基于姿态对齐的行人重识别方法（英文）

发布时间：2018-05-20 09:50

本文选题：行人重识别 + 姿态对齐　；参考：《控制理论与应用》2017年06期

【摘要】：行人重识别是指根据输入的某个行人图片,在视频监控网络中对该行人目标进行检索.行人的姿态变化和监控场景的亮度变化是该任务的两个主要挑战.针对行人的姿态变化问题,本文首先对训练集中行人图片进行稠密图像块采样获得图像块集合,然后对每一个图像块提取其局部表观空间特征,最后在此特征集上聚类得到通用的行人部件字典.由于该部件字典编码了行人的部件信息,因此通过该字典内的每一个码元可以建立两幅行人图像中特定图像块之间的对应关系.将两幅行人图片的图像块集合分别向部件字典投影,可以获得2幅行人图片姿态对齐后的图像块序列.针对监控场景的亮度变化问题,本文在姿态对齐后的图像块上分别提取4种颜色描述子,并将不同颜色描述子下的图像块相似性进行分数级组合以获得更好的亮度不变性.其中不同颜色描述子之间的组合系数通过结构化输出支持向量机学习得到.在常用的视点不变行人重识别(viewpoint invariant pedestrian recognition,VIPe R)数据集上的实验结果表明,该方法在存在行人姿态变化和场景亮度变化干扰时获得了较好的行人重识别效果.
[Abstract]:Pedestrian recognition is to retrieve the pedestrian target in video surveillance network according to the input of a pedestrian image. The change of pedestrian posture and the brightness change of monitoring scene are the two main challenges of this task. In order to solve the problem of pedestrian attitude change, this paper firstly collects the dense image blocks of pedestrian images, and then extracts the local apparent spatial features of each image block. Finally, a general pedestrian component dictionary is obtained by clustering on this feature set. Because the component dictionary encodes the component information of pedestrians, it is possible to establish the corresponding relationship between the specific image blocks in two pedestrian images by each symbol in the dictionary. By projecting the image blocks of two pedestrian images to the component dictionary, the image block sequences of the two pedestrian images can be obtained after the pose alignment of the two pedestrian images is obtained. To monitor the luminance change of the scene, four color descriptors are extracted from the image blocks after pose alignment, and the similarity of the image blocks under different color descriptors is combined in a fractional order to obtain better brightness invariance. The combination coefficients between different color descriptors are obtained by structured output support vector machine learning. The experimental results on the viewpoint invariant pedestrian recognition VIPe R data set show that the proposed method can achieve a better result in the presence of pedestrian posture changes and scene brightness changes.
【作者单位】：华中科技大学自动化学院图像信息处理与智能控制教育部重点实验室;
【分类号】：TP391.41

【相似文献】