计算机工程 ›› 2020, Vol. 46 ›› Issue (3): 246-253.doi: 10.19678/j.issn.1000-3428.0054092

• 图形图像处理 • 上一篇    下一篇

基于姿态与并行化属性学习的行人再识别方法

陶飞, 成科扬, 张建明, 汤宇豪   

  1. 江苏大学 计算机科学与通信工程学院, 江苏 镇江 212013
  • 收稿日期:2019-03-04 修回日期:2019-04-12 发布日期:2019-05-22
  • 作者简介:陶飞(1993-),女,硕士研究生,主研方向为行人再识别、属性学习;成科扬,副教授、博士;张建明(通信作者),教授、博士;汤宇豪,硕士研究生。
  • 基金项目:
    国家自然科学基金"基于数据驱动与语义建模的多摄像拓扑推理与行人再识别研究"(61602215);江苏省科学基金"大数据驱动下基于深度学习与语义建模的行人再识别研究"(BK20150527);社会安全风险感知与防控大数据应用国家工程实验室主任基金"跨摄像头逻辑推理与多模态行人再识别技术研究"。

Pedestrian Reidentification Method Based on Posture and Parallel Attribute Learning

TAO Fei, CHENG Keyang, ZHANG Jianming, TANG Yuhao   

  1. School of Computer Science and Communication Engineering, Jiangsu University, Zhenjiang, Jiangsu 212013, China
  • Received:2019-03-04 Revised:2019-04-12 Published:2019-05-22

摘要: 行人再识别是当前图像识别领域的一个重要研究分支,在取得众多研究成果的同时,在实际场景中的应用也存在诸多挑战。摄像设备和拍摄场景的差异,以及穿着、尺度、部分遮挡、姿态等对行人外观的影响,给行人再识别带来较大的困难。为此,提出一种行人再识别方法,通过基于姿态的并行化属性学习任务对行人姿态信息进行标注,并将其作为语义属性融入到行人再识别任务中,降低实际场景中属性缺失对模型的影响,加速训练过程。实验结果表明,该方法在VIPeR数据集上达到了90%的识别率。

关键词: 深度学习, 行人再识别, 姿态, 属性学习, 并行化

Abstract: Pedestrian reidentification is an important research branch in the current image recognition field.While many breakthroughs are made,there are also many challenges in the application of actual scenes.The differences of camera equipment and shooting scenes,and the influence of wearing,scale,partial occlusion,posture,etc on pedestrian appearance,bring greater difficulties to pedestrians reidentification.Therefore,this paper proposes a pedestrian reidentification method.The method marks the pedestrian attribute information through the parallel attribute learning task based on posture,and then the labeled attitude information is integrated into the pedestrian reidentification task as a semantic attribute.The method reduces the impact of attribute missing on the model in the actual scenes and accelerates the training process.Experimental results show that the method achieves a 90% recognition accuracy on the VIPeR data set.

Key words: deep learning, pedestrian reidentification, posture, attribute learning, parallelization

中图分类号: