Pedestrian Re-identification Based On Clustering And Unsupervised Learning

Pedestrian re-identification is one of the more popular research directions in the field of computer vision in recent years.It is a specific recognition technology that can effectively judge pedestrians appearing in pictures or videos.With the development of science and technology,pedestrian re-identification technology has been widely used in security,video surveillance and other fields.Traditional face recognition technology has been difficult to meet the needs of people in today's society.Popularization and better service to society is the key to the development of this technology.At present,pedestrian re-identification has achieved a relatively large breakthrough in labeled supervision areas,and the performance has been significantly improved.However,it is very difficult to perform unsupervised training on unlabeled data.This is mainly reflected in the fact that it is challenging to adapt the target domain pictures without id annotation to the domain adaptation.Therefore,the technical difficulties in the unsupervised field are the main problems faced by current pedestrian re-identification.In the early days of research,the pedestrian re-identification technology in the unsupervised field only had three methods: dictionary-based,manual extraction of features,and exploration of local features.However,due to the inability to avoid the interference of external conditions,unsupervised technology has always been impossible.Go beyond a supervised approach.Therefore,how to effectively apply the unsupervised technology to the problem of pedestrian re-identification is also an important development direction in the field of pedestrian re-identification in the future.Based on the problem of imperfect current unsupervised technology,fromthe perspective of the problem,we improve it based on the bottom-up clustering method based on the unsupervised state to gradually improve the performance of the model in the unsupervised field.Specifically,it is divided into three steps:the initial stage uses convolutional neural networks to extract features from the picture,and the SE module and self-attention mechanism are added to the convolutional neural network to optimize the network;then clustering and merging of similar features is performed.;Then retrain the model based on the combined results until the convergence training is over.The experiment was performed on two major pedestrian re-identification datasets Market-1501 and DukeMTMC-reID.The final results show that our proposed improved method has a significant improvement in the recognition effect compared with the original unsupervised clustering method.