光学学报, 2019, 39 (6): 0615007, 网络出版: 2019-06-17
基于视角信息嵌入的行人重识别 下载: 1263次
Person Re-Identification Based on View Information Embedding
机器视觉 光计算 行人重识别 视角信息嵌入 深度残差卷积神经网络 深度可分离卷积 machine vision optics in computing person re-identification perspective information embedding deep residual convolution neural network depthwise separable convolution
摘要
提出一种基于视角信息嵌入的行人重识别模型。结合行人图像视角朝向特点,对PSE (pose-sensitive embedding)网络结构进行了优化。首先将PSE特征向量融合部分由特征的融合改成更符合不同视角特征空间性质的三个视角单元特征向量的拼接;其次视角单元从骨架网络更浅层的blocks-3进行分离,增加三个视角单元特征空间的差异性;最后利用改进的深度可分离卷积,设计了一个深度可分离模块,对视角单元进一步进行提取特征,防止模型参数过大的同时提高网络非线性能力,从而提高网络的泛化能力。利用Market1501、Duke-MTMC-reID和MARS数据集对所提的算法进行有效性验证实验,结果表明所提的改进方法取得了更好的识别效果。
Abstract
In this study, we propose a person re-identification model based on view information embedding. In particular, a pose-sensitive embedding (PSE) network structure is optimized based on the perspective towards characteristics of pedestrian images. First, the fusion part of the PSE feature vector is changed from feature fusion into the concatenation of the feature vectors of three view units, which is considerably reasonable for utilizing different view feature spaces. Second, the view units are separated from the shallow blocks-3 of the skeleton network, which improves the difference of the view feature space. Finally, we design a depthwise separable module based on the improved depth separable convolution to extract features of perspective units, preventing the model parameters from being considerably large and improving the network nonlinearity. The results of the validation experiments conducted using the Market1501, Duke-MTMC-reID and MARS datasets demonstrate that the proposed method can achieve a better recognition accuracy when compared with several advanced algorithms.
毕晓君, 汪灏. 基于视角信息嵌入的行人重识别[J]. 光学学报, 2019, 39(6): 0615007. Xiaojun Bi, Hao Wang. Person Re-Identification Based on View Information Embedding[J]. Acta Optica Sinica, 2019, 39(6): 0615007.