基于多层级特征融合的行人重识别算法

针对行人遮挡、姿态变化等现象造成当前行人重 识别算法精度不高的问题,提出一 种基于多层级特征融合的行人重识别算法。首先通过自注意力机制骨干网络ResNeSt提取图 像特征中的短距离信息；其次通过多尺度 金字塔卷积(pyramid convolution,Pyconv) 分支网络提取图像中长像素关 联特征信息,提高模型表达能力；最后使用一种统一形式且可学习的广义均值池化 (generalized mean pooling,GEM) 替代传统平均池化层,达到关注不同区域特征差异性目的。测试阶段添加平均逆消极惩罚 (mINP)作为新评价指标。实验结果表明,本文所提算法在多个数据集上均展现出优势,在 DukeMTMC-ReID数据集上Rank-1达到了90.9%,mAP达到了89.8%。

Pedestrian re-identification algorithm based on multi-level feature fusion

In order to solve the problem of low accuracy of current pedestrian re-identification algorithms caused by phenomena such as pedestrian occlusion a nd pose changes,a pedestrian re-identification algorithm based on multi-level feature fusion is proposed.Firstly,the short-distance information in the image features is extr acted by the self-attention mechanism backbone network ResNeSt;secondly,the multi-scale pyramid convolution (pyconv) branch network is used to extract the associated feature in formation of the long pixels in the image to improve the expression ability of the model;fin ally,a unified form and the learnable generalized mean pooling (GEM) is used to replace the tra ditional average pooling layer for the purpose of focusing on the variability of differen t regional characteristics.The average inverse negative penalty (mINP) is added as a new e valuation metric in the testing phase.Experimental results show that the proposed algorit hm exhibits advantages on multiple datasets,with Rank-1reaching 90.9% and mAP reaching 89.8% on the DukeMTMC-ReID dataset.