顾及视频地理映射的人群密度估计方法

针对复杂场景下人群计数及地图可视化问题,提出了一种顾及视频地理映射的人群密度估计方法。首先,通过迁移学习的方式,构建了适合于复杂场景的人群语义分割模型;将视频与GIS相结合,求解摄像机和人群场景的单应矩阵;基于人群语义分割结果及单应矩阵,可将人群多边形映射至二维地图。然后,设计等距分区和格网分区2种分区方案;根据人群语义分割结果,统计不同分区方案下各子区人群密度。最后,基于训练的各子区人群密度值,计算视域内人口总数;通过人群目标点符号均匀填充人群多边形,进行地图可视化。实验结果表明：(1)本文构建的人群语义分割模型,可实现大场景下人群的高精度分割,精度为94.11%;(2)将视频与GIS相结合,实现了人群的地图映射和地图可视化表达,达到了人群可定位、可量测和可空间分析的目标;(3)实现了监控视频人群的精准计数,将视域划分更多子区有利于提高计数精度。与基于密度图的人群密度估计方法相比,本文方法在高空、高密集度的大场景中具有优势,能有效解决人群难以精准计数及地图可视化的问题,可用于大型活动、车站、商场、运动场馆的人群监管。     

利用C_SURF配准的空基视频运动目标检测

目的 针对传统的车辆检测算法的性能易受低空移动平台影响造成相机自运动以及外界的干扰等问题,提出了一种基于改进的C_SURF彩色特征稳像和光流法向量相结合的方法来解决低空视频中的运动车辆检测问题。通过图像稳像消除了相机的自运动和外界干扰问题,提高了运动车辆的检测性能。实验结果显示,该方法不仅在检测车辆方面可以获得更好的检测性能,在复杂的背景环境下也能有效地检测运动车辆。     

错误多发异质网络环境下视频编码系统的总体设计

描述了针对错误多发异质的网络环境进行视频编码系统总体设计的全过程。分析了外部需求,提出了改进视频表现、增强容错能力和提供可分级编码能力的系统设计目标。在简要介绍H.263+协议的全部16个可选的高级编码模式的基础上,重点分析了设计中对系统是否采用某个高级编码模式进行取舍的理由。对视频数据流的打包、错误掩蔽和错误跟踪的机制等方面进行了讨论。实际应用表明,根据本文设计进行的系统实现完全达到了原定的系统设计目标。     

Long-term change in the structure of a Posidonia oceanica landscape and its reference for a monitoring plan

Abstract On the basis of a detailed cartographic survey carried out by Side Scan Sonar and a towed underwater video camera during 2005, and from a series of historical maps (1959, 1980, 1990), an extensive regression of Posidonia oceanica (L.) Delile beds was evidenced for a vast area of the central Tyrrhenian Sea (Latium coast, Italy). The total loss of P. oceanica surface was assessed through GIS estimate. In 1959, the Posidonia beds extended over 7290 ha, while in the 2005 survey they had regressed to 2899 ha, a loss of about 60% of their coverage. Also the seagrass lower limit showed a general depth decrease in time. Total seagrass coverage loss and lower limit regression were not uniform along the whole investigated areas and three main sub-areas have been identified with different degrees of regression somehow related with coastal potential human-mediated impacts. From different coverage estimates of the present survey and of the previous maps, minimum sampling areas were calculated through bootstrapping simulation routines from small sampling areas (Landscape Units) to reach the nearest estimate of the observed condition in the different periods.     

关系挖掘驱动的视频描述自动生成

视频的自动描述任务是计算机视觉领域的一个热点问题.视频描述语句的生成过程需要自然语言处理的知识,并且能够满足输入（视频帧序列）和输出（文本词序列）的长度可变.为此本文结合了最近机器翻译领域取得的进展,设计了基于编码-解码框架的双层LSTM模型.在实验过程中,本文基于构建深度学习框架时重要的表示学习思想,利用卷积神经网络（CNN）提取视频帧的特征向量作为序列转换模型的输入,并比较了不同特征提取方法下对双层LSTM视频描述模型的影响.实验结果表明,本文的模型具有学习序列知识并转化为文本表示的能力.     

Origin of the fumarolic fluids of Vulcano Island,Italy and implications for volcanic surveillance

Variations in D and ¹⁸O values with H₂O contents and outlet temperatures indicate that the fumaroles of La Fossa crater have discharged mixtures of magmatic water and marine hydrothermal water, since 1979. The contribution of meteoric water was low in the period 1979–1982 and very low afterwards. The ¹⁸O values of the marine-hydrothermal component of +5 to +7.2 are due to isotopic exchange with the ¹⁸O-rich silicates of the rocks under high-temperature and low-permeability conditions. The ¹⁸O value of the magmatic end-member is generally +3.5 to +4.3, although values as high as +5.5 to +6.5 were reached in the summer of 1988, when magma degassing appears to have extended into the core of the magma body. The D values of the end-member were close to -20, typical of andesitic waters. Both the isotopic values and chemical data strongly support a dry model, consisting of a central magmatic gas column and a surrounding hydrothermal envelope, in which marine hydrothermal brines move along limited fracture zones to undergo total evaporation on approaching the conduits of magmatic fluids. The vents at the eastern and western boundaries of the fumarolic field are fed by fluids whose pressure is governed by the coexistence of vapor, liquid and halite, giving rise to a high risk of phreato magmatic explosions, should magma penetrate into these wet environments. Most La Fossa eruptions were triggered by an initial hydrothermal blast and continued with a series of phreatomagmatic explosions. The fluids discharged by the Forgia Vecchia fumaroles are mixed with meteoric water, which is largely evaporated, although subordinate loss of condensed steam may be responsible for scrubbing most of the acidic gas species. The temperatures and pressures, and the risk of a sudden pressure increase, are low. A boiling hydrothermal aquifer at 230° C is present underneath the Baia di Levante beach. This area has a minor risk of hydrothermal explosions.     

一种基于道路监控视频图像质量评估的新方法

针对近些年图像质量评价的重要性,提出一种针对道路监控视频的图像质量评估方法。因为道路监控视频具有连贯性、图像稳定性强、单个监控视频场景单一等特点,本文采用一种结合多幅图像创建物理模型对比恢复天气影响的图像质量评估方法,通过多幅图像比较恢复道路视频图像清晰度,加入噪声亮度比的因子来增加图像的鲁棒性和准确性,分析该方法的优势和不足。     

YOLO v4框架下Multi-Patch多帧增量式交通视频目标检测

提升目标检测模型的泛化能力是计算机视觉领域的研究热点和关键难点。本文提出了一种Multi-Patch方法和多帧增量式预测策略,提升了不同场景下交通视频目标检测的稳健性,有效解决了目标尺度多变导致的视频中目标召回率低的问题。根据视频分辨率和目标尺寸,基于Multi-Patch方法自动将视频帧分割成最佳输入尺寸,使用YOLO v4神经网络并关联连续帧的上下文信息,采用增量式预测策略降低视频目标检测的漏检率,提升不同场景下视频目标的检测置信度得分和召回率。采集不同拍摄条件下的交通视频,验证该方法的有效性。试验结果表明,本文提出的目标检测方法召回率在80%以上,置信度平均得分在0.84以上。