Research On The Re-Identification Methods Of Suspect Target For Night Video Investigation Scenes

With the widespread use of networked and intelligent video surveillance networks,video image information has been widely used in detecting the activities of suspected targets.By analyzing the video image information of suspected targets and completing the association of information such as person and vehicle photos across cameras,the activity trajectory of suspected targets can be obtained,enabling tracking and capture of suspected targets,which is called suspect target re-identification.Although significant progress has been made in suspect target re-identification under natural light environments in recent years,research in night surveillance scenes is rare.On the one hand,data acquisition is difficult: the difficulty of collecting and labeling nighttime images leads to a lack of available datasets;on the other hand,target matching is difficult: images captured at night lack sufficient detail,making it difficult to identify unique features for visual processing.To address these challenges,this study focuses on re-identifying person and vehicles in night scenes,with the following details:(1)A night person re-identification algorithm based on generative adversarial network is proposed to address the problem of missing identifiable features of person in night environments.Firstly,this method uses an image degradation module to generate multi-level night images.Then,an adversarial recovery module is used to restore the original,well-lit images as labels and an improved U-Net model is used as the generator to preserve all information in the original images.Finally,the recovered data is input into a convolutional neural network to extract person identification features.(2)A night vehicle re-identification algorithm based on mean average precision loss is proposed to address the problem of significant appearance changes of vehicles in night environments that can interfere with retrieval.This method consists of two parts: metric learning and mean average precision optimization.Metric learning optimizes the model using triplet loss and cross-entropy loss functions,while mean average precision optimization replaces the classification task with the sorting tasks and completes the sorting problem using a mean average precision loss function to directly optimize the mean average precision value during training.To evaluate these methods,the re-identification datasets for person and vehicles in night scenes were constructed,and experiments were conducted.The results show that innovative technologies such as generative adversarial network and mean average precision loss can significantly improve recognition performance of suspect target reidentification in nighttime surveillance scenarios,which is of great value and relevance to the development of criminal investigation and smart security.