基于外观统计特征融合的人体目标再识别

Fusing Appearance Statistical Features for Person Re-identification

人体目标再识别是视频监控等应用的关键问题之一。该文从外观统计特征融合的角度,利用人体的颜色和结构信息,基于空间直方图和区域协方差两种优秀的统计描述方法,研究了再识别问题的特征构建和测度选择等内容。构建特征时从图像多个层次的统计区域中提取了多类互补性较好的统计向量,设计测度时使用了简单的1l距离进行加权组合。两类统计方式融合而成的再识别方法不需要进行预处理和监督性训练过程。该文进行了广泛的实验比较和分析,验证了该文方法优异的识别性能和较强的实用性能。

Person re-identification is among the key issues in video surveillance. From the viewpoint of fusing appearance statistical features, human color and structure information are exploited; two statistical descriptors named spatiogram and region covariance are both explored on feature designing and metric choosing. Several complimentary feature vectors are extracted from a proper number of hierarchical image layers and regions. The simplest 1l norm distance is chosen to form the proposed weighted combining distance. The fused method with such two descriptors requires neither preprocessing nor supervised training. Extensive experiments by comparisons and analysis show that the proposed method not only achieves the state-of-the-art re-identification performance, but also enjoys a great applicability.