激光与光电子学进展, 2020, 57 (8): 081503, 网络出版: 2020-04-03
基于多层级特征的行人重识别 下载: 1158次
Person Re-Identification Based on Multi-Layer Feature
机器视觉 行人重识别 残差网络 多层级特征 相似性度量 machine vision person re-identification residual network multi-layer feature similarity measurement
摘要
针对现有的行人重识别方法提取行人特征过程中存在因信息缺失导致鲁棒性和判别力较差的问题,提出了一种基于残差神经网络提取行人图像多层级特征的方法。该方法在训练阶段使用残差网络分别在4个卷积残差模块之后提取阶段特征,以此来弥补信息丢失,使用三元组损失函数对每个特征向量进行监督训练。在相似性度量阶段,针对4个特征向量分别计算特征相似度,使用映射函数进行求和,并对求和结果进行相似度匹配。将该方法在Market-1501和DukeMTMC-ReID数据集上进行仿真,首中准确率(Rank-1)分别达到了91.7%和84.9%,平均准确率(mAP)分别达到了86.8%和80.7%。结果表明所提方法提取的多层级特征具有较好的鲁棒性和判别力,提高了行人重识别的准确度。
Abstract
To address the issue that existing person re-identification (Re-ID) algorithms have low robustness and discriminative capability when extracting pedestrian features with information loss, a novel Re-ID algorithm based on residual neural network is proposed for extracting multi-layer features of pedestrian images. During training phase, the residual network is used to extract the phase features after the four convolutional residual modules, to compensate for the information loss. And the triple loss function is used to supervise training of each feature vector. During the similarity measurement phase, the feature similarity is calculated according to the four feature vectors, the similarity of each stage is calculated by the summation of mapping function, and then the result of the summation is used to perform similarity matching. During the experiment, we validate the proposed algorithm on the Market-1501 and DukeMTMC-ReID datasets. The accuracy (Rank-1) of our algorithm reaches 91.7% and 84.9% and mean average precision (mAP) reaches 86.8% and 80.7%, respectively. Experimental results show that the multi-layer features extracted by our algorithm have considerable robustness and discriminative capability, which improves the accuracy of Re-ID.
刘可文, 房攀攀, 熊红霞, 刘朝阳, 马圆, 李小军, 陈亚雷. 基于多层级特征的行人重识别[J]. 激光与光电子学进展, 2020, 57(8): 081503. Kewen Liu, Panpan Fang, Hongxia Xiong, Chaoyang Liu, Yuan Ma, Xiaojun Li, Yalei Chen. Person Re-Identification Based on Multi-Layer Feature[J]. Laser & Optoelectronics Progress, 2020, 57(8): 081503.