Research On Pedestrian Fine-grained Recognition And Re-identification Technology

With the development of Internet technology,intelligent video surveillance systems have penetrated into every aspect of life.Pedestrian detection and re-identification are the major researches in video image analysis.However,the camera's location,angle,background information,lighting conditions,and pedestrian posture bring challenges to pedestrian detection and re-identification tasks.The existing detection and re-identification methods often have the disadvantages of inaccurate expression of pedestrian image features,low learning efficiency and low accuracy of feature distance measurement.In view of the deficiencies of the existing research methods,The author scombine the theory and methods of deep learning to study some key issues in pedestrian video recognition,pedestrian reidentification,and pedestrian's hash retrieval in intelligent video surveillance systems.The main tasks of this paper are as follows:1.The main difficulty in pedestrian fine-gained recognition is the identification of pedestrian attributes.The author propose a pedestrian fine-gained recognition method based on multi-task joint and multi-model fusion learning to improve the accuracy of pedestrian recognition.First of all,the method completes pedestrian detection and secondary detection of sub-components based on YOLOv2 to achieve accurate positioning of pedestrian subcomponents.Secondly,the recognition of pedestrian attributes is completed.Due to the higher semantics of pedestrian gender attributes,The author propose a method of identifying gender attributes based on multi-task learning,and use a combination of pedestrian gender attributes and other attributes to improve learning ability.In addition,the method integrates VGG16 and ResNet50 to achieve identification of pedestrian jacket style,undershirt style,shirt color or undershirt color.Experiments show that the recognition accuracy of pedestrian gender attribute reaches 78%,and the accuracy of the attribute of long and short hair reaches 88%.The recognition accuracy of the clothing style attribute reaches more than 80%,and the identification accuracy of clothes color attribute is as high as 92%.2.The main difficulty of pedestrian re-identification is the extraction of pedestrian signatures with distance metrics.The author propose a pedestrian re-identification model combining with subcomponents,and extract pedestrian signatures in combination with subcomponents.The model takes the triple image as input,and extracts the global ID and sub-component ID of the pedestrian image based on the idea of pedestrian target natural region division and ResNet-50 structure.Finally,adjust the image feature ID by minimizing the "difficult sample sampling triplet loss function" so that the spatial distances of feature IDs of the same pedestrian target image are closer,and the spatial distance of feature IDs of the different target image is farther.The model not only considers partial image information but also can learn global image information,and can make a more comprehensive description of the pedestrian features.The experimental verification of the model on the data set VIPeR shows that the Rank-1 is 53.3%.Experiments show that the pedestrian reidentification model of combining with subcomponents has higher recognition accuracy.3.The purpose of the image hash retrieval is to retrieve the image efficiently,and the main difficulty is how to convert the image into a hash code.The author propose a pedestrian hash retrieval model based on metric loss.The model extracts pedestrian image features based on convolutional neural network and considers the quantization loss in image feature learning.The author joint measurement loss and classification loss to achieve the automatic learning of the hash code.Experiments show that the classification loss makes the image can learn semantic features with strong distinguishing ability.The introduction of measurement loss makes the information loss of the hash code lower.The model improves the CMC value of pedestrian search effectively.