Spatiotemporal variation of velocity is impor- tant for debris flow dynamics. This paper presents a new method, the trace projection transformation, for accurate, non-contact measurement of a debris-flow surface veloc...Spatiotemporal variation of velocity is impor- tant for debris flow dynamics. This paper presents a new method, the trace projection transformation, for accurate, non-contact measurement of a debris-flow surface velocity field based on a combination of dense optical flow and perspective projection transformation. The algorithm for interpreting and processing is implemented in C ＋＋ and realized in Visual Studio 2012. The method allows quantitative analysis of flow motion through videos from various angles （camera positioned at the opposite direction of fluid motion）. It yields the spatiotemporal distribution of surface velocity field at pixel level and thus provides a quantitative description of the surface processes. The trace projection transformation is superior to conventional measurement methods in that it obtains the full surface velocity field by computing the optical flow of all pixels. The result achieves a 90% accuracy of when comparing with the observed values. As a case study, the method is applied to the quantitative analysis of surface velocity field of a specific debris flow.展开更多
行为识别在语义分析领域具有很高的学术研究价值和广泛的市场应用前景.为了实现对视频行为的准确描述,提出了2类构建稠密轨迹运动描述子的方法.1)通过光流约束和聚类,实现对运动区域的稠密采样,以获取行为的局部位置信息;2)选取目标运...行为识别在语义分析领域具有很高的学术研究价值和广泛的市场应用前景.为了实现对视频行为的准确描述,提出了2类构建稠密轨迹运动描述子的方法.1)通过光流约束和聚类,实现对运动区域的稠密采样,以获取行为的局部位置信息;2)选取目标运动角点为特征点,通过对特征点的跟踪获取运动轨迹;3)在以轨迹为中心的视频立方体内,分别构建三维梯度方向直方图(3Dhistograms of oriented gradients in trajectory centered cube,3DHOGTCC)描述子和三维光流梯度方向直方图(3Dhistograms of oriented optical flow gradients,3DHOOFG)描述子,用以对运动的局部信息进行准确描述.为了充分利用行为发生的场景信息,提出了一种融合动态描述子和静态描述子的行为识别新框架,使得动态特征与静态特征相互融合支撑,即使在摄像头运动等复杂场景下,亦能取得较好的识别效果.在Weizmann和UCF-Sports数据库采用留一交叉验证,在KTH和Youtube数据库采用4折交叉验证.实验证明了提出新框架的有效性.展开更多
针对现有光流估计方法实时性不够的问题,提出轻量化的深度可分离卷积的PWC-Net改进模型(depth separable pyramid,warping and cost volume,DS-PWC)。其改进是将常规二维卷积网络层解耦为深度可分离卷积层,并且DS-PWC在金字塔层增加基...针对现有光流估计方法实时性不够的问题,提出轻量化的深度可分离卷积的PWC-Net改进模型(depth separable pyramid,warping and cost volume,DS-PWC)。其改进是将常规二维卷积网络层解耦为深度可分离卷积层,并且DS-PWC在金字塔层增加基于层数的权重系数,从而使得网络结构在不损失精度的情况下大幅减少模型参数量。在训练过程中,使用图像及对象感知数据随机擦除(image and object-aware random erasing,I+ORE)等数据增强技术,进一步提升估计预测结果泛化能力。实验结果表明,在数据集测试DS-PWC模型,在保持质量的同时运行效率达到约58 fps(frame per second)。同时为了验证算法有效性,进行了模型结构和数据增强的消融实验。结果证明了DS-PWC模型的有效性。展开更多
文摘Spatiotemporal variation of velocity is impor- tant for debris flow dynamics. This paper presents a new method, the trace projection transformation, for accurate, non-contact measurement of a debris-flow surface velocity field based on a combination of dense optical flow and perspective projection transformation. The algorithm for interpreting and processing is implemented in C ＋＋ and realized in Visual Studio 2012. The method allows quantitative analysis of flow motion through videos from various angles （camera positioned at the opposite direction of fluid motion）. It yields the spatiotemporal distribution of surface velocity field at pixel level and thus provides a quantitative description of the surface processes. The trace projection transformation is superior to conventional measurement methods in that it obtains the full surface velocity field by computing the optical flow of all pixels. The result achieves a 90% accuracy of when comparing with the observed values. As a case study, the method is applied to the quantitative analysis of surface velocity field of a specific debris flow.
文摘行为识别在语义分析领域具有很高的学术研究价值和广泛的市场应用前景.为了实现对视频行为的准确描述,提出了2类构建稠密轨迹运动描述子的方法.1)通过光流约束和聚类,实现对运动区域的稠密采样,以获取行为的局部位置信息;2)选取目标运动角点为特征点,通过对特征点的跟踪获取运动轨迹;3)在以轨迹为中心的视频立方体内,分别构建三维梯度方向直方图(3Dhistograms of oriented gradients in trajectory centered cube,3DHOGTCC)描述子和三维光流梯度方向直方图(3Dhistograms of oriented optical flow gradients,3DHOOFG)描述子,用以对运动的局部信息进行准确描述.为了充分利用行为发生的场景信息,提出了一种融合动态描述子和静态描述子的行为识别新框架,使得动态特征与静态特征相互融合支撑,即使在摄像头运动等复杂场景下,亦能取得较好的识别效果.在Weizmann和UCF-Sports数据库采用留一交叉验证,在KTH和Youtube数据库采用4折交叉验证.实验证明了提出新框架的有效性.
文摘针对现有光流估计方法实时性不够的问题,提出轻量化的深度可分离卷积的PWC-Net改进模型(depth separable pyramid,warping and cost volume,DS-PWC)。其改进是将常规二维卷积网络层解耦为深度可分离卷积层,并且DS-PWC在金字塔层增加基于层数的权重系数,从而使得网络结构在不损失精度的情况下大幅减少模型参数量。在训练过程中,使用图像及对象感知数据随机擦除(image and object-aware random erasing,I+ORE)等数据增强技术,进一步提升估计预测结果泛化能力。实验结果表明,在数据集测试DS-PWC模型,在保持质量的同时运行效率达到约58 fps(frame per second)。同时为了验证算法有效性,进行了模型结构和数据增强的消融实验。结果证明了DS-PWC模型的有效性。
