计算机应用 ›› 2018, Vol. 38 ›› Issue (5): 1299-1303.DOI: 10.11772/j.issn.1001-9081.2017102581

• 人工智能 • 上一篇    下一篇

复杂环境中基于视觉词袋模型的车辆再识别算法

王茜1,2, 陈一民1, 丁友东3   

  1. 1. 上海大学 计算机工程与科学学院, 上海 200072;
    2. 上海市公安局刑事侦查总队 科技信息科, 上海 200083;
    3. 上海大学 影视学院, 上海 200072
  • 收稿日期:2017-10-31 修回日期:2017-12-12 出版日期:2018-05-10 发布日期:2018-05-24
  • 通讯作者: 王茜
  • 作者简介:王茜(1981-),女,上海人,高级工程师,博士研究生,主要研究方向:视频图像智能分析、模式识别、深度学习网络、生物特征识别;陈一民(1961-),男,上海人,教授,博士,CCF高级会员,主要研究方向:多媒体、计算机网络、计算机控制、机器人控制;丁友东(1967-),男,上海人,教授,博士,CCF高级会员,主要研究方向:计算机图形学、数字视觉媒体、电影大数据处理。

Vehicle re-identification algorithm based on bag of visual words in complicated environments

WANG Qian1,2, CHEN Yimin1, DING Youdong3   

  1. 1. College of Computer Engineering and Science, Shanghai University, Shanghai 200072, China;
    2. Information Center, Criminal Investigation Department of Shanghai Public Security Bureau, Shanghai 200083, China;
    3. Shanghai Film Academy, Shanghai University, Shanghai 200072, China
  • Received:2017-10-31 Revised:2017-12-12 Online:2018-05-10 Published:2018-05-24
  • Contact: 王茜

摘要: 根据公共安全部门在复杂环境中搜索出特定目标的迫切需求,将目标再识别(re-ID)技术应用到车辆识别领域,提出了一种基于视觉词袋(BoVW)模型的车辆再识别解决方案。首先,为解决复杂环境中遮挡、目标物位姿变化、目标物在图片中的大小位置存在差异等问题,提取出可基于不同尺度、不同位姿的改进基于部件的一对一局部特征(POOF);其次,通过基于欧氏距离的聚类算法获取视觉词袋中的词汇集合;接着,将训练和测试集中的每张图像或目标转换为词袋中的词汇表述集;最后,利用基于改进保持直接简单原则的度量方法(KISSME)上的再排序方法分离出类间距离和类内距离,通过最近邻方法(KNN)获得识别结果。实验结果显示,在基础特征构建环节上该算法比冒泡银行算法(BB)识别率提升了3.85个百分点;其基于KISSME距离度量的改进再排序算法比贝叶斯再访问算法提升了3.14个百分点。最后,算法对目标位姿变化和局部遮挡具有的适应性和整体时效指标,进一步验证了其可适应于复杂环境应用的特色和优越性。

关键词: 车辆再识别, 视觉词袋模型, 一对一局部特征, 距离度量, 再排序

Abstract: To meet the demands of the public security department to search out specific target in complicated real environment, the target re-IDentification (re-ID) technique was introduced in vehicle identification field, and a vehicle re-ID algorithm based on the model of Bag of Visual Words (BoVW) was proposed. Firstly, in order to solve the probloms of occlusion pose change, target size and position difference in images, the improved scales and poses adaptive Part-based One-vs-One Feature (POOF) were extracted. Secondly, a set of visual words was clustered as a vocabulary by using k-means algorithm based on Euclidean distance, and the features of each image (or target) were expressed as the composition of visual vocabularies. Thirdly, the improved Keep It Simple and Straightforward Metric (KISSME) method followed with re-rank step was used to separate the between-classes and within-classes distances. Finally, the result was obtained by using K-Nearest Neighbor (KNN) method. The experimental results show that the algorithm has 3.85 percentage points increasement of identification rate in feature representation step compared with Bubble Bank (BB) and 3.14 percentage points increase in metric learning step compared with Bayesian face revisited. Furthermore, it is proved that the proposed algorithm is economical in time-consuming and has strong adaptability to target pose change and small portion of occlusion, which further domonstrates that it can adapt to complicated environments.

Key words: vehicle re-identification, Bag of Visual Words (BoVW), Part-based One-vs-One Feature (POOF), distance metric, re-rank

中图分类号: