基于Transformer的无人机多目标跟踪算法研究
Transformer-based multi-object tracking algorithm for UAV
苑玉彬 1吴一全1
作者信息
- 1. 南京航空航天大学 电子信息工程学院,南京 210016
- 折叠
摘要
针对无人机多目标跟踪面临目标遮挡、尺度变化、快速运动、复杂环境等问题,提出了一种基于Transformer的无人机多目标跟踪算法,采用Focal Transformer来捕获高分辨率输入的Transformer层中的局部和全局交互.该算法能够生成目标的检测信息以及目标的外观特征,从而提高了跟踪性能.在轨迹预测方面,采用了卡尔曼滤波方法,以准确地预测目标的运动轨迹,有助于提高跟踪的准确性和鲁棒性.在数据关联过程中,同时考虑了检测置信度、外观嵌入距离和IOU距离3 个因素,以更有效地处理数据关联过程,提高了多目标跟踪模型的鲁棒性,使其能够在复杂场景中更好地跟踪目标.此外,还使用了轨迹的二次匹配方法,进一步提高了算法的性能.在 VisDrone 和UAVDT数据集上进行了对比验证,证明了该算法在实际应用中的有效性和可行性.本研究为无人机多目标跟踪提供了一种新的解决方案,具有广泛的应用前景.
Abstract
In response to the challenges faced in UAV multi-object tracking,including target occlusion,scale variations,rapid movements,and complex environments,this study introduces a UAV multi-object tracking algorithm based on the Transformer architecture.It leverages the Focal Transformer to capture both local and global interactions within the Transformer layers for high-resolution input.This algorithm is capable of generating target detection information and appearance features,thereby significantly enhancing tracking performance.For trajectory prediction,it incorporates the Kalman filtering method to accurately forecast target motion paths,contributing to improved tracking accuracy and robustness.In the data association process,it simultaneously considers three factors:detection confidence,appearance embedding distance,and IOU distance.This enhances the robustness of the multi-object tracking model and enables it to better track targets in complex scenarios.Furthermore,a secondary matching approach for trajectories is employed to further boost the algorithm's performance.Comparative validation on the VisDrone and UAVDT datasets demonstrates the effectiveness and feasibility of this algorithm in practical applications.This research presents a novel solution for UAV multi-object tracking,with promising applications across a wide range of scenarios.
关键词
多目标跟踪/Transformer/卡尔曼滤波/检测置信度/多特征匹配Key words
multi object tracking/transformer/Kalman filtering/detection confidence/multiple feature matching引用本文复制引用
基金项目
江苏省研究生科研与实践创新计划项目(KYCX24_0583)
南京航空航天大学博士学位论文创新与创优基金项目(BCXJ24-10)
出版年
2024
NETL
NSTL
万方数据