基于显著特征的行人重识别方法研究

发布时间：2018-04-29 11:49

本文选题：行人重识别 + 超像素　；参考：《南京邮电大学》2017年硕士论文

【摘要】：在视频监控领域,视频监看人员需要跨越多个摄像头对人群进行分析,从而识别出某个特定的已经出现过的人。研究人员将这种在多个监控视频环境下行人目标检索问题称为行人重识别问题。由于监控视频中的行人受视角变化、姿态变化、光照变化等因素影响,常常导致行人在不同摄像头下外貌发生较为明显的变化。围绕这些问题,本文研究了一种基于显著特征的行人重识别方法。在特征提取阶段,本文研究了一种基于超像素的特征表示方法。该方法首先将行人图像分割成若干超像素块,在此基础上结合颜色直方图和加速鲁棒特征构造特征空间。实验证明通过将该特征描述子与现有的块匹配方法相结合,极大地提高了行人重识别问题的运行效率与算法精度。传统的基于显著特征重识别方法通过不同样本的差异表征自身权重。然而这种计算结果不够稳定,它可能随着比较样本的变化而变化。因此本文介绍了一种基于元胞自动机的方法计算行人图像的内在显著特征。为了充分利用上述两种方法的优点,本文利用多层元胞自动机将它们融合从而在实验中获得了更好的效果。最后在相似性度量过程中仅仅根据各子块的显著与否决定匹配权重过于简单,本文在此基础上研究了一种学习排序方法衡量各图像之间的相似性。实验结果表明,与现有的算法相比在i LIDS数据库上本文算法表现出了更好的性能。
[Abstract]:In the field of video surveillance, video monitors need to analyze people across multiple cameras to identify a particular person who has already appeared. The problem of pedestrian target retrieval in multiple surveillance video environments is called pedestrian recognition problem by researchers. Due to the change of visual angle, posture, illumination and other factors in the video, the appearance of the pedestrian changes obviously under different cameras. To solve these problems, a pedestrian recognition method based on salient features is studied in this paper. In the phase of feature extraction, a feature representation method based on hyperpixel is studied in this paper. Firstly, the pedestrian image is divided into several super-pixel blocks, and then the color histogram and the accelerated robust feature are combined to construct the feature space. The experimental results show that by combining the feature descriptor with the existing block matching methods, the running efficiency and the algorithm accuracy of the pedestrian recognition problem are greatly improved. Traditional recognition methods based on salient features represent their weight by different samples. However, the results of this calculation are not stable and may vary with the variation of the comparison samples. Therefore, this paper introduces a cellular automaton based method for calculating the inherent salient features of pedestrian images. In order to make full use of the advantages of the above two methods, this paper uses multilayer cellular automata to fuse them and obtain better results in experiments. Finally, in the process of similarity measurement, it is too simple to determine the matching weight based on the salience or not of each sub-block. On the basis of this, we study a learning sorting method to measure the similarity between images. Experimental results show that the proposed algorithm performs better on I LIDS database than the existing algorithms.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【参考文献】