Deep Learning Based Person Recognition Methods Under Surveillance Cameras

Person recognition under the perspective of surveillance is a key research direction in the field of intelligent security.Efficient and accurate person recognition algorithms can play an important role in scenarios such as public security criminal investigation that need to track specific targets.However,the pictures taken by the surveillance camera have complicated background information and noise interference,which makes traditional people recognition methods unable to effectively recognize them.Compared with traditional person recognition methods,deep learning methods have greatly improved accuracy and robustness in the field of person object detection and recognition.Therefore,person recognition methods based on deep learning have attracted a lot of attention.This paper improves the single modality person re-identification and cross modality person re-identification from three aspects: convolutional network structure,ranking loss,and identity loss.Compared with the original method,the improved method has improved accuracy in the person re-identification dataset experiment.At the same time,based on the deep learning method,this paper designs a set of people recognition system from the perspective of monitoring,and provides a feasible solution for intelligent security projects.The main work of this article is as follows:(1)A mid-level feature expansion method based on representation learning is proposed to alleviate the problem of overfitting the convolutional network model to the training data set in single modality person re-identification.In this paper,the convolutional network model is modified,and the middle-level features are extracted based on the backbone network for grouping and pooling operations,and then they are stitched with the high-level features to output.Experiments prove that the proposed method can effectively improve the generalization ability of the model and improve the accuracy of person recognition.(2)A cross modality person re-identification training framework based on metric learning is proposed,which improves the problem that person cannot be effectively identified under poor light conditions.This paper uses a dual-flow structure to extract person pictures of different modalities,and transforms the feature extraction module.Based on the transformation of the network structure,this paper proposes a cross modality dual-flow hard triplet loss function to improve the ranking loss to improve the training effect,and uses the focus loss instead of the traditional identity loss to improve the learning weight of the difficult samples in training.The above three methods are tested on the Reg DB dataset and achieved excellent results.(3)A set of person recognition system from the perspective of monitoring was set up,which verified the feasibility of deep learning algorithms for intelligent security projects.The system is built in an indoor environment with multiple cameras and uses YOLOv3 target recognition algorithm and multiple recognition algorithm to perform person target detection and person recognition,respectively.In order to ensure the effective operation of the system,this paper not only designs a set of recognition processes for multiple recognition algorithms,but also designs a multi-thread version recognition process for real-time requirements.The system collected and produced a single modality person dataset through experiments to provide experimental data for the middle-level feature expansion method,which verified the feasibility of the person recognition system.