Research And Implementation Of Person Identification System Based On Intelligent Video Analysis Technology

Person identification system under video surveillance is an intelligent identification system based on computer vision technology for automatic detection and identification of individuals within the camera surveillance range.Such systems are widely used in security,commercial and transportation fields,improving surveillance efficiency and security.Face recognition and pedestrian re-identification are two key technologies for person identification systems.As a strong discriminative biometric feature,face is widely used for identification,but it is usually difficult to capture a clear frontal view of a person’s face due to the complexity of the surveillance environment.Pedestrian images under video surveillance are easier to acquire than faces,and pedestrian re-identification achieves identification purposes by deep mining and extraction of pedestrian image features to match the same object at different times and scenes.In intelligent surveillance,pedestrian re-recognition and face recognition can complement each other.When the pedestrian face cannot be clearly captured,the pedestrian can be identified by comparing the pedestrian candidate library through pedestrian re-recognition technology;when the pedestrian face can be clearly captured,the pedestrian can be identified by face recognition technology,and the recognition result and the pedestrian image can update the pedestrian candidate library.In order to improve the accuracy and robustness of person recognition,this thesis designs and implements a person recognition system in which pedestrian re-identification and face recognition work together,and the main research and work are as follows:(1)In this paper,an improved lightweight multiscale feature model MHSA-OSNet is proposed.MHSA-OSNet replaces the maximum local perceptual convolutional stream in OSNet with a multi-head self-attention(MHSA)structure,which enables the model to dynamically fuse global and local features and facilitates the extraction of discriminative features.The mean average precision(mAP)of MSMT1501 dataset is improved from 52.9%to 55.96%.The lightweight network structure brings smaller number of parameters and computation,which can be well suited for intelligent video surveillance scenarios.(2)In this paper,we propose a pseudo-labeled cluster generation method suitable for unsupervised domain adaptation for pedestrian reidentification,called divided clustering secondary combination DCSC.DCSC utilizes the camera IDs and relative acquisition times that come with the pedestrian reidentification dataset at the time of acquisition,groups the pedestrian images according to the different acquisition cameras,and clusters them within each group relying on the acquisition relative times and feature distances,so that the clustering results of each group have better intra-cluster tightness.The clustering results are recombined in the secondary combination using all images clustering results and group clustering results against each other,which makes the clustering results have better inter-cluster separability.The clustering results of DCSC are applied as pseudolabels in Self-paced Contrastive Learning(SpCL),which can improve the mAP of pedestrian re-identification domain adaptation learning from MSMT17 to Market1501 from 77.5%to 81.6%.(3)In this paper,a personnel recognition system with pedestrian re-identification and face recognition working in concert is designed and implemented.The pedestrian rerecognition algorithm of the system uses MHSA-OSNet as the backbone model.The pedestrian images are first collected in the application scenario,and the collected pedestrian image dataset is used as the target domain data without labeling,and the four labeled pedestrian reidentification datasets,Market1501,DukeMTMC-reID,CUHK03 and MSMT17,are used as the source domain data,and unsupervised domain adaptation learning is performed using SpCL and DCSC to make the backbone model target scenario to extract discriminative pedestrian features.The system runs with face recognition results to dynamically update the pedestrian candidate pool,and detected pedestrians are identified by comparing the pedestrian candidate pool.The designed and implemented person identification system is tested to have good accuracy,real-time and stability.