基于跨纬度交互注意力机制的行人重识别方法

Research on person re-identification based on triplet attention mechanism

为解决由行人姿态、环境等复杂因素导致的行人特征表达能力弱、识别率低等问题,本文通过对AlignedReID++模型进行改进,提出了基于跨纬度交互注意力机制的行人重识别方法。首先,在特征提取部分,将跨纬度交互注意力Triplet Attention模块嵌入到ResNet50网络中,捕获空间维度和通道维度之间跨纬度的交互信息;其次,引入基于空间特性的视觉激活函数Funnel ReLU,解决激活函数的空间不敏感问题,增强网络模型的非线性表达能力;最后,在3个主流的行人重识别数据集Market1501、DukeMTMC-ReID、CUHK03上对改进模型进行效能评估,首位命中率Rank-1分别提高了1.6%、1.4%、2.8%,平均精度均值mAP分别提高了2.1%、2.3%、3.1%。结果表明所提算法具有良好的性能。

In order to solve the problem of weak person feature expression ability and low recognition rate caused by complex factors such as person attitude and environment,this paper proposes a person re-identificaton method based on cross-latitude interactive attention mechanism by improving AlignedReID++model.Firstly,in feature extraction,the Triplet Attention module is embedded in ResNet50 network to capture the cross-latitude interaction information between spatial dimensions and channel dimensions;Secondly,Funnel ReLU,a visual activation function based on spatial characteristics,was introduced to alleviate the spatial insensitivity of the activation function and enhance the nonlinear expression ability of the network model;Finally,the effectiveness of the improved model was evaluated on the three mainstream person re-identification datasets Market1501,DukeMTMC-ReID and CUHK03.The first shot Rank-1 increased by 1.6%,1.4%and 2.8%,the mean average precision mAP increased by 2.1%,2.3%and 3.1%,respectively.The results show that the proposed algorithm has good performance and can achieve higher recognition accuracy.