仇耀宗,李琳,郭皓捷,于清泽.面向船闸船舶的在线多目标跟踪技术研究[J].装备环境工程,2024,21(3):73-79. QIU Yaozong,LI Lin,GUO Haojie,YU Qingze.Ship Online Multi-object Tracking in Lock Approach Channel[J].Equipment Environmental Engineering,2024,21(3):73-79. |
面向船闸船舶的在线多目标跟踪技术研究 |
Ship Online Multi-object Tracking in Lock Approach Channel |
投稿时间:2023-12-12 修订日期:2024-02-01 |
DOI:10.7643/issn.1672-9242.2024.03.010 |
中文关键词: 在线多目标跟踪 船闸船舶 改进FairMOT 上下文联系 Contextual Transformer 上下文注意力中图分类号:U675.79 TP391.41 文献标志码:A 文章编号:1672-9242(2024)03-0073-07 |
英文关键词:online multi-object tracking ship lock improved FairMOT context information Contextual Transformer context attention |
基金项目:研究所产业资助项目(MYXM22020) |
|
|
摘要点击次数: |
全文下载次数: |
中文摘要: |
目的 满足船闸船舶在线跟踪要求,改善由于复杂背景、遮挡等因素导致轨迹不连续和身份变更的问题,提出一种增强上下文联系和上下文注意力的多目标跟踪方法。方法 基于设计的在线系统,采集连续帧图像,改进FairMOT多目标跟踪模型。首先,通过在骨干网络设计基于Bottleneck和Contextual Transformer的上下文建模模块,以加强上下文联系,增强场景理解的能力。其次,在迭代聚合后的特征图上应用全局上下文注意力,提高定位船舶目标的能力。结果 相对于原生的FairMOT方法,设计上下文建模模块后,多目标跟踪准确度指标MOTA提高2.1%,继续添加全局上下文注意力MOTA,共计提高3.5%,同时在多项指标中取得了最佳表现。结论 改进的FairMOT方法不仅拥有更强的轨迹保持能力,而且在身份维持方面更胜一筹。 |
英文摘要: |
The work aims to propose a method of multi-object tracking to enhance contextual connection and attention to meet the requirements of ship online tracking in lock approach channel, and to ameliorate the problem of discontinuous trajectories and identity changes caused by complex backgrounds, occlusion, and other factors. The multi-object tracking model named of FairMOT was improved by continuous frame images captured from the online monitoring system. Firstly, a block based on Bottleneck of FairMOT and Contextual Transformer (BoCoT), was constructed in the backbone to exploit contextual information and strengthen the representative capability. Secondly, Global Context Attention (GCA) module was embedded after the iterative aggregation layer to assist in discriminating the object locations. The experimental results showed that, Multiple Object Tracking Accuracy (MOTA) index after context modeling was increased by 2.1% compared with the original FairMOT method, and it obtained a 3.5% increase totally after continuing to embed GCA module. The improved model also achieved the best performance in multiple evaluation indexes. In conclusion, the improved FairMOT not only has stronger trajectory retention ability, but it also excels in identity maintenance. |
查看全文 查看/发表评论 下载PDF阅读器 |
关闭 |
|
|
|