CORC  > 自动化研究所  > 中国科学院自动化研究所  > 毕业生  > 博士学位论文
题名航拍视频运动目标检测关键技术研究
作者申浩
学位类别工学博士
答辩日期2013-05-28
授予单位中国科学院大学
授予地点中国科学院自动化研究所
导师杨一平 ; 常红星
关键词无人飞行器 运动目标检测 图像配准 显著性 基于跟踪的检测 目标检测 航拍视频 unmanned aerial vehicle moving object detection image registration saliency detection-by-tracking object detection aerial video
其他题名Research on Key Technologies of Moving Object Detection in Aerial Video
学位专业计算机应用技术
中文摘要在航拍视频中自动检测运动目标,是无人机执行空中侦察、火力打击等高层任务的基础,是提升无人机自主能力的关键技术之一。相对于摄像机固定的视频监控,无人机航拍具有机动灵活,监视范围广的优点,但同时也面临着更多的挑战,如目标尺寸较小、图像模糊、相机运动和计算资源有限等。现有的航拍视频运动目标检测算法在环境适应性、检测结果稳定性和检测耗时方面都有待提高,距离实用仍有很大的差距。本文针对无人机运动目标检测中所涉及的图像配准、运动目标分割和检测结果修正三部分关键内容,展开深入研究。论文主要工作如下: (1)针对无人机航拍场景的特点,提出了一种适用于航拍视频的快速、鲁棒的帧间图像配准算法。通过空间约束和角点量限制获得稳定的、空间分布均匀的FAST特征点,保证在图像背景平坦的情况下仍能获得足够的背景特征描述,增强算法对环境的适应性;利用视频中相邻帧间图像内容的连续性,对特征检测阈值进行自适应调整,以获得合适数量的特征点。采用训练得到的不相关采样点集对特征点进行二值描述,并通过最近邻算法根据汉明距离获得特征匹配对,进而实现准确快速的特征描述与匹配;最后运用RANSAC方法得到帧间仿射变换模型参数。实验结果表明该算法快速、稳定,具有较高的环境适应性,能够满足无人机航拍视频帧间配准的要求。 (2)提出了一种基于时空显著性的运动目标检测算法。航拍视频中,运动目标具有较强的时间显著性,但由于目标较小,其空间显著性在全局图像中并不明显,而仅在局部区域较为突出。本文基于时空显著性信息,提出了一种递进式的运动目标检测方法。首先利用时间显著性获得初步的目标分割结果即候选区域,然后在候选区域中计算目标的空间显著性,最后通过融合时间显著性和空间显著性信息获得最终的目标检测结果。其中为全面描述目标的空间特性,空间显著性信息分别在像素和区域两个不同的层面提取,由像素显著性获得目标的细节描述,采用区域显著性反映目标的整体特性。时间、像素和区域三类显著性信息互为补充,保证了运动目标检测算法的实时性和准确性。 (3)提出了一种基于跟踪的运动目标检测算法。通过融合跟踪信息,对运动目标检测结果进行修正,减少单一的目标检测算法因帧与帧间结果没有关联而出现的目标多检或漏检问题。该方法分别在候选区域提取和最终结果融合两个阶段运用跟踪信息对运动目标检测算法进行优化。在候选区域提取阶段,采用前序帧的目标跟踪结果对目标在当前帧中可能出现的位置进行预测,弥补前向运动历史图像将来信息累积不足的缺点,得到更为准确的目标候选区域;在检测跟踪结果融合阶段,一方面利用跟踪轨迹信息对目标检测结果进行修正,消除多检或漏检错误,另一方面根据检测跟踪目标关联结果,对最终的目标检测结果进行精确定位。此外,对经典mean shift 算法进行了改进,在跟踪算法中加入了全局跟踪策略。在大尺度区域内统计候选目标模型,使其精度低于目标模型,以获得稳定的跟踪权重结果图;通过在全局范围内搜索最佳初始位置,保证跟踪的全局特性。实验结果表明,通过融合跟踪信息,能有效去除目标多检、...
英文摘要Automatically moving targets detection in aerial video is a foundation step for the high level tasks such as aerial reconnaissance, fire attack and so on. Also, it is one of key technologies to improve the automatic ability of Unmanned Aerial Vehicles(UAV). In contrast to applications with fixed cameras, aerial surveillance has the advantages of higher mobility and larger surveillance scope. Meanwhile, more challenges are involved in aerial video, such as small object size, blurring image and camera motion. There is still a big gap between the existing detection algorithms and the utility till now,and many aspects need to be improved, such as environmental adaptability, results stability and time consuming. In this paper, we mainly focus on three aspects of moving object detection: image registration, object segmentation and detection result refinement. The main contributions are summarized as follows: (1) According to the characteristics of the UAV aerial scenes, a fast and reliable interframe registration method for aerial videos is proposed. Firstly, the stable FAST corners are selected via the constraints of spatial displacements and cornerness measurements, which can enhance the algorithm environmental adaptability, and make sure the algorithm can extract sufficient background feature descriptions in the case of flat scene. Secondly, via taking advantage of the continuity of image content in video, the threshold of feature detection is adjusted adaptively, so as to obtain suitable number of feature points. The binary descriptions of the detected features are generated by using the uncorrelated sample point set, which is obtained by training, and the matched points are established using the NN (Nearest Neighbor) algorithm based on hamming distances. Finally, the affine transformation parameters between adjacent frames are estimated using the matched points by RANSAC. Experimental results show that the proposed algorithm is fast and reliable, has high environmental adaptability, and thus can meet the image registration requirements in UAV systems. (2) A novel spatiotemporal saliency-based moving object detection algorithm is proposed. In aerial video, the moving object has strong temporal saliency in global image, while due the small size of object, the spatial saliency is more prominent in the local area. To deal with this issue, we propose a novel hierarchical moving object detection method based on spatiotemporal saliency. Firstly, the tempor...
语种中文
其他标识符201018014629091
内容类型学位论文
源URL[http://ir.ia.ac.cn/handle/173211/6523]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
申浩. 航拍视频运动目标检测关键技术研究[D]. 中国科学院自动化研究所. 中国科学院大学. 2013.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。


©版权所有 ©2017 CSpace - Powered by CSpace