无人机目标跟踪技术中的研究进展

Research Progress of UAV Object Tracking

无人机因其体积小,移动灵活等特点,常用于目标跟踪场景上,同时随着人工智能的迅速发展,越来越多的深度学习算法应用于无人机跟踪上。该文着力于探索近年无人机跟踪中所使用的算法,通过分析归纳出各种算法的特点并进行分类,通过实验分析比较各种算法的优缺点。把常用的无人机跟踪算法分为四类,分别为基于判别相关滤波的目标跟踪算法、基于孪生网络的目标跟踪算法,判别式预测模型和基于Transformer的目标跟踪算法。首先通过分析各种算法的共同点而得出它们的基础框架;然后通过举例分析近年来的代表性算法以回顾其发展,并对各算法的特点进行分析;最后使用四个无人机数据集进行算法评估。根据评估结果提出未来研究方向,得出不仅要保证算法同时具有较高的跟踪精度和跟踪速度,而且还要加强算法的鲁棒性以应对无人机跟踪场景中的挑战。

Unmanned aerial vehicles are commonly used in target tracking scenarios due to their small size and flexible movement.With the rapid development of artificial intelligence,more and more deep learning algorithms are applied to UAV tracking.We focus on exploring the algorithms used in UAV tracking in recent years,summarizing the characteristics of various algorithms and classifying them,and comparing the advantages and disadvantages of various algorithms through experiments.The commonly used UAV tracking algorithms are classified into four categories,which are DCF-based model for UAV tracking,Siamese-based for UAV tracking,discriminative prediction model and Transformer-based for UAV tracking.Firstly,we analyze the common points of various algorithms and derive their basic framework.Then we review the development of recent representative algorithms by citing examples and analyze the characteristics of each algorithm.Finally,we use four UAV datasets to evaluate the algorithms and propose the future research direction.It is derived not only to ensure that the proposed algorithm has both high tracking accuracy and tracking speed,but also to enhance the robustness of the algorithm to cope with the challenges in UAV tracking scenarios.