无人机视频运动目标检测跟踪关键技术研究

Unlike traditional fixed view surveillance video, UAV video has the characteristics of flexible view, continuous view and wide monitoring scope. Its effective analysis can improve the intelligence level of monitoring. However, the UAV video is dynamically captured by overlooking, and the appearance feature of object is further reduced. Meanwhile, in such non-controlled environment, the uncertain size of object, scene changes, strong motion noise, as well as the dim small object in long-distance observation, these all bring new challenges to video analysis and understanding. This project is oriented to the demand of public security video monitoring and application, and studies the detection and tracking of moving object in UAV video. A fast and precise feature matching method based on texture and scale difference is proposed, and image alignment method based on distortion correction using multi-structure scene is implemented. They both guarantee the precise image registration, which is anti-interference even in complex scenes. The change detection based on the difference stability of spatial random pixels is proposed to get the slight change of pixel level. At the same time, the accuracy of detection is further improved under the guidance of scene perception and object recognition in high-level semantics. The detailed description of the object in case of limited features and the method of measuring the similarity of the object in multiple perspectives are proposed to realize the long time tracking of the object.

有别于传统的固定视角监控视频，无人机视频具有视角灵活、视域连续、监视范围广等特点，对其有效分析可提高监控的智能化程度。然而，无人机视频由空中动态俯拍获取，俯视导致目标表观信息进一步减少，同时，非受控环境下存在目标尺寸不定、场景变化大、运动噪声强等问题，远距离观察下目标尤其弱小，为视频分析与理解带来新的挑战。为此，本项目面向公共安全视频监控应用需求，对无人机视频中动目标检测跟踪开展研究。提出兼顾纹理与尺度差异的特征快速精确匹配、基于多结构场景区域畸变校正的图像对齐方法，实现复杂场景下抗干扰的视频帧间精确配准；提出基于空间随机像素值差异稳定性的变化检测，得到像素级的细微变化，同时在场景感知、目标识别高层语义指导下进一步提升检测的准确性；提出特征有限情况下目标的精细化描述，以及多视角下目标相似度度量方法，实现目标广域长时跟踪。

针对现有无人机视频运动目标检测跟踪方法在平台运动、目标弱小、视角多变等复杂情况下稳健性差、准确性低的问题，本项目提出兼顾尺度与纹理差异的特征快速精确匹配，并建立多结构场景区域畸变校正模型，解决现有特征分布不均匀、高污染下特征易误匹配的问题，实现了局部畸变区域对齐；通过整合像素级检测与高层语义识别的优势，结合场景上下文提出复杂场景下准确快速的运动目标检测方法，克服目标弱小、训练样本少的缺陷，挖掘细微线索实现了准确检测；通过建立多尺度多视角目标结构化描述模型，并提出多视角视觉特征相似性度量方法，实现了目标的高效准确表达与度量，提高了广域长时目标跟踪的准确性和稳健性。.在研究过程中，项目组构建了数十小时的无人机数据集并标注了20万以上的目标，以及基于UAV123改进的运动目标数据集，这些数据集后期将陆续发布，推进本领域发展；突破了语义分割、精确快速配准、弱小目标检测、长时跟踪等关键技术，提出了新算法并发表了论文、申请了专利、软件著作权；上述技术在公安部门得到了应用示范，获得了某地公安局的好评及应用证明；这些在无人机视频分析的模型方法与关键技术方面的突破与创新，获得了国家科学技术进步奖二等奖一项（2019年12月，面向公共安全的大规模监控视频智能处理技术及应用，本项目负责人排名第10）；负责人承担的军委科技委的创新基金“XXXX弱小目标准确XX与XX关键技术”（2019年9月-2021年9月）也正是基于上述理论研究基础进行的装备技术深化，该创新基金于2021年底顺利完成了验收；同时，本项目研制的无人机目标检测跟踪系统在中电科53所、中国兵器工业第205所、航天9院无人机平台上进行了验证，取得了很好的效果。公安部门、军事领域的应用都表明本项目具有良好的应用前景。