基于视频的行人再识别

行人再识别是指在无交叉区域的多摄像机视频监控系统中,匹配不同摄像机中的相同行人目标。本文提出了一种基于视频的行人再识别方法,用HOG3D来描述一组视频的时空特征,在训练集上用预训练的DenseNet来微调模型参数,利用迁移学习得到的模型来提取视频中行人的表观特征,融合两种特征来描述视频序列中的行人。最后将融合的高维特征降维,并用度量学习方法计算行人对之间的距离。本文在PRID 2011和iLIDS-VID这两个视频数据集上进行了使用,实验结果表明本文的方法取得了较高的累积匹配得分。

Video-Based Person Re-identification

The task of person re-identification is to match pedestrian images observed from different cameras in a non-overlapping multi-camera surveillance systems. In this article, a video-based person re-identification method is proposed. HOG3D is extracted as temporal and spatial feature and the DenseNet model pre-trained on MSCOCO is adopted to fine-tune the parameters for person re-identification, and the fine-tuned model is used to extract the feature from person image. The two features are combined to describe the person video clip.Finally the metric learning model is applied to measure distance between person pairs. We evaluate our approach by operating in-depth experiments in two video-based benchmarks, and the experimental results on the two benchmark show significant and consistent improvements over the state-of-the-art methods.