Person Re-identification In Complex Scenes

Person re-identification is to determine whether a pedestrian appears in other cameras in the monitoring network given a specific pedestrian situation under a monitoring camera.It is valuable to real applications and has received more and more attention in recent years.Although many methods have achieved high recognition rates on general data sets,these methods are still not adaptive to complex practical scenarios.The existing research on person re-identification still has the following problems.1.The absence of real-scene data problem: Existing research data scenarios are simple and cannot truly reflect the actual monitoring data scenarios.At present,person reidentification researchers only focuses on simple scene data such as walking pedestrians during the day.Then in the actual monitoring scene,in addition to the daytime scene data,it also includes the night scene data;in addition to the walking pedestrian data,it also includes the non-walking cyclist data.The latter two scenarios are rarely studied at present.2.Deep network model’s ability for feature representation problem: The feature extraction model based on deep learning has good results,but this type of method has poor interpretability and cannot fully meet the robustness requirements of actual scenes.Therefore,how to design an effective interpretable feature extraction module becomes a key to person re-identification.3.Measurement problem: A robust person re-identification system not only needs to distinguish strong features,but also a robust metric function to measure the differences between features.This problem is often solved by the method of metric learning,so what kind of metric function is designed and how to obtain the parameters of the metric function are the research focus of person re-identification metric problems.Focusing on the above three issues,this paper has conducted in-depth research on the problem of person re-identification in complex scenarios.Based on the actual problem requirements,multiple open source datasets have been constructed and a series of model frameworks have been proposed including:1.Aiming at the problem of absence of non-walking person data,we propose research the bike-person re-identification problem for the first time and we construct the first bike-person re-identification dataset.Besides,a new pipeline for this problem was constructed.Through investigation and distribution,bike-person data accounted for a large proportion in the real monitoring environment.At the same time,bike-person data contains more information to assist person re-identification,so the article first proposed and studied bike-person re-identification.In order to promote the research of this problem,the article first constructs a large-scale real-life scene re-identification dataset,which contains more than 4500 pedestrians and more than 200,000 images.Based on this,the article proposes a symmetry-based adaptive segmentation framework to solve this problem.It performs adaptive segmentation of pedestrians and vehicles on bike-person data,and then calculates the features of each part for final matching.By comparing many feature extraction and metric learning methods in the BPReid dataset,the proposed pipeline achieves more than 50.0 m AP value,which shows the effectiveness of the proposed method.2.Aiming at the problem of absence of night scene data,the problem of night person re-identification in night scene is studied,and the first real-scene night person reidentification dataset is proposed.Besides,a deep learning solution to this problem is proposed.Recently,only a few researchers focuses on the problem of night person reidentification.In order to meet the needs of night pedestrian re-identification,we first constructed the first large-scale real-life nighttime person re-identification dataset,which includes 971 pedestrians under 3 cameras and ovner 300 k pedestrian images.On this basis,the article proposes a framework for night person re-identification based on deep learning.Night pedestrian data is often affected by noise and its quality is reduced,which reduces the effectiveness of the model.Based on the above considerations,this paper proposes a deep model combining joint learning of denoising networks and pedestrian re-identification networks.The proposed model achieves more than 10.0 rank-1 value,which shows the effectiveness of the method.3.Aiming at the problem of poor explanation of deep feature extraction model,an interpretable deep Gabor convolution module is proposed,and a deep Gabor convolution network based on this module is constructed.The person re-identification feature extraction models based on deep networks have strong discriminative power,but the existing deep modules are not interpretable and hinder the stability of the person re-identification system.Traditional Gabor filters have interpretable parameters,but need to be designed manually,while deep networks have the ability to automatically learn parameters.Based on this consideration,the article proposes a deep Gabor convolution module,which automatically learns Gabor convolution parameters through a deep network,while making the learned parameters interpretable.Based on the deep Gabor convolution module,the article builds a deep Gabor convolution network.Experiments on benchmark datasets show the effectiveness of the method.4.Aiming at the problem of imbalance of positive and negative samples in traditional metric learning,an adaptive category center criterion was proposed,and a metric learning objective loss function based on difference minimization was constructed,which was successfully applied to the person re-identification problem.Traditional metric learning techniques often utilize the pairwise constrained criterion,which leads to the problem of imbalance of positive and negative samples.In view of this problem,we introduce the unknown class center to be an optimal variable and a criterion based on the variable is proposed.With the proposed criterion,a metric learning model is constructed.Besides,an efficient algorithm based on matric optimization is proposed to solve the proposed model.It achieves good performance and is superior to most metric learning method.