首页|基于多重信息融合与轨迹关联修正的多目标跟踪方法

基于多重信息融合与轨迹关联修正的多目标跟踪方法

A multi-object tracking method based on multi-information fusion and trajectory correlation correction

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
一阶段多目标跟踪框架由于可以有效提升算法跟踪效率而备受关注,然而该框架在提升效率的同时忽略了检测与关联任务间信息的交互,且目标遮挡的频发会导致轨迹碎片的增加,从而影响跟踪效果.针对这些问题,提出基于多重信息融合与轨迹关联修正的多目标跟踪方法.通过无锚一阶段主干网络,在检测器上另外建立跟踪分支预测跟踪偏移量和嵌入特征信息;设计中和匹配关联模块优化跨帧特征匹配方式,协调检测与关联任务,提升两任务间信息交互能力;采用多重信息融合模块,对时空多层次特征进行融合以获得更加丰富的特征信息;提出轨迹关联修正网络处理因遮挡造成的轨迹碎片,通过改进数据关联方式评估碎片与检测低分目标关系,尝试找回遮挡目标轨迹;将提出的算法在MOT16和MOT17数据集上进行评估,并与其他优异的算法定量比较.通过分析实验结果可以发现,所提出的方法能有效缓解关键性问题,提升算法整体性能.
The one-stage multi-object tracking framework has attracted much attention because it can effectively improve the tracking efficiency of the algorithm.However,the framework ignores the information interaction between detection and associated tasks while improving the efficiency,and the frequent object occlusion will lead to the increase of trajectory fragmentation,which affects the tracking effect.To address these problems,a multi-object tracking method based on multi-information fusion and trajectory correlation correction is proposed.The tracking branch prediction tracking offset and embedded feature information are built on the detector through the anchor-free one-stage backbone network.The designed neutralization matching association module optimizes the cross-frame feature matching mode,coordinates the detection and association tasks,and improves the information interaction capability between the two tasks.The multi-information fusion module is used to fuse spatio-temporal multi-level features to obtain richer feature information.The trajectory correlation correction network is proposed to deal with the trajectory fragmentation caused by occlusion,and to try to retrieve the trajectory of occluding objects by evaluating the relationship between fragmentation and detection of low-scoring objects by improving the data association method.The proposed algorithm is evaluated on MOT 16 and MOT 17 datasets and compared quantitatively with other excellent algorithms.By analyzing the experimental results,it can be found that the proposed methods are very effective,which effectively alleviate key problems and improve the overall performance of the algorithm.

multi-object trackingtrajectory fragmentationembedded featureneutralization matchingmulti-information fusiontrajectory correlation correction

田嘉意、李辉、李赛宇、陈双敏、刘云

展开 >

青岛科技大学信息科学技术学院,山东青岛 266061

多目标跟踪 轨迹碎片 嵌入特征 中和匹配 多重信息融合 轨迹关联修正

国家自然科学基金国家自然科学基金中国高校产学研创新基金新一代信息技术创新项目

62002190617022952021ITA05047

2024

10.13195/j.kzyjc.2022.1673

控制与决策
东北大学

控制与决策

CSTPCD北大核心
影响因子：1.227
ISSN：1001-0920
年,卷(期)：2024.39(6)