Research On Deep Learning Based Person Re-Identification Technology

Person re-identification technology is to determine whether the pedestrian appearing under different surveillance cameras belong to the same one.Faced with massive growth of surveillance video,the demand of using computers to re-identify pedestrians in the video came into being.However,the existing person re-id algorithm is mainly in the cropped pedestrian picture to match the search set and candidate set,which is impractical,cropped pedestrian in the practical consideration can't be given directly,the target pedestrian needs to be detected.At present,deep learning has achieved excellent results in image recognition,speech recognition,natural language processing and other fields.Compared with the artificial features,deep learning can automatically learn better features from the original image data to classify objects,have more practical significance.The application of the deep learning to the person re-identification has become the current research hotspot,but because of the current person re-identification has some problems that make it from the practical application of a long distance.This paper summarizes some commonly used features,algorithms and deep neural network structures for pedestrian detection and re-id,and conducts in-depth research and analysis.This paper proposes a pre-training network model for end-to-end pedestrian re-recognition.The model combines the two network structures of verification and classification,and uses the spatial pooling operation to normalize the input images of different scales.On the basis of this pre-training model and ResNet-50 network structure with good performance,the end-to-end person re-identification network structure has been promoted.Then,the improved model is trained on the caffe deep learning framework and the experiment is carried out,including the effectiveness of the pre-training model and the effect of different feature dimensions on the effect of the network model.In the different size candidate sets,low resolution and occlusion subsets performance analysis,and compared with some advanced algorithms.The experiments show that the model trained in this paper can learn high robustness features and improve the recognition rate of person re-identification.