Research On Person Search Algorithm Based On Attention Mechanism And Distance Calculation

Person search technology utilizes video surveillance to recognize and search for person images,and is widely used in public safety,transportation,commercial,and other fields.Compared to using traditional digital modeling methods,person search algorithms based on deep learning can more quickly retrieve specific persons in video surveillance,helping the police to search for suspect images among a large amount of surveillance videos,and providing significant assistance in maintaining public safety and investigating crimes in cities.However,in practical applications,existing person search technology has problems with detection and re-identification conflicts,as well as matching errors between text features and image library features,and the accuracy of search results still needs to be improved.To address these issues,this paper aims to improve person search accuracy by focusing on features.(1)The image-based target person search algorithm consists of two sub-tasks: person detection and person re-identification.However,person detection focuses on the commonality of persons,while person re-identification focuses on the uniqueness of persons,which creates a contradiction in the attention points of these two tasks.To address this issue,a person search algorithm based on attention mechanism(Efficient Channel Attention,ECA)is proposed.This algorithm utilizes the idea of attention mechanism to guide the model to focus on features that are beneficial for model performance.By adding ECA modules to the residual blocks of the feature extraction backbone network Res Net50,which uses one-dimensional convolution operation to interact between channel features,the model adjusts the weights based on the importance of the features to achieve a balance between commonality and uniqueness.(2)In text-image-based person search,the mismatch between text features and image features is one of the biggest challenges.To address this issue,a person search algorithm based on the Dynamically Matching Local Information(DMLI)distance calculation module is proposed.This algorithm optimizes the alignment of local features by combining cosine similarity and DMLI methods.Cosine similarity is used to calculate the affinity between two modal local features,while DMLI is used to calculate the distance between local features of each person image and text.With the idea of DMLI,the model can assist in reducing the distance between high-similarity local features and pushing away two modal features that do not match.The experimental results show that the person search algorithm based on the attention mechanism ECA can guide the model to autonomously focus on the key features of the image,reduce the negative impact of detection on re-identification,and balance the two sub-tasks of person detection and re-identification.On the public dataset CUHKSYSU,the algorithm achieves the highest m AP of 91.29% and top-1 accuracy of 92.55%.Furthermore,the person search algorithm based on the distance calculation module DMLI can better align the local features of text and person image,optimizing the matching accuracy of the two modalities.On the public dataset CUHK-PEDES,the top-1 accuracy of this algorithm reaches 65.25%.In summary,the two improved algorithms proposed in this paper can effectively improve the accuracy of person search.