一种特征增强的Tri-CNN行人再识别方法

针对行人再识别中遮挡导致提取的高层特征分辨率低而影响识别率的问题，建立一种基于Tri-CNN的特征增强行人再识别方法。首先，对池化层提取的图像特征进行PCA降维，根据典型相关分析策略(CCA)融合特征，提取更具判别力的行人特征。其次，引入空间递归模型(SRM)对遮挡行人特征进行空间多向检测，提高对遮挡行人的识别率。最后，根据欧氏距离度量准则，分别验证正、负样本对间的距离，联合Softmax损失函数和Triplet损失函数优化网络模型，进而判别是否为同一行人。在MARS和ETHZ这2个数据集上进行实验，结果表明本文方法有效解决了一般遮挡识别问题，并显著提高了行人再识别精度。

A Feature-enhanced Tri-CNN Pedestrian Re-identification Method

Aiming at the problem of low resolution of extracted high level features and low recognition rate caused by occlusion in person re-identification, a feature enhanced pedestrian re-identification method based on Tri-CNN is established. Firstly, PCA dimensionality reduction is performed on the image features extracted from the pooling layer, and more discriminating pedestrian features are extracted according to CCA fusion features. Secondly, the spatial recursive model (SRM) is introduced to detect the features of occluded pedestrians in multiple directions, so as to improve the recognition rate of occluded pedestrians. Finally, according to the Euclidean distance measurement criterion, the distance between positive and negative sample pairs is verified respectively, and the loss function of Softmax and Triplet is combined to optimize the network model, so as to determine whether it is the same pedestrian. Experiments on MARS and ETHZ data sets show that the proposed method can effectively solve the problem of general occlusion recognition and significantly improve the accuracy of pedestrian re-identification.