摘要
动目标检测是从时序图像中获取时间维度动态变化信息的有效手段,通过快速获取、分析和利用卫星遥感数据中的现势信息,将遥感数据中时空域运动和变化信息表征为对象和知识,可以有效提升卫星数据的利用效率,实现“现势数据、即时传输”的应用要求。现有的卫星视频和多光谱遥感图像用于动目标检测领域,无法满足大范围动目标发现、运动速度适应性以及硬件加速兼容性等应用需求。本文以利用光学遥感卫星实现大范围动目标快速感知为目标,通过双线阵推扫光学遥感相机采集包含时域动目标信息的双条带图像,以船舶目标为例,经过在轨处理形成动目标关键信息。相机原理样机已经搭载“泰景四号01星”于2022年2月27日成功发射,验证了大范围动目标检测的技术路线。本文在提出双线阵推扫成像方法以及动目标时空特性数学模型基础上,构建基于显著性区域建议的动目标检测方法,利用时空域变化信息表征和提取动目标显著性区域,基于深度学习目标检测技术实现动目标判别,获取动目标检测结果。利用在轨拍摄的双条带图像,验证了算法精度,对比传统算法,本文方法在单体目标分割方面提升明显,动目标分割结果较为完整、清晰,有效提升了动目标两个时刻间的匹配精度。结果表明,算法能够有效实现双条带图像中的动目标检测,具有以下2个方面的优势:(1)提出双条带推扫成像模式和方法,扩展图像时域信息,解决卫星无机动条件下大范围动目标观测问题;(2)提出基于显著性区域建议的动目标检测模型,解决了大范围双条带图像中复杂背景下动目标检测问题。通过获取动目标关键信息,大幅度减少遥感信息的对地传输带宽需求,可以为大规模对地观测系统数据在轨处理与地学应用提供新型的数据获取与处理思路。
Moving target detection plays a pivotal role in extracting temporal information from time-series images,particularly from satellite data.This method enables the rapid acquisition,analysis,and utilization of dynamic change information,meeting the demand for"real-time target discovery and delivery."In the processing of optical image-based moving target detection,existing methods often fall short of meeting the requirements for large-scale target discovery,accommodating diverse speeds,and ensuring hardware acceleration compatibility.This study aims to achieve swift perception of large-scale moving targets using optical remote sensing satellites,with a primary focus on both camera innovation and algorithm research in terms of target discovery and target information processing.This paper proposes a novel imaging mode,leveraging a dual-linear array push-broom optical remote sensing camera to capture dual-strip images containing temporal changes associated with moving targets.The camera principle prototype was successfully deployed on the"Taijing-4 Satellite"on February 27,2022,thereby validating the technical approach for large-scale detections.Furthermore,this paper introduces a pioneering approach for detecting moving targets based on saliency region proposal for dual-band images,which significantly enhances the temporal information captured in dual-linear array push-broom imaging.Subsequently,we employ a sophisticated saliency region proposal method to extract the prominent regions of moving targets by utilizing the temporal and spatial change information within the image.These salient regions encompass dynamic targets across the entire image,effectively reducing the amount of intermediate data processed by the algorithm.Finally,a lightweight and efficient deep learning object detection model is leveraged to classify moving targets and eliminate false positives from the initial detection outcomes.The results indicate that the proposed method can efficiently detect moving targets in dual-strip images,substantially im
作者
杨灿坤
李小娟
李韦
钟若飞
李清扬
杜鑫
YANG Cankun;LI Xiaojuan;LIWei;ZHONG Ruofei;LI Qingyang;DU Xin(College of Resource Environment and Tourism,Capital Normal University,Beijing 100028,China;Key Laboratory of 3D Information Acquisition and Application(Ministry of Education),Capital Normal University,Beijing 100028,China)
出处
《地球信息科学学报》
EI
CSCD
北大核心
2024年第4期1040-1056,共17页
Journal of Geo-information Science
基金
国家自然科学基金项目(42271487)。
关键词
双线阵推扫
双条带影像
动目标检测
在轨智能处理
时空域变化信息
单体目标分割
深度学习目标检测
显著性区域建议
dual-linear array push-broom
dual-strip images
moving target detection
on-orbit intelligent processing
single-object segmentation
spatiotemporal change information
deep learning object detection
saliency region proposal
