高级检索

    王潇潇, 张雪芹. 基于CNN-GRU度量网络的多目标跟踪算法[J]. 华东理工大学学报(自然科学版), 2021, 47(4): 483-493. DOI: 10.14135/j.cnki.1006-3080.20200331001
    引用本文: 王潇潇, 张雪芹. 基于CNN-GRU度量网络的多目标跟踪算法[J]. 华东理工大学学报(自然科学版), 2021, 47(4): 483-493. DOI: 10.14135/j.cnki.1006-3080.20200331001
    WANG Xiaoxiao, ZHANG Xueqin. Multi-Object Tracking Algorithm Based on CNN-GRU Metric Network[J]. Journal of East China University of Science and Technology, 2021, 47(4): 483-493. DOI: 10.14135/j.cnki.1006-3080.20200331001
    Citation: WANG Xiaoxiao, ZHANG Xueqin. Multi-Object Tracking Algorithm Based on CNN-GRU Metric Network[J]. Journal of East China University of Science and Technology, 2021, 47(4): 483-493. DOI: 10.14135/j.cnki.1006-3080.20200331001

    基于CNN-GRU度量网络的多目标跟踪算法

    Multi-Object Tracking Algorithm Based on CNN-GRU Metric Network

    • 摘要: 针对复杂场景下多目标跟踪算法存在目标标识切换率高、目标轨迹误报率高的问题,提出了一种基于行人重识别网络和CNN-GRU(Convolutional Neural Networks-Gated Recurrent Unit)度量网络的多目标跟踪算法。通过构建一个CNN和双GRU网络结合的深度度量模型,同时预测跟踪目标轨迹框外观特征和运动特征的时间特性,使提取的目标特征更具有判别性,降低目标的标识切换率。基于CNN-GRU网络自动学习历史目标轨迹框正确匹配的概率,给同一目标的不同轨迹框分配不同的注意力,以此来抑制目标轨迹中误检的目标框对目标整体特征的影响,在降低误报率的同时有效聚合轨迹框的特征。该算法将行人重识别网络输出的特征计算得到的检测框和轨迹框的相似度,以及CNN-GRU网络直接输出的相似度作为数据关联部分的匹配成本。在标准多目标跟踪数据集上的实验结果验证了本文算法的有效性。

       

      Abstract: Aiming at the shortcomings in multi-object tracking in complex scenes, e.g., high object identification switching rate, high object trajectory false alarm rate, etc., this paper proposes a multi-object tracking algorithm based on pedestrian re-identification network and CNN-GRU (Convolutional Neural Networks-Gated Recurrent Unit) metric network. By constructing a deep metric model combining CNN and dual GRU network, the time characteristics of both the appearance and the motion features of the tracking object trajectory boxes can be predicted simultaneously so that the extracted objects have more discriminative features and the ID switch rate of object can be reduced. The CNN-GRU network is utilized to automatically learn the correct matching probability of historical object trajectory and different attentions are assigned to different track trajectory boxes of the same object. Thus, the proposed algorithm can effectively attenuate the influence of mis-detected object boxes in the object trajectory on the overall features of the object, meanwhile, aggregate the features of the object trajectory box. Moreover, it combines the similarity of the detection boxes and the calculated trajectory boxes via the features of pedestrian re-identification network, and takes the similarity CNN-GRU network output as the matching cost of data association part. Finally, the experimental evaluation results on a standard multi-object tracking dataset verify the effectiveness of the proposed algorithm.

       

    /

    返回文章
    返回