Research Of Weakly Supervised Person Re-identification Algorithm Based On Deep Learning

With the development of society,more and more public places have deployed video surveillance networks,but it is the key to ensure public safety to analyze the information conducive to criminal investigation from a large number of surveillance videos.Person reidentification(Re-id)task can retrieve the image consistent with the identity of the target person from multiple non-overlapping cameras,so as to locate the target person quickly and effectively.In recent years,with the development of artificial intelligence technology,the fully supervised person re-id proposed by researchers based on deep learning can achieve impressive recognition performance.However,expensive manual labeling cost limits the application of person re-id algorithm in practical scenes.In order to address this problem,the paper explores and studies the weakly supervised person re-id models.The setting of weakly supervised person re-id refers to annotating each video clip in the dataset with a video label which indicates the identities appearing in the video clip.The labeled video clip is called a sample bag.The precisely corresponding relationships are unknown between the person images and the identities in the video label.Therefore,this paper focuses on the setting of weakly supervised person re-id.The main research contents and innovations of this paper are as follows:Firstly,the intra-bag tracklet feature clustering-based algorithm is proposed for weakly supervised person re-identification.Specifically,due to the weak nature of the label information,the associations between the person images and the identities in the video label is unknown.Therefore,this model first assigns a pseudo label by leveraging the prediction probability and video labels for each person image,and then the discrimination person features are learned by using the assigned pseudo labels.In order to reduce the distance between person image features in a person tracklet,person tracklet feature clustering learning is further introduced to pull the feature distance between images in the tracklet closer and push the negative person images away from each other.Then,the inter-bag tracklet association learning-based algorithm is proposed to solve the problem of noisy tracklet in weakly supervised person re-id task.Specifically,when the detection and tracking algorithm is not perfect,some noisy tracklets will be introduced into bags,which will interfere with the performance of the model.In order to address this problem,a inter-bag tracklet association learning-based weakly supervised person reidentification is proposed in this paper.The proposed algorithm is composed of person noisy tracklet filtering and tracklet association learning.Specifically,each person identity is associated with an effective tracklet through the person noisy tracklet filtering,and the effective tracklet is used for person identity feature learning,so the interference of noisy tracklet is avoided.However,for person re-id task,the person noisy tracklet filtering only considers each bag in isolation without exploring inter-bag tracklet associations.We address this problem by introducing a tracklet association learning,which mine reliable positive tracklet pairs and hard negative pairs between bags.And then we pull the positive tracklet pairs closer and push the hard negative pairs away in the feature space.Finally,a large number of experiments are carried out on the WL-MARS and WLDuke V weakly labeled datasets.Compared with the person re-id models proposed in recent years,the proposed models can achieve better recognition performance.And the robustness of the proposed model is verified from many aspects.