Research On Person Re-identification Approaches For Intelligent Monitoring

With the development of artificial intelligence and smart cities,the demand for intelligent monitoring is increasing.As one of the key technologies for intelligent monitoring systems,person re-identification has received extensive attention with important scientific meanings and broad application fields.For a specific person image,the person re-identification technology applies machine learning and computer vision algorithms to identify and retrieve images of the pedestrian in different cameras and different scenes,which is the important primary work for intelligent video analysis.However,due to changing illumination conditions,complex background clutters,various camera viewing angles,and occlusions,accurate and robust feature extraction is challenging.Therefore,person re-identification is still a challenging topic in computer vision.At present,the application of deep learning to person re-identification has become a research trend.Compared with traditional methods,deep learning technology can obtain more robust and discriminate models.Based on deep learning tools,the research contents and main innovations of this paper are as follows:(1)To solve the problem of low accuracy,which is caused by image space misalignment and occlusion in person re-identification tasks,the paper proposes a novel self-attention model based on deep convolutional neural networks.The model takes advantages of the dependence of adaptively learn image features at different positions and a non-linear combination of multi-scale features.So it effectively highlights the detailed features and information of different images and overcomes the problems of space misalignment and over-reliance on local features.At the same time,the model combines cross-entropy loss with triplet loss for supervised learning,which enables the network to capture common characteristics within the same individuals and significant differences between distinct persons.(2)As a result of generalization of deep learning models is poor,person re-identification has low accuracy in multi-scenario applications.To solve the above problem,this paper proposes a domain adaption model based on transfer learning that a source-to-target transfer model can be trained using annotated source domain and unlabeled target domain.Considering the problem of inconsistent person numbers in different domains and the differences of camera views in domains,this model combines the source and target domains,which uses the limited supervision information to perform transfer learning.It well explores the intra-domain correlations and inter-domain differences.Finally,it jointly unites the supervised term and feature invariance loss function to train a robust feature extractor.(3)Directing at the recognition problem caused by the visual similarity in person re-identification images,a method of hard negative mining was proposed.In this method,the source domain data is used as the reference.Moreover,the similarity among visual features and the invariance of features are also used to mine the positive/negative sample of the data,and the feature memory in the target domain is optimized and updated.Meanwhile,according to the similarity score among samples,multi-label for pedestrians are generated.Finally,soft-label is performed on unlabeled images in the target domain,which combined a variety of self-supervised information for the feature extractor of the training model.At last,the proposed strategy is evaluated on challenging large-scale datasets: CUHK03,Market1501,Duke MTMC-re ID,and MSMT17.Extensive experiments and comparative evaluations show that our proposed model outperforms most of the state-of-the-art methods.