欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 图像与视频处理 • 上一篇    下一篇

结合 Faster RCNN 和相似性度量的行人目标检测

  

  1. 1. 中国石油大学(华东)计算机与通信工程学院，山东 青岛 266580； 
    2. 中国科学院大学中科院计算所，北京 100190
  • 出版日期:2018-10-31 发布日期:2018-11-16
  • 基金资助:
    国家自然科学基金项目(61379106，61379082，61227802)；山东省自然科学基金项目(ZR2013FM036，ZR2015FM011)

Pedestrian Object Detection Based on Faster RCNN and  Similarity Measurement

  1. 1. College of Computer and Communication Engineering, China University of Petroleum, Qingdao Shandong 266580, China; 
    2. University of Chinese Academy of Sciences, Institute of Computing Technology Chinese Academy of Sciences, Beijing 100190, China
  • Online:2018-10-31 Published:2018-11-16

PDF (PC)

322

摘要： 行人检测是计算机视觉领域的一个研究热点，针对目前算法中常采用非极大值抑 制和硬阈值筛选的方法作为后处理，容易造成误检和漏检的问题，提出一种基于相似性度量的 行人目标检测方法。首先，采用 Faster RCNN 生成一系列的目标候选集，应用非极大值抑制对 候选集进行初步筛选，然后由较高置信度的目标区域建立特征模板，再根据特征相似性对较低 置信度的目标区域做进一步判别，最后将筛选后的目标候选集和模板区域作为检测结果。在 VOC、INRIA、Caltech 数据集的实验结果证明，基于相似性度量的算法提高了行人检测的准确率。
 

关键词: 行人检测, 目标候选集筛选, 特征相似性度量, 模板匹配

Abstract: Pedestrian detection has become a hot topic in the field of computer vision. Non-maximal suppression combined with hard threshold is the most common post-process method in pedestrian detection, whereas it is easy to cause false positive and false negative. As to this problem, this paper presents a pedestrian-object detection method based on similarity measurement. Firstly, Faster RCNN is used to build a series of candidate proposals among which initial selection is made based on non-maximal suppression. Then the authors create feature templates by target areas with high confidence, and make a further selection in the low-confidence proposals according to the feature similarity. Lastly, the detection results are composed of the reserved proposals and the templates. The experimental results from VOC, INRIA, Caltech datasets demonstrate that similarity measurement method can achieve higher pedestrian detection performance.

Key words: pedestrian detection, object proposals selecting, feature similarity measurement, template matching