摘要
由于无人机体积小、机动性强,非常适合于难以收集信息的区域执行工作,因此,无人机航拍影像的目标检测和跟踪在许多领域发挥了巨大的辅助作用。然而,由于无人机本身的特殊位置和观察角度,实际目标跟踪面临许多困难。例如,小目标、复杂背景和遮挡等问题。其中目标遮挡问题最为普遍,一旦出现目标被遮挡的情况,原有的跟踪算法会发生偏移,并持续影响后续的跟踪任务。为了解决传统相关滤波器算法的轨迹漂移问题,本文提出了一种基于相关滤波器与重检测机制的无人机航拍目标跟踪方法。该算法使用Faster-RCNN或YOLO模型作为重检测模块,利用相关滤波器的跟踪过程来计算响应图的重检测指数,以在线确定难以跟踪的目标位置。无人机和大规模跟踪数据集上进行的大量实验,实验表明,该方法取得了比基于相关滤波跟踪器更高的跟踪准确率和跟踪成功率,并进一步降低了视频遮挡和形态变化带来的目标跟踪漂移风险。
Due to their small size and high maneuverability,drones are highly suitable for performing tasks in areas where information collection is difficult.Therefore,drone object visual detection and tracking play an important role as auxiliary means in many fields.However,due to the unique positioning and observation angle of drones,practical object tracking faces many challenges,such as small objects,complex backgrounds,and occlusions.Among these challenges,object occlusion is the most common.Once an object is occluded,the existing tracking algorithm will experience drift and affect subsequent tracking tasks.To address the trajectory drift issue in traditional correlation filter algorithms,this paper proposes a drone object tracking method based on correlation filtration and a re-detection mechanism.The algorithm uses the Faster-RCNN or YOLO model as a re-detection module and utilizes the tracking process of correlation filters to calculate the re-detection score of the response map,thereby determining the position of difficult-to-track targets online.Extensive experiments on drone and large-scale tracking datasets have been conducted.Results demonstrate that this method achieves higher tracking accuracy and success rate compared to correlation filter-based trackers.It further reduces the risk of object tracking drift caused by video occlusion and shape variation.
作者
仇甲柱
方亮
姚睿
周勇
赵佳琦
QIU Jiazhu;FANG Liang;YAO Rui;ZHOU Yong;ZHAO Jiaqi(School of Computer Science and Technology,China University of Mine and Technology,Xuzhou 221116,China)
出处
《中国体视学与图像分析》
2023年第2期155-167,共13页
Chinese Journal of Stereology and Image Analysis
基金
国家自然科学基金(No.62172417,62272461)
徐州市重点研发计划(KC22287)。
关键词
UAV
目标跟踪
目标检测
相关滤波
重检测
UAV
object tracking
object detection
correlation filter
re-detection
