无监督学习三元组用于视频行人重识别研究  

Unsupervised learning triplets for video-based pedestrian reidentification

作　　者：蔡江琳 韩华[1] 王春媛 潘欣宇 芮行江 CAI Jianglin;HAN Hua;WANG Chunyuan;PAN Xinyu;RUI Xingjiang(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《智能计算机与应用》2022年第11期18-25,共8页Intelligent Computer and Applications

基　　金：国家自然科学基金(61305014);上海市自然科学基金(22ZR1426200);上海市教育委员会和上海市教育发展基金会“晨光计划”(13CG60)。

摘　　要：在智能交通中,对于目前产生的海量视频通过人工来标记行人图像不切实际,使无监督学习得到更多的关注。针对在无监督学习数据中缺少详细的身份信息,无法知晓目标图像对应的正负样本问题,提出一种无监督学习三元组用于视频行人重识别研究的方法。该方法从无标签的数据集中挖掘三元组、即目标图像,与目标图像身份相同的轨迹和与目标图像身份不同的轨迹。首先根据单相机内轨迹的时空一致性,即构成轨迹的任意帧图像具有相同的身份,将行人轨迹特征表示成图像特征均值后,通过计算rank-1轨迹作为判断三元组的条件,用于设计特殊的三元组损失函数。并根据特征距离大小分配样本权重,着重学习困难样本,使模型动态调整正、负样本对之间的距离,加速模型的收敛速率,降低过拟合风险。然后通过计算跨相机rank-1,合并高度关联的轨迹作为跨相机三元组的锚样本用于损失计算。最后联合单相机和跨相机的损失评估模型。经过实验证明,该方法在PRID2011、iLIDS-VID和MARS上的结果都表明了该模型的有效性和可靠性。In view of the massive videos of the intelligent transportation system,it is impractical to manually label pedestrian images,making unsupervised learning get more attention.Aiming at the lack of detailed identity information in the unsupervised learning data and the inability to know the positive and negative samples corresponding to the target image,an unsupervised learning triplet method is proposed for video pedestrian re-identification research.From an unlabeled dataset,the method mines triples,namely target images,trajectories with the same identity as the target image and trajectories with different identities from the target image.First,according to the spatio-temporal consistency of the trajectory within a single camera,that is,any frame images that constitute the trajectory have the same identity.After the pedestrian trajectory feature is expressed as the feature mean of the image,the rank-1 trajectory is calculated as the condition for judging the triplet and is used for designing a special triplet loss function.Based on this,the paper assigns the sample weight according to the feature distance,focuses on learning difficult samples,makes the model dynamically adjust the distance between positive and negative sample pairs,accelerates the convergence rate of the model,and reduces the risk of overfitting.Then,by computing the cross-camera rank-1,the highly correlated trajectories are merged as anchor samples for the cross-camera triples for loss computation.Finally,joint single-camera and cross-camera losses are evaluated.Experiments show that the results of this method on PRID2011,iLIDS-VID and MARS all demonstrate the validity and reliability of the model.

关 键 词：无监督学习 行人轨迹 关联排序 时空一致性 三元组损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]