that the collected person image may have inconsistent resolution,that is,the person image collected from one perspective is of low resolution,while the person image collected from another perspective is of high-resolution,a method combining the a single image super-resolution reconstruction with person re-identification is proposed.Firstly,a person image super-resolution network is constructed.The convolutional layer is used in each layer of the network,and the super-resolution network of images is trained jointly with the loss of pixel mean square error,perception loss and texture matching loss.Then the trained model is used to sample the low-resolution person image from a perspective into a high-resolution person image.Then,the high-resolution image from the upper sample and the high-resolution image from another perspective were input into the person network Res Net50 to extract feature information,and the Softmax loss and Triplet loss monitoring model were trained respectively.Experimental results show that the reconstruction of high-resolution person image by this method is effective in both objective image quality evaluation criteria and subjective vision.Compared with the existingadvanced person re-identification method,this method has higher accuracy and wider application range.