Research On Person Re-identification Technology Based On Deep Learning

With the development of artificial intelligence,deep learning has been playing an important role in person re-identification(re-id),among which the convolutional neural network has become the mainstream method to extract features of images.However,due to the lack of samples in some person re-identification datasets,a deep convolutional neural network model is hard to be trained and prone to be overfitting.Furthermore,person attributes recognition has been one of the most crucial tasks in person re-identification.Focusing on the above problems,this paper designs a deep convolutional neural network model for person re-id.The model contains several specific multi-layers convolutional structures,which concatenates the input and output of convolutional layers on the channel dimension to make the feature be reused and enhance the flow of information in the whole network.Such architecture is capable to restrain the overfitting and make the extracted features more discriminative.Moreover,a combination of three loss functions is employed to update the parameters of network model.Such combination of loss functions could extend the inter-class distance and decrease the intra-class distance which makes features of people more distinguishing.Furthermore,this paper transforms person attributes recognition problem into feature extraction problem in time series and designs a person attributes recognition network based on long-short term memory and attention mechanism.On the basis of the identification results from convolutional neural network,an attributes-aided re-ranking algorithm is proposed according to attributes recognition result to improve accuracy of person re-id.Finally,this work designs a person re-id system based on person image database to verify the practicability of methods in this paper.Experiments on three public datasets of person re-id demonstrate that the designed deep convolutional neural network could reduce overfitting caused by less train samples effectively and obtain outstanding accuracy.Experimental results also verify the availability of person attributes recognition and attributes assist re-ranking algorithm.