| 张立国,张升,章玉鹏,耿星硕,金梅.基于无锚框的孪生网络目标跟踪改进算法[J].高技术通讯(中文),2023,33(6):610~619 |
| 基于无锚框的孪生网络目标跟踪改进算法 |
| An improved target tracking algorithm based on frameless twin networks |
| |
| DOI：10. 3772/ j. issn. 1002-0470. 2023. 06. 006 |
| 中文关键词: 目标跟踪； 特征提取； 孪生神经网络； 精度 |
| 英文关键词: target tracking, feature extraction, siamese neural network, accuracy |
| 基金项目: |
| 作者 | 单位 | | 张立国 | （燕山大学电气工程学院秦皇岛 066000 ） | | 张升 | （燕山大学电气工程学院秦皇岛 066000 ） | | 章玉鹏 | （燕山大学电气工程学院秦皇岛 066000 ） | | 耿星硕 | （燕山大学电气工程学院秦皇岛 066000 ） | | 金梅 | （燕山大学电气工程学院秦皇岛 066000 ） |
|
| 摘要点击次数: 980 |
| 全文下载次数: 880 |
| 中文摘要: |
| 视觉目标跟踪在车辆、人机交互以及监控等领域应用广泛，虽然近年来取得了很大的进展，但是在跟踪过程中，仍然存在许多的干扰因素。针对跟踪过程存在目标尺度和长宽的比例会随着目标或跟踪设备的变化而变化以及背景干扰的问题，设计了一种基于无锚框的孪生神经网络的跟踪方法。首先，改进了特征提取网络，提高了跟踪的准确性。其次，增加了非局部感知网络，能够更好地利用模板和搜索分支更深度的特征。对于分类来说，增加了选择分支，用于抑制较低的得分，选择更高更准确的得分，从而能够进行更好的回归预测。其采样策略也不同于之前的网络，并对损失部分进行了优化。在对网络进行整体的训练及实验之后，该算法能够很好地跟踪目标，提高了跟踪的成功率和精确度。 |
| 英文摘要: |
| Visual target tracking technology is widely applied in vehicles, human-computer interactions, monitoring and other fields. Despite great progress that has been made in recent years, the current visual target tracking methods still suffer from many interference factors that affect the tracking process. To cope with the problem that the scale and the length-to-width ratio of the target vary with changes of targets or the tracking devices and background interference in the tracking process, a tracking method based on frameless twin neural network is designed. First of all, the feature extraction network is improved to increase tracking accuracy. Meanwhile, non-local sensing network is introduced, which can make better use of the template and deeper features of search branch. For classification, the selection branch is incorporated to suppress low scores and select higher and more accurate scores, which enable better regression prediction. In addition, the sampling strategy differs from the previous network, and the loss function is optimized. With the whole network training and experiments conducted on the network, the algorithm performs better in target tracking with higher success rate and accuracy. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
