| 基于Wasserstein度量的目标数据关联算法 |
| |
| 引用本文： | 刘洋,郭春生.基于Wasserstein度量的目标数据关联算法[J].教育技术导刊,2019,18(10):74-77. |
| |
| 作者姓名： | 刘洋 郭春生 |
| |
| 作者单位： | 杭州电子科技大学 通信工程学院，浙江 杭州 310018 |
| |
| 摘 要： | 针对目前多目标跟踪中目标数据关联度量方式的不足，以及Wasserstein度量值衡量概率测度间差异程度的性质，提出基于Wasserstein度量的目标数据关联算法，即利用Wasserstein距离衡量目标外观特征向量之间的相似度，将目标外观特征向量看成一个分布，计算分布之间的 Wasserstein距离，再用Wasserstein距离判断目标是否相似。但是Wasserstein距离表达式比较复杂，难以直接计算，因此通过训练一个深度网络计算Wasserstein距离，并使相同目标特征向量之间Wasserstein距离缩小、不同目标特征向量之间的Wasserstein距离增大；然后，利用目标运动匹配度进一步筛选满足外观匹配度的目标，最终得到最佳目标关联。实验结果表明，该算法能较好地解决多目标跟踪中的漏报问题，与原算法相比，MT提高了6.7%，ML减少了4.9%，FN减少了6 627个。
|
| 关 键 词： | 多目标跟踪 深度网络 Wasserstein距离 数据关联 |
| 收稿时间： | 2019-02-26 |
Target Data Association Algorithm Based on Wasserstein Metric |
| |
| LIU Yang,GUO Chun-sheng.Target Data Association Algorithm Based on Wasserstein Metric[J].Introduction of Educational Technology,2019,18(10):74-77. |
| |
| Authors: | LIU Yang GUO Chun-sheng |
| |
| Institution: | School of Communication Engineering，Hangzhou Dianzi University，Hangzhou 310018， China |
| |
| Abstract: | In view of the shortcomings of the current target data association metrics in multi-target tracking and the nature of Wasserstein metrics to measure the degree of difference between probability measures， this paper proposes the target data association algorithm based on Wasserstein metric which uses Wasserstein distance to measure the similarity between target appearance eigenvectors. In this paper， we consider the appearance feature vector of the target as a distribution， calculate the Wasserstein distance between the distributions， and then use the Wasserstein distance to judge whether the targets are similar. However， the Wasserstein distance expression is more complicated and difficult to calculate directly. In this paper， a deep network is trained to calculate the Wasserstein distance， and the Wasserstein distance between the same target feature vectors is reduced， and the Wasserstein distance between different target feature vectors is increased. Then， the motion matching degree of the target reuse target that satisfies the appearance matching degree is further filtered， and finally the best target association is obtained. The experimental results show that the proposed algorithm can better solve the problem of underreporting in multi-target tracking. Compared with the original algorithm， MT has increased by 6.7%， ML decreased by 4.9% and FN decreased by 6 627. |
| |
| Keywords: | multi-object tracking deep network Wasserstein distance data association |
|
| 点击此处可从《教育技术导刊》浏览原始摘要信息 |
| 点击此处可从《教育技术导刊》下载免费的PDF全文 |
