Multi-object Tracking Based on Discriminative Correlation Fi

发布时间：2021-03-28 16:43

　　在计算机视觉领域,多目标跟踪（MOT）一直是困难且具有挑战性的研究。在该领域中,计算数据量很大,并且算法的处理速度要求很高,因此跟踪速度、效率和精度常常难以很好地平衡。由于这个原因,将单目标跟踪中使用的具有通道和空间可靠性的判别相关滤波器（DCF-CSR）用于MOT,以在保持跟踪速度的同时确保跟踪效果。本文分为以下四个部分:（1）目标检测。由于2D MOT 2015和MOT16数据集目标的检测存在错漏,提出将基于更快区域的卷积神经网络（Faster R-CNN）用于目标检测,并且将数据集提供的结果替换为网络检测的结果。（2）目标跟踪。为了在MOT中添加目标外观数据,并结合速度和准确性因素,本文使用DCF模型。但是,已经发现执行多目标跟踪数据集时跟踪结果并不理想,因为DCF算法不支持多尺度。为克服此困难,本文解决了计算检测目标框和预测目标框之间的交并比（IOU）并使用检测目标框代替预测的方法。₃数据关联。对于第（2）部分的IOU的计算,这是一批检测框和一组预测框之间的相互计算问题,本质上是任务分配和优化问题。为此,匈牙利算法被用于确定目标之间的最佳相关性。（4）对象... 

【文章来源】：大连海事大学辽宁省 211工程院校

【文章页数】：77 页

【学位级别】：硕士

【文章目录】：
摘要
Abstract
1 Introduction
    1.1 Background and significance of the topic
    1.2 Research Status of target tracking in China and Abroad
        1.2.1 Research Status of Single-target tracking with correlation filtering
        1.2.2 Research Status of Multi-target tracking with correlation filtering
    1.3 Research content and chapter arrangement
        1.3.1 Research content
        1.3.2 Chapter Arrangement
2 Multi-target tracking pre-detection
    2.1 Introduction of Multi-target tracking data sets
        2.1.1 2D MOT 2015 dataset
        2.1.2 MOT16 Data Set
        2.1.3 Comparison of differences between data sets
        2.1.4 Inaccurate target detection
    2.2 Target detection models
        2.2.1 Traditional models
    2.3 Evaluation of target detection model
        2.3.1 Target detection and evaluation indicators
        2.3.2 Comparative analysis of test results
    2.4 Summary of this chapter
3 Discriminative Correlation Filter Tracker with Channel and Spatial Reliability
    3.1 Spatially constrained correlation filters
    3.2 Constructing spatial reliability map
    3.3 Channel detection reliability
    3.4. Channel reliability estimation
    3.5. Tracking with channel and spatial reliability
    3.6. Spatial and channel reliability ablation study
        3.6.1 The OTB 100 benchmark
        3.6.2. The VOT2015 benchmark
        3.6.3. The VOT2016 benchmark
    3.7 Inadequacies
    3.8 Summary of this chapter
4 Core-related filter tracking incorporating data association strategies
    4.1 Target transition status process
    4.2 Target Association Strategy
        4.2.1 Calculating the correlation matrix
        4.2.2 Determining the optimal association
    4.3 Target similarity judgment method
    4.4 Algorithm Flow
        4.4.1 Troubleshooting
        4.4.2 Outline of proposed algorithm
    4.5 Summary of this chapter
5 Multi-target tracking experiments
    5.1 Experimental Environment
    5.2 Experimental analysis
        5.2.1. Implementation details and parameters
        5.2.2 Non-axis-aligned target initialization robustness
    5.3 Evaluation indicators
    5.4 Summary of this chapter
6 Summary and Outlook
    6.1 Summary of work in this thesis
    6.2 Future work prospects
References
DECLARATION
ACKNOWLEDGEMENT



本文编号：3105900