显著关注区域局部与整体融合的行人再识别方法研究

Research on Pedestrian Re-recognition Method with Significant Focus on Regional Local and Global Integration

在行人再识别中,传统的行人再识别方法在进行行人再识别时,会受到背景、面纱、服装等变化的影响使识别效果下降。为了减少背景、面纱、服装等变化对识别效果的影响,论文提出了基于显著关注区域的局部与整体融合的行人再识别方法,该方法通过比较两个行人的度量距离完成行人再识别。首先,该方法包含整体特征提取和局部显著关注特征提取两个支路,然后将整体特征和局部显著关注特征融合得到融合特征。并用融合特征进行对比度量学习,并计算出相似度分数对样本进行排序。其次,论文改进了三元组损失并用于基于显著关注区域的局部与整体融合的模型。在大型数据集CUHK03和VIPER上的大量实验结果表明,该方法减少了背景、面纱、服装等变化对识别效果的影响。

In pedestrian re-identification,the traditional pedestrian re-identification method will be affected by the background,veil,clothing and other changes to make the recognition effect decline.In order to reduce the influence of background,veil,clothing and other changes on the pedestrian re-identification effect,pedestrian re-identification method is proposed.Based on local attention area feature and global feature fusion for pedestrian re-identification,which achieves the purpose of pedestrian re-identification by comparing the measured distance between two pedestrians.Firstly,this method consists of two branches,which are global feature extraction and local attention area feature extraction,then the integrated features are obtained by fusion of the global features and the local attention area features.The integrated features are used for comparative measurement learning and the similarity score was calculated to discriminate the samples.Secondly,in order to make the method have better performance,the triple loss is improved and used to train the model of this method.Extensive experimental results on the large-scale,CUHK03 and VIPeR data sets demonstrate that the proposed method reduces the influence of background,veil and clothing on the recognition effect.