Person Re-identification Technical Research Via Transfer Learning

Person re-identification points to the subject that searches whether a specific target character in a video also appears in other video with regard to surveillance video data that does not overlap in time domain and space domain.In recent years,person re-identification has aroused the interest of a vast number of researchers in the field of deep learning,and they have put a lot of time and effort into this field to carry out corresponding research,putting forward many excellent algorithms.The character images in the surveillance video data are often have a low resolution,strong light changes and various character movements and gestures,which will cause great differences between images for the same person in different video data and add great challenges to person re-identification.In addition,there exist several datasets for multi-shot person re-identification in academia,however,most of them have only hundreds of video sequences,which is far from sufficient for training a high performance multi-shot person re-identification model and is unfavorable to enhance the training effectiveness of deep neural network.This paper explores the problem of person re-identification based on video sequences.We first construct a deep neural network by the convolutional neural networks and recurrent neural networks,which takes video sequence data as input data.After that,based on this deep neural network,the training effect is improved through the method of transfer learning,which is used to solve the adverse effect on network training due to insufficient training data.There are two kinds of transfer learning methods that are adopted: transfer learning based on pre-training method and transfer learning method based on cross-modality alignment module.This paper first proposes a baseline network for video-based person re-identification network.The three-depth convolutional neural network is used to extract static features of each image in the video sequence,and learn distinguished features from local to whole and from abstract to concrete.Then,recurrent neural network is used to extract temporal information of the serialized input features,and the input video sequences are expressed by more expressive features.The temporal pooling layer connected to the recurrent neural network further makes the abstracted feature more expressive,and enter features into the error functions to prepare for the back propagation calculation.On this basis,two kinds of person re-identification network structures based on transfer learning are proposed in order to solve the lack of video sequence data in existing video dataset.The first method is to use external person re-identification dataset to pre-train the network,and then use the target dataset to further fine-tune the entire network.In the pre-training stage,the network parameters have been preliminarily trained and the training of next stage is initialized by these parameters.The training in the fine-tune stage is effectively guided by the pre-train method,accelerating the convergence process and improving the learning ability of the neural network.The second method is to utilize the pseudo sequence data which is generated by using single-frame image data,and the pseudo sequence is trained in the way of joint training with the real sequence data.The learned knowledge of the pseudo sequence is transferred to the sub-network which utilizes the real video sequence,to enhance the effectiveness of deep neural network.In order to eliminate the mismatch between single-shot images and video sequences,a cross-modality alignment module is proposed to generate sequence data from single-shot images,which guarantees the reliability of transfer learning.The experimental results of the proposed methods are compared with the results of some other existing excellent algorithms.The practicability and effectiveness of the proposed methods are proved.