SIA OpenIR  > 机器人学研究室
基于深度学习的多行人目标跟踪及行人重识别算法研究
Alternative TitleDeep Learning based Multi-person Tracking and Person Re-identification
刘金文
Department机器人学研究室
Thesis Advisor田建东
Keyword多行人目标跟踪 人群密度图 行人重识别 注意力机制 深度学习
Pages63页
Degree Discipline模式识别与智能系统
Degree Name硕士
2021-05-21
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract多行人目标跟踪及行人重识别作为计算机视觉领域的两个重要分支,有着广阔的应用前景和重要的应用价值,并且二者都是以人为研究对象,面临的挑战和难题有一定的相似之处。近年来,随着深度学习的发展,多行人目标跟踪及行人重识别研究取得了一定的进展,但是,行人姿态多变以及行人遮挡等问题仍然是目前所面临的技术难题和挑战。针对上述问题,本文在多行人目标跟踪研究中提出了一种融合人群密度的自适应深度多目标跟踪算法;在行人重识别研究中提出了一种融合注意力机制的行人重识别网络模型。本学位论文的主要研究内容和创新点如下:(1)提出了一种新的多目标跟踪算法。本文算法首先对人群密度图和目标检测结果进行融合,利用人群密度图的位置和计数信息对检测器输出结果进行修正,有效地消除了漏检和误检,可得到更加精确的检测结果;然后使用自适应Triplet Loss改进行人重识别模型的损失函数,提高了重识别特征的辨别能力,提升了跟踪器的数据关联精度;最后,使用行人表观和运动信息进行数据关联,得到最终的跟踪结果。实验验证了本文所提出的多目标跟踪算法能够有效地解决目标遮挡场景下的多目标跟踪问题。(2)提出了一种新的行人重识别网络模型。本文模型融合了注意力机制以增强网络获得行人空间信息的能力,所设计的注意力模块包含了空间注意模块和通道注意模块,使得注意力模块可以筛选出值得关注的特征。在本文模型中,选择了ResNet-101作为行人重识别模型的主干网络。为了网络提取的特征包含丰富的空间信息和语义信息,本文在主干网络的不同输出阶段构建了分支结构。为了网络提取的行人特征更具判别性,本文使用ID分类损失和度量学习方法同时对网络进行训练。对于注意力模块的最佳融合方式,本文进行了分析并确定了最佳融合方式。实验结果表明本文提出的行人重识别模型可以提取判别性行人特征,具有良好的性能。
Other AbstractMulti-person object tracking and person re-identification, as two important branches of computer vision, have broad application prospects and value. They both regard person as research objects, and the challenges and problems they face have certain similarities. In recent years, with the development of deep learning, the research of multi-person tracking and person re-identification has made good progress. However, the changeable person posture and the problem of person occlusion are still challenges. In response to the above-mentioned problems, this thesis proposes an adaptive deep multi-object tracking algorithm that integrates crowd density in the multi-person tracking research, and a person re-identification network model that integrates the attention mechanism in the person re-identification. The main innovations and research contents of this thesis are as follows: (1) A multi-object tracking algorithm is proposed in this thesis. It firstly fuses the crowd density maps and the object detection results, and uses the location and counting information of the crowd density maps to correct the output results of the detectors, effectively eliminating missed detections and false detections, and get more accurate detect results. Then this algorithm uses the adaptive Triplet Loss to improve the loss function of the re-identification model, which improves the discrimination ability of re-identification features, and then improves the accuracy of the data association of the tracker. Finally, in the algorithm, the person appearance and motion information are used for data association, and the final tracking result is obtained. The experiment results verify that the multi-object tracking algorithm proposed in this thesis can effectively deal with the problem of multi-object tracking in severely occluded scenes. (2) A person re-identification network model is proposed in this thesis.The model incorporates an attention mechanism to enhance the ability of the network to obtain person spatial information. The attention module designed in this model includes the spatial attention module and the channel attention module, so that the attention module can filter out the features worthy of attention. In this model, ResNet-101 is choosen as the backbone network of the person re-identification model. In order that the features extracted by the network contain rich spatial and semantic information, branch structures in different output stages of the backbone network are constructed. In order to make the person features extracted by the network more discriminative, ID classification loss and metric learning methods are used to train the network at the same time. For the best fusion of attention modules, analysis and experiments are made to explore and determine the best fusion method. The experimental results show that the person re-identification model proposed in this paper can extract discriminative person features and has good performance.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/28952
Collection机器人学研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
刘金文. 基于深度学习的多行人目标跟踪及行人重识别算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的多行人目标跟踪及行人重识别(4739KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[刘金文]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘金文]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘金文]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.