光学学报, 2019, 39 (9): 0915006, 网络出版: 2019-09-09
联合多级深度特征表示和有序加权距离融合的视频行人再识别方法 下载: 901次
Video-Based Person Re-Identification via Combined Multi-Level Deep Feature Representation and Ordered Weighted Distance Fusion
机器视觉 视频行人再识别 多级深度特征 距离融合 卷积神经网络 循环神经网络 有序加权 machine vision video-based person re-identification multi-level deep feature distance fusion convolutional neural network recurrent neural network ordered weighted
摘要
针对目前视频行人再识别中存在视角、光线变化,背景干扰与遮挡,行人外观与行为相似,以及相同行人在不同模态特征下距离的差异性而导致的匹配不正确问题,提出一种联合多级深度特征表示和有序加权距离融合的视频行人再识别方法。在行人特征表示阶段,提出了行人多级深度特征表示网络,该网络不仅能学习视频序列中行人的时空特征,还能获取行人的全局外观特征和局部外观特征。在有序加权距离融合阶段,将行人的特征表示输入到距离测度学习中,分别计算行人在三类特征下的独立距离,并将距离排序后,根据距离的排名优化距离权值,最后融合三类距离得到最终距离,从而准确匹配行人。通过在公共数据集中的实验表明,所提方法不仅能够提高视频行人再识别的识别率,还具有丰富和完整的行人特征表示能力。
Abstract
Video-based person re-identification problems are caused by perspective changes, lighting variations, background clutter, occlusion, appearance similarity, motion similarity, and mismatch resulting from the distance difference of same person with different modal features. This study proposes a video-based person re-identification method that combines multi-level deep feature representation and ordered weighted distance fusion. During the stage of person feature representation, the multi-level deep feature representation network proposed herein not only learns the space-time features of the persons in video sequences but also acquires the persons' global and local appearance features. In the stage of the ordered weighted distance fusion, the feature representations of persons are firstly input into distance metric learning, and the independent distances of persons under three types of features are calculated. The fusion algorithm then sorts the distances to optimize distance weights according to distance ranking. Finally, to accurately match a person, the algorithm fuses the three types of distances to obtain the final distance. Experimental results compared with the results of related methods in public datasets show that the proposed method not only improves the recognition rate of video-based person re-identification but also possesses abundant and integral ability for person feature representation.
孙锐, 黄启恒, 陆伟明, 高隽. 联合多级深度特征表示和有序加权距离融合的视频行人再识别方法[J]. 光学学报, 2019, 39(9): 0915006. Rui Sun, Qiheng Huang, Weiming Lu, Jun Gao. Video-Based Person Re-Identification via Combined Multi-Level Deep Feature Representation and Ordered Weighted Distance Fusion[J]. Acta Optica Sinica, 2019, 39(9): 0915006.