Research On Person Re-identification Based On Siamese Networks

With the development of urban intelligence,social public safety is received more and more attention by the public.Intelligent detection technologies such as positioning,target identification and comparison of video controlled devices are playing a more and more important role in criminal investigation urban management and tracing.As an important part of intelligent video analysis technologies,person re-identification method has gradually become a research hotspot for scholars in recent years.This paper proposes a solution strategy to solve the problem of low recognition rate of cross-domain person re-identification.Due to differences in camera shooting angles and lighting conditions,pedestrian re-identification has become a challenging pattern recognition task.Aim is to serve the mismatch problem caused by appearing indistinguishable sample pairs,which due to illumination changing,similar wearing and different shooting angles,a person re-identification optimization algorithm based on joint loss and Siamese network is proposed.First,a random erasing algorithm is added to improve the robustness of the model.Secondly,features of images are extracted by the convolutional neural network.Supervised training of these features is applied under a joint loss function which combines focal loss and cross-entropy loss,increasing the attention of the model to the indistinguishable pairs.Finally,a re-ranking algorithm is adopted to reduce the mismatch rate.In order to solve the problem of low recognition rate of cross-domain person re-identification,this paper proposes a solution strategy.First a framework that based cycle Generative Adversarial Network(C ycle GAN)is constructed to perform style conversion between shots of pedestrian images in the dataset.Then the converted images and the original images are combined as the Siamese network input.The network is trained with joint loss optimization.Also a re-ranking algorithm is added to jointly improve the recognition accuracy of cross-domain person re-identification.To evaluate the performance of the algorithms proposed in this thesis,experiments are verified and evaluated on two public datasets,i.e.Market-1501 and Duke MTMC-re ID.The experimental results show that after adding random erasure algorithm to process the image,the Siamese networks with joint loss have higher recognition accuracy than the classical algorithms.By training the pedestrian images generated by the cyclic generation antagonism network,the accuracy of the model for cross-domain pedestrians' re-identification is improved,the proposed network can improve the accuracy of cross-domain person re-identification,and it is better than the recognition accuracy of some current cross-domain algorithms.