| Person re-identification is a technology that determines whether a specific pedestrian that appears belongs to the same target through cross-domain cameras,and has been widely used in intelligent security,traffic safety and other fields.The traditional person reidentification method mainly uses the color,texture and other feature information of the image for identification.Although it is simple and fast,the feature extraction ability is limited and it is easily disturbed by the external environment.In recent years,person reidentification technology based on the feature extraction method of convolutional neural network has developed rapidly and showed a good development prospect.However,in real scenes,the recognition effect of person re-recognition technology is poor due to factors such as diverse pedestrian poses,occlusion by obstacles,and low image recognition rate.Therefore,how to accurately and effectively extract the characteristic information of pedestrians from the image is the key and difficult point to improve the recognition effect.Based on deep learning technology,this paper conducts the following researches on person re-identification in unoccluded and occluded pedestrian images:(1)Aiming at the problem of ignoring the lack of connection between pedestrian parts when the image is not occluded,a person re-identification method based on feature fusion is proposed.This method establishes the connection between pedestrian body parts by fusing the average pooling and local features of the image,and introduces the pedestrian global feature to optimize the feature extraction process,so as to eliminate the influence of the background noise of the image,and then obtain a more discriminative feature expression,further Improve model performance.In addition,triplet loss and cross-entropy loss are used to jointly optimize the model to improve the generalization ability of the model.(2)Aiming at the problem that the local features contain invalid information of the occluded area when the image is occluded,a person re-identification method based on relational network is proposed.This paper uses pose estimation to extract pedestrian key point information,weakens the influence of background factors,and builds a relationship network of key points through relational feature fusion to maintain the intrinsic relationship between key points.The feature similarity measure is used to reduce the influence of the occluded parts of the image,so that the network can focus on the effective information of the unoccluded area,and then extract more discriminative pedestrian matching features.In addition,the classification loss and triplet loss are used to jointly optimize the model to improve the robustness and generalization ability of the model.(3)Based on the above method,this paper designs a person re-identification system and realizes the visualization of the algorithm running results.Through the simple input of person images to be detected,the system intuitively feeds back the results of person reidentification,which verifies the effectiveness of the method in this paper and meets the needs of person re-identification in practical application scenarios. |
