Research On Person Re-identification Based On Multi-task Learning

With the gradual increase of people's safety awareness,surveillance cameras have been installed in crowded shopping malls and even in individual residences.The main task of person re-identification is to extract a valid representation of a specific pedestrian from the image captured by a certain camera,and then retrieve the pedestrian again from the image captured by other cameras.At present,the person re-identification algorithm based on deep learning is mainly researched in two aspects: pedestrian feature expression and pedestrian similarity measurement.As the extracted pedestrian information is not detailed enough,obtaining a more comprehensive pedestrian feature representation is the focus of this thesis.This thesis proposes two neural network models,namely MTL-Res Net50 model and FA-MTL-Res Net50 model.In response to the problem of only focusing on pedestrian identification in the past,the MTL-Res Net50 model designs a multi-task network,where multiple related tasks provide supplementary information from different angles to improve the learning ability of the network.The model integrates three different tasks,including pedestrian identity classification tasks,identity verification tasks,and pedestrian attribute classification tasks that assist feature extraction.Three network branches are used to train a robust multi-task network.In order to improve the accuracy of pedestrian description,the model uses the correlation between each pair of attributes to design an attribute prediction module to correct inaccurate attributes and restore missing attributes.The introduction of attribute classification tasks not only improves the accuracy of attribute recognition,but also provides assistance for identity recognition and improves the accuracy of person re-identification.The FA-MTL-Res Net50 model introduces an attention mechanism on the basis of the MTL-Res Net50 model.Aiming at the problem of pedestrian image misalignment,this method divides the features into blocks,combines channel attention and spatial attention in the residual module,and assigns different weights to different areas of pedestrians.By extracting better pedestrian feature representations,the effect of person re-identification is effectively improved.The thesis compares the above two models with the current more advanced algorithms on two large-scale pedestrian image datasets Market-1501 and Duke MTMC-re ID.The experimental results show that the method proposed in this thesis obtains good results in both person re-identification tasks and attribute recognition.The recognition accuracy of the multi-task network is better than the existing multiple methods for a single task,and the fusion attention mechanism can make the extracted features more stable,and the recognition effect is further improved.