Person Re-identification Using Deformable Part Models
Person Re-identification Using Deformable Part Models
复制标题
DOI：
10.1007/978-3-642-42051-1_76
复制
发表时间：
2013-11
期刊：
影响因子：
--
通讯作者：
Vu-Hoang Nguyen;Kien Nguyen;Duy-Dinh Le;D. Duong;S. Satoh
中科院分区：
文献类型：
--
作者：
Vu-Hoang Nguyen;Kien Nguyen;Duy-Dinh Le;D. Duong;S. Satoh
人员重新识别是在非重叠摄像机网络中匹配人员的问题。挑战之一是如何将身体部分与身体部分匹配，以用于在不同视点以及可变形人体的背景下的两个人的图像之间的比较。现有的方法通常使用固定的模型来定位身体部位或检测人体形状以从形状中提取身体部位。因此，很难改变到新的模型或身体部位的结构。此外，这些方法不能同时处理多个人体姿势。我们提出了一种基于机器学习的方法来提取身体部位，这是基于可变形的部分模型（DEFORM）。它很容易训练，并且具有强大的性能。此外，使用mathematics，我们可以同时为多个人体姿势使用多个模型。在标准数据集ETHZ1上的实验表明，该方法优于现有的方法。
Person Re-Identification is the problem of matching people across a network of non-overlapping cameras. One of the challenges is how to match body parts to body parts for comparison between images of two people in the context of different viewpoints as well as deformable human bodies. Existing approaches usually use fixed models to localize body parts or detect human shapes to extract body parts from the shapes. Therefore, it is difficult to change to a new model or structure of body parts. Moreover, those approaches could not deal with multiple human poses simultaneously. We propose a machine learning-based method to extract body parts that is based on Deformable Part Models (DPM). DPM is easy to train and has robust performance. In addition, with DPM, we could use multiple models for multiple human poses concurrently. Experiments on standard dataset ETHZ1 show that the proposed method outperforms state of the art methods.