无人机视频中目标跟踪技术研究

Different from traditional surveillance video data, UAV videos have new features including large scale changes, complex motion patterns, continuous changes in angle, etc., and the existing methods can hardly achieve good tracking results. In order to improve intelligent level of UAV in the monitoring application, effective target detection, and tracking technology is needed for accurate, fast, long-term tracking. To this end, this project aims to carry out research on the problems related to the moving object detection and tracking for UAV video analysis. Firstly, we jointly perform view estimation and multi-scale object detection to detect and classify the objects. Secondly, we will study the semi-supervised cross-domain robust appearance modeling methods and temporal evolution model to build discriminative appearance model for the target. Then, to predict the movement of the target, we propose to mine motion model from the obtained tracklets using vector field clustering. Finally, an adaptive multi-clue fusion method is proposed to integrate the appearance and motion information to obtain a comprehensive similarity of the tracklets. This project is potential to achieve long-term tracking in UAV video, and provide important clues for high-level UAV video analysis, making UAV to play a more important role in sports, urban monitoring and other aspects.

与传统的监控视频数据不同，无人机视频具有尺度变化大、运动模式复杂、角度连续多变等特点，现有的方法难以取得好的跟踪效果，亟需有效的目标检测、跟踪技术来对无人机视频中的运动目标进行准确、快速、长时的跟踪，以提高其在监控应用中的智能化水平。为此，本课题拟针对无人机视频分析中的运动目标检测及跟踪相关的问题开展研究，首先拟采用基于联合视角估计的多尺度目标检测方法对目标进行检测及分类；其次拟研究半监督跨域的鲁棒表观建模技术和时序演化预测模型，为目标构建有判别力的表观模型；然后，根据轨迹片段，拟构建基于向量场聚类的运动模式挖掘模型，准确预测目标的运动趋势；最后，提出自适应的多线索融合方法，有效融合表观及运动的相似性，得到轨迹片段的准确关联。本项目有望实现无人机视频中目标的长时跟踪，为高层的视频分析任务提供重要的线索，使无人机在体育、城市监控等方面发挥更加重要的作用。

经济的发展促使着人们不断的提高安防意识，传统的视频监控通过安装摄像机对固定区域实施监控，已逐渐不能满足某些行业大范围、远距离监控、移动性强的室外监控。无人机具有机动性高、灵活性强、结构简单、造价便宜、适应能力强、隐蔽性好、便于携带等优点，在视频监控领域得到广泛的应用和关注。与传统的监控视频数据不同，无人机视频具有尺度变化大、运动模式复杂、角度连续多变等特点，现有的方法难以取得好的跟踪效果，亟需有效的目标检测、跟踪技术来对无人机视频中的运动目标进行准确、快速、长时的跟踪，以提高其在监控应用中的智能化水平。项目围绕无人机视频中运动目标检测及跟踪技术的相关问题开展研究，研究内容多视角多尺度特征融合方法、运动目标检测方法及跨域目标表观建模技术和目标运动模式挖掘及自适应的多线索融合算，实现了无人机视频中运动目标的检测与跟踪，为更高层的视频分析与理解提供轨迹线索。项目实施期间共发表和录用学术论文24篇：其中IEEE/ACM汇刊论文和CCF-A类论文14篇。发表的国际期刊论文中7篇包括International Journal of Computer Vision 1篇、IEEE Trans. on Image Processing 2篇、IEEE Trans. on Multimedia 1篇、 Information Sciences 2篇、 Pattern Recognition Letter 1篇；国际会议论文中包括IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) 3篇；ACM Multimedia（长文）4篇; IEEE International Conference on Computer Vision(ICCV) 2篇, AAAI Conference on Artificial Intelligence (AAAI) 1篇；European Conference on Computer Vision 2篇等。申请国家发明专利5项。项目组1人获选IEEE Senior Member，2人入选CCF高级会员。培养的硕士生、博士生各一名。依托研究成果，搭建了无人视频目标跟踪平台，并在相关企业进行了验证。