【作者】 胡涛；

【导师】 李开；

【作者基本信息】 华中科技大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 近年来,无人机凭借着其灵活轻巧的特性,逐渐在交通管理、军事行动等场景中展现出优势,面向无人机场景的多目标跟踪算法在中间扮演着重要角色。目前该算法的研究主要面临两个问题:一方面,无人机航拍视频中的目标物体较小,跟踪算法无法准确检测出小目标;另一方面,无人机的运动是不规则的,无人机和目标物体的运动耦合导致跟踪算法的数据关联阶段准确率不高。基于上述背景,本课题的工作内容有以下三点。(1)针对无人机视角下普通跟踪算法难以检测小目标导致跟踪效果不佳的问题,对网络结构进行改进,提出了基于特征融合和任务解耦的无人机视频跟踪算法。首先设计了多支路特征融合模块,从多个分支中学习不同感受野的特征图信息,提升模型对各种尺寸目标的辨别力。同时加入了任务解耦模块,让模型自发学习不同任务的特征图信息,解决检测和重识别两者之间的优化竞争问题。然后设计了尺度自适应损失函数,再一次增强了模型对小目标的检测能力。(2)针对无人机和目标物体存在运动耦合导致相邻帧轨迹关联不准确的问题,仅修改网络结构无法有效解决,因此对数据关联部分进行改进,提出了基于运动自适应和轨迹恢复的无人机视频跟踪算法。首先设计了运动自适应滤波器,根据无人机的不同运动模式,自适应地选择对应的关联匹配方法,提高数据关联的准确率。其次提出了基于插值的轨迹恢复策略,使用线性插值技术计算出目标丢失的轨迹位置信息,并基于阈值筛选出合适的轨迹进行恢复,减少无人机视频跟踪过程中身份切换的次数。(3)在VisDrone2019和UAVDT两个公开的无人机视频跟踪数据集上,对基于特征融合和任务解耦的无人机视频跟踪算法和基于运动自适应和轨迹恢复的无人机视频跟踪算法分别进行了消融实验、对比实验和可视化分析,实验结果验证了上述两个算法在无人机视频跟踪场景中具有明显优势。更多还原

【Abstract】 In recent years,unmanned aerial vehicle(UAV)have been gradually used in traffic management,military operations and other scenes due to their flexibility and lightness,in which multi-object tracking algorithms for UAV scenes play an important role.At present,the research of this algorithm mainly faces two problems: on the one hand,the target object in the UAV videos is small,and the tracking algorithm cannot accurately detect the small target,on the other hand,the movement of the UAV is irregular,and the coupling of the movement of the UAV and the target object leads to the low accuracy of the data association stage of the tracking algorithm.Therefore,based on the above background,the work content of this topic has the following three points.(1)In order to solve the problem that the ordinary target tracking algorithm is difficult to detect small targets from the perspective of UAV,resulting in poor tracking effect,the network structure is improved,and an UAV video tracking algorithm based on feature fusion and task decoupling is proposed.Firstly,a multi-branch feature fusion module is designed to learn the feature map information of different receptive fields from multiple branches,so as to improve the discrimination of the model for targets of various sizes.At the same time,the task decoupling module is added to allow the model to spontaneously learn the feature graph information of different tasks,and solve the problem of optimal competition between detection and re-identification.Then,the scale adaptive loss function is designed,which once again enhances the detection ability of the model for small targets.(2)In order to solve the problem that the data association of adjacent frames is inaccurate due to the motion coupling between the UAV and the target object,which cannot be effectively solved by modifying the network structure alone,the data association part is improved,and an UAV video tracking algorithm based on motion adaptation and trajectory recovery is proposed.Firstly,a motion adaptive filter is designed to adaptively select the corresponding association matching method according to the different motion modes of the UAV to improve the accuracy of data association.Secondly,a trajectory recovery strategy based on interpolation is proposed,which uses linear interpolation technology to calculate the trajectory position information of the target loss,and selects the appropriate trajectory for recovery based on the threshold,so as to reduce the number of identity switching in the process of UAV videos tracking.(3)On the two public UAV video tracking datasets of VisDrone2019 and UAVDT,ablation experiments,comparative experiments and visualization analysis are carried out on the UAV video tracking algorithm based on feature fusion and task decoupling and the UAV video tracking algorithm based on motion adaptation and trajectory recovery,respectively,and the experimental results verify that the above two algorithms have obvious advantages in the scene of UAV video tracking.更多还原