Design And Implementation Of Person Re-Identification System Based On Deep Learning

As the main body of many places and events,pedestrians are the focus of attention in the intelligent monitoring system.Therefore,person re-identification is the main technology support of intelligent monitoring system.This technology is used to re-identify whether a target person from a surveillance camera appears in other places.Person re-identification technology has extremely high application value in the fields of security monitoring,smart retail and military reconnaissance.However,under the influence of complex environment,the visual features extracted from network are vulnerable to interference of background,which reduces the accuracy of system.At the same time,with the rapid increase in the number of monitoring devices,massive image data need to be stored,and the amount of computation is huge,existing Re-ID systems are difficult to meet the real-time demand.In addition,for some specific scenes that can provide both the person image and the semantic attribute descriptions,existing Re-ID systems have limited functions and can't make full use of the attribute description.Therefore,it is necessary to construct a person Re-ID system that can meet the high retrieval accuracy and real-time requirements,and can also provide extended function for specific scenes.In view of the above questions,this paper proposes solutions respectively and builds a person re-identification system based on deep learning.Specific works are as follows.(1)From the aspect of accuracy,the local regions of the person image contain both the target person and the background information,which lead to the problem that the features learned by the deep network contain background noise,thus affecting the performance of the Re-ID algorithm.In this paper,an adaptive local feature aggregation network is proposed.This paper designs an adaptive aggregation module to guide the network to automatically focus on the target person and weaken the influence of background information on feature learning.The validity of the network is verified on three person datasets.(2)In terms of real-time performance of the system,this paper introduces a deep hashing learning module on the basis of the adaptive local feature aggregation network,to meet the real-time performance of algorithm application and small computation storage in real scenes.The algorithm uses pyramid structure and self-distillation learning to learn multi-dimensional discriminative hashing codes for person images,which can improve the calculation speed and reduce the memory space.At the same time,in the retrieval stage,the Coarse-to-Fine hashing retrieval strategy is implemented to meet the real-time speed and ensure the performance of Re-ID.(3)From the aspect of function expansion,this paper proposes a multi-attribute and multi-modal person re-identification function for the specific scenes which provide both the target person image and the semantic attribute description.Semantic attributes and visual images are complementary to each other.This paper intends to combine them to realize the person re-identification function in multi-modal.(4)Based on the above three works,this paper designs and implements a person re-identification system based on deep learning,which meets the requirements of realistic application of accuracy,real-time performance and functional expansibility.The experiments and test results show the effectiveness of the solutions proposed in this paper.