| With the rapid growth of intelligent monitoring technology,person reidentification has become an area which gets a lot of attention.person re-identification means a technology which accurate cross-camera recognition and matching of the same pedestrian in multiple different scenes or times.In the field of intelligent monitoring,person re-identification technology can assist in completing labor-intensive tasks such as floating population survey,tracking suspicious persons,and searching for missing persons.In real-life scenarios,the implementation of person re-identification technology faces many challenges due to factors such as lighting,viewing angle,and occlusion in different scenarios.In addition to data engineering work,many researchers have devoted themselves to researching more reliable and effective algorithms.Currently,domestic and foreign researchers have made many useful explorations and attempts in person re-identification algorithms,including methods based on global features,local features,and a combination of global and local features.However,most methods have complex network structures and concatenate multiple branch features,which significantly reduce the retrieval speed during the inference stage and are difficult to meet the stringent requirements for real-time performance in actual scenarios.This article is based on a simple structured benchmark network model—BNNeck,and improves the structure of its loss function without increasing the complexity of the model parameters to ensure the speed of model inference.The main research contents are as follows:(1)Through research,it was found that the traditional triplet loss function focuses on the relative size of positive and negative sample pairs,while ignoring the absolute distance between positive and negative sample pairs.Additionally,the distance between positive and negative samples is not introduced as a parameter,which generally has a mediocre effect on the recognition of difficult samples.This article proposes a new triplet loss function inspired,which views triplet samples as a triangular structure,and redesigns the loss function based on this.The new loss function outperforms the benchmark model on multiple datasets,effectively improving the recognition rate of difficult samples.(2)The loss function structure of the benchmark model includes cross-entropy loss,triplet loss,and center loss.The center loss function was introduced to compensate for the problems with the triplet loss,However,the above improvements to triplet loss have effectively alleviated this issue,and to avoid duplicate calculations,it has been removed.The results of the experimental showed that after removing the center loss,the performance of the indicator did not fluctuate significantly.At the same time,due to the randomness of center selection,it had a significant impact on the results,and the removal effectively enhanced the robustness of the model.(3)A desktop system for person re-identification has been developed based on the Py Qt framework,the newly designed model in this article was integrated into the pe re-identification system and achieved good results in different scenarios. |
