A Research For Pedestrian Re-Identification Based On Parallel Attribute Learning Incorporated Posture

With the development of projects such as“smart city"and"safe city",the task of pedestrian re-identification in actual scenes has become increasingly onerous.Pedestrian re-identification has become a hot topic in the field of intelligent image analysis.Due to the huge number of pedestrians in the surveillance videos,the task of examining data in the videos would be hard,inefficient and feasibly missed if only depend on manual examination,especially in the field of criminal investigation which requires real-time identification of the target person.Therefore,the introduction of deep learning methods to analyze data breaks through the bottleneck of processing pedestrian data set.However,most of these methods are based on pedestrian low-level visual features or middle-level attribute semantic features,and the effects caused by the differences of cameras are not taken into account.In addition,the appearance of pedestrians is susceptible to the wearing,the partial occlusion,the posture,the shooting angle of the camera and other factors,and there may be one or more similar target pedestrians in the surveillance field of view,affecting recognition.To solve these problems,we conduct a more in-depth study and propose a parallel attribute learning method that incorporated posture.The main works of this thesis are described as follows:(1)A method of pedestrian re-identification based on posture and diversified attributes is proposed.Currently,there is a lack of study of pedestrian posture among the existing pedestrian re-identification methods,which leads to the problems such as false detections and missed detections of pedestrian attributes.We propose to integrate the pedestrian posture into the category of pedestrian attributes.Postures of pedestrian are divided in order to reduce the occlusions of the postures on the pedestrian features and the effects of various attributes of the bodies First,we obtain the pedestrian joint point images.The overall image of the fourteen joint points and the images of each joint point are all taken as input to describe the fine-grained posture feature of pedestrians.Then,by comparing the inputs of coarse and fine-grained posture and their classification effects of the model on the target dataset,one of the granularities is determined as the final input of the posture learning model.Finally,the obtained gesture label is integrated into the attribute diversification learning process,and then a new attribute learning model is established.The experimental results show that our method uses the convolutional neural network to learn the diverse attributes of pedestrian,which not only optimizes the classification of pedestrian attributes,but also improves the accuracy of pedestrian re-identification(2)A method of pedestrian re-identification based on attribute and parallel gradient descent algorithm is put forward.For the shortage of research on attribute categories for the attribute learning,and the bottlenecks of parameter exchange caused by complex network calculations,we establish a parallel structure based on pedestrian posture and diversified attributes.First of all,the pedestrians are divided according to the body structure,avoiding the attribute loss caused by the segmentation.Then,a six-hierarchy parallel network model is established,of which five block data and pedestrian images are simultaneously used as training targets to solve the problem of slow calculation.Finally,a parameter exchange period is proposed to accelerate network training.Six slave nodes and one master parameter server are established as well.In addition,the combination of momentum correction and adaptive learning rate in training can solve some problems,for example,the previous parameter training is difficult to adapt to the features of the data set,and the inappropriate learning rate results in unstable convergence process.The experimental results show that our method improves the training speed without affecting the accuracy of pedestrian recognition(3)A re-identification prototype system combining posture and attributes is designed and realized.Aiming at the proposed method,we design modularly and realize a pedestrian re-identification prototype system combining posture information and attributes.The above posture and attribute learning are taken as the core module.In addition,the image input module,the target pedestrian module,the targetless pedestrian module and the recognition result module are also included.The feasibility and practicability of the system are verified by experiments.