Research On Large Scale Pedestrian Retrieval

Large scale pedestrian retrieval has a wide range of applications in real scenarios,such as monitoring security,check-in access control,and group event recognition and early warning.Related research issues play an important role in artificial intelligence field,which have attracted wide attention in academic and engineering fields for a long time,and become one of the frontier directions with both theoretical and application value.Two main tasks of pedestrian retrieval are pedestrian low-level visual feature retrieval and pedestrian high-level semantic attribute retrieval.For pedestrian low-level visual feature retrieval problem,how to achieve efficient coding and fast retrieval for pedestrian lowlevel features among large-scale dataset is the difficult key problem for large-scale application of pedestrian retrieval technology.For pedestrian high-level semantic attribute retrieval,due to the limited number and the convenience of structured storage of semantic attributes,there is often no speed problem for retrieval.However,the accuracy of pedestrian semantic attribute extraction is unsatisfied.Pedestrian images under surveillance video often have many disadvantages,such as multi-view,fuzziness,noisy and illumination variation,which greatly limit the accuracy of semantic attribute recognition.How to build a proper model which can fully mine the information in training data and improve the performance of pedestrian attributes recognition under the limited conditions is the key problem for research.In addition,in order to satisfy the accuracy and speed requirements of pedestrian retrieval applications,establishing a general query expansion framework to refine the retrieval results of low-level visual features and high-level semantic features is also an key part of pedestrian retrieval research.In response to these challenges,the key contributions of this paper are as follow.Firstly,this paper attempts to use binary hash coding to improve the speed of pedestrian retrieval,and makes a theoretical analysis of the Deep Supervised Hashing(DSH)methods,exploring the reasons why the DSH methods are always with high correlations among bits,which makes it difficult to expand.And then we propose a deep supervised hashing framework without bit correlation.Based of this,we establish a mixed explicit and implicit attribute hashing feature extraction model,which performs well in pedestrian hashing retrieval tasks.Secondly,this paper proposes Grouping Recurrent Learning(GRL)framework for pedestrian attribute recognition to mine both intra-group and inter-group correlations among pedestrian high-level semantic attributes.Under this framework,we first propose a Grouping Recurrent Learning model based on body part generation.And then we propose Recurrent Convolutional model with Conv LSTM to optimize this model.And attention mechanism is added in the Reccurrent Convolutional model.This Recurrent Convolutional model can achieve higher recognition accuracy in pedestrian attribute recognition task without training independent Body Joint Detection Network.This end-to-end attribute recognition framework performs well on accuracy,speed and usability.Thirdly,this paper proposes a query expansion framework based on discriminant model for relevance,using discriminant model learnt from data to predict correlation during the sample selection procedure in query expansion,which optimizes the query expansion process.The method proposed in this paper involves deep learning technology in pedestrian retrieval post-processing.This method achieves similar improvement for retrieval performance compared to the state-of-the-art hand-craft pedestrian reranking algorithm with less time cost.