基于视频预训练和注意力特征融合的行人重识别方法

Person re-identification based on video pre-training and attentional feature fusion

行人重识别是跨摄像头追踪的关键环节之一,主流方法多采用ImageNet进行预训练,忽视了数据集的域间差异,且以结构庞大的多分支模型居多,模型复杂度较高。本文设计一种行人重识别方法,采用基于原始视频带噪声标签参与监督的方式进行预训练,减少域间差异以提升特征表达能力;以基于注意力的特征融合方式取代残差网络的跳接映射,增强网络的特征提取能力;在网络中嵌入坐标注意力机制,在低复杂度的情况下强化关键特征,抑制低贡献特征;采用随机擦除对输入数据做数据增强以提高泛化能力,联合分类损失、三元组损失和中心损失函数对网络进行监督训练。在公开数据集Market-1501和Duke-MTMC上完成了消融实验,与主流方法对比实验表明本方法在不需要复杂多分支逻辑结构的前提下,仍可达到较高的精度。

Person re-identification is one of the key steps in cross camera tracking.Most mainstream methods use ImageNet for pre training,ignoring the difference between domains of data sets,and most of them are multi branch models with large structures,which have high complexity.In this paper,a pedestrian re recognition method is designed,which adopts the method of pre training based on the original video with noisy tags to participate in the supervision,and reduces the difference between domains to improve the feature expression ability;The attention based feature fusion method is used to replace the jump mapping of the residual network,which enhances the feature extraction ability of the network;Embed coordinate attention mechanism in the network to strengthen key features and suppress low contribution features in the case of low complexity;At the same time,random erasure is used to enhance the input data to improve the generalization ability.The network is supervised by combining classification loss,triple loss and central loss functions.Ablation experiments have been completed on Market-1501 and Duke MTMC public datasets.The comparison experiments with mainstream methods show that this method can still achieve high accuracy without complex multi branch logic structure.