基于注意力机制与图网络的多目标跟踪
投稿时间:2022-08-14  修订日期:2022-09-13  点此下载全文
引用本文:
摘要点击次数: 21
全文下载次数: 0
作者单位邮编
杨坤 西安工程大学 710048
任劼 西安工程大学 
基金项目:1.陕西省自然科学基础研究计划(2022JM-394)资助 2.陕西省教育厅科研计划项目(19JK0364)资助
中文摘要:在基于检测的多目标跟踪算法中,为了获取更具鉴别性的特征以及解决复杂场景下目标的频繁带来的目标丢失以及身份切换问题,提出的方法为一种基于图网络与注意力机制多目标跟踪算法。该算法利用Resnet-34-CBAM网络作为外观特征提取网络,分别将相邻帧的外观特征、位置信息利用特征融合网络进行融合,将获得的融合特征与运动特征分别使用不同更新策略的图网络进行更新,分别获得融合特征与运动特征相似度,使用超参数将两种相似度结合,进而获得相邻帧目标之间的相似度。最终使用匈牙利算法完成关联实现跟踪任务。最后在MOT17数据集进行实验,相较MOTDT算法,MOTA指标提升2.7%,MOTP指标提升6.4%,IDF1指标提升5.9%。实验结果证明,提出的基于图网络与注意力机制的多目标跟踪算法可以有效提高多目标跟踪的整体性能,并有效降低身份切换。
中文关键词:注意力机制  多目标跟踪  图网络
 
Multi-target tracking based on attention mechanism and graph network
Abstract:In multi-target tracking algorithm based on detection, in order to obtain more discriminative features and solve the problem of target loss and identity switching caused by frequent targets in complex scenes. This paper proposed a multi-target tracking algorithm based on graph network and attention mechanism. The method used Resnet-34-CBAM network as the appearance feature extracting network, fuses the appearance feature of adjacent frames and location information with feature fusion network, updates the obtained fusion feature and motion feature using graph network with different update strategies, and obtains appearance and motion similarity respectively. The similarity between adjacent frames is obtained by combining the two similarity degrees with hyperparameters. Finally, the Hungarian algorithm is used to complete the tracking task. Finally, experiments were carried out in MOT17 dataset. Compared with MOTDT algorithm, MOTA index improved 2.7%, MOTP index improved 6.4%, and IDF1 index improved 5.9%. Experimental results show that the proposed multi-target tracking algorithm based on graph network and attention mechanism can effectively improve the overall performance of multi-target tracking and reduce identity switching.
keywords:attentional mechanism  multiple object tracking  graph network
查看全文   查看/发表评论   下载pdf阅读器