Research On Person Re-identification Based On Fusion Of Deep Features And Traditional Features

As people pay more and more attention to public safety,the video surveillance system is gradually applied in various public situations.When the amount of video surveillance data increases,the traditional human-eye viewing mode seems to be inadequate,and non-overlapping multi-camera person re-identification technology came into being.However,person re-identification is still one of the most challenging research directions due to the influence of light changes,perspective changes,background differences and pedestrian attitude differences in the actual monitoring network till now.At present,the researches on person re-identification are mainly divided into three aspects:traditional features,deep features and metric learning.In order to reduce the impact of occlusion,different perspectives,light changes and other factors,this thesis designs more distinctive feature expressions from traditional and deep learning directions.The main research work and innovation points are as follows:1.Firstly,according to the influence of light and the limitation of the single feature,this thesis designs a traditional hand-crafted feature descriptor which is a combination of color feature and texture feature.Secondly,this thesis proposes the concept of Damaged Region to handle the occlusion problem to some extend,and further proposes a person re-identification algorithm based on local region partition.Specifically,the pedestrian images have been enlarged and divided directionally to separate damaged areas and weaken the impact of viewpoint changes.Finally,an effective area judgment strategy is adopted to reduce the proportion of the damaged area in the final score to solve problems such as occlusion and background differences,and improve the final pedestrian recognition rate.2.The traditional feature is the observer's capture of one particular aspect of the target object,while the deep feature is the intrinsic and abstract feature automatically learned by the deep neural network based on the large amount of input data.In order to exploit the advantages of both and explore the synergy between them,this thesis learns the deep feature expression of the target by optimizing the deep network,and fuses the traditional features with the deep features.At the same time,with the purpose of improving the adaptability of the deep network,this thesis fuses the sample images from the three public datasets we used.Compared with the results of the method which only use traditional features,the experimental results of this these further prove the effectiveness of the method.