CN102708182B - Rapid video concentration abstracting method - Google Patents

Rapid video concentration abstracting method Download PDF

Info

Publication number
CN102708182B
CN102708182B CN 201210142026 CN201210142026A CN102708182B CN 102708182 B CN102708182 B CN 102708182B CN 201210142026 CN201210142026 CN 201210142026 CN 201210142026 A CN201210142026 A CN 201210142026A CN 102708182 B CN102708182 B CN 102708182B
Authority
CN
Grant status
Grant
Patent type
Prior art keywords
video
target
concentrated
collision
method
Prior art date
Application number
CN 201210142026
Other languages
Chinese (zh)
Other versions
CN102708182A (en )
Inventor
尚凌辉
刘嘉
陈石平
张兆生
高勇
Original Assignee
浙江捷尚视觉科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Grant date

Links

Abstract

本发明涉及一种快速视频浓缩摘要方法,现有的视频浓缩技术对运动目标的检测率和跟踪率效果较差,不能有效浓缩视频长度。 The present invention relates to a method for fast video summary concentrating, poor concentration techniques to existing video moving target detection and tracking rate of the effect, the length of the video can not be effectively concentrated. 本发明采其特征在于由服务器端对预处理视频中的运动目标进行检测跟踪,根据视频的长度或视频中检测目标的数量进行判断,将视频切为多个浓缩段,对每个浓缩段内的目标轨迹进行碰撞检测和重排,之后记录浓缩段信息进入索引文件中;由客户端对存入服务器端内的索引文件进行分析,获取已处理的浓缩段,逐帧渲染浓缩段,形成视频序列,并对播放中的浓缩视频动态调整目标密度。 The present invention is adopted wherein detecting tracking moving objects in a video preprocessing by the server, be determined according to the number or length of video detection target video, video cut into a plurality of segments and concentrated, the concentrate for each segment collision detection target trajectory and rearrangements, and concentrated after recording into the index file, segment information; analysis of the index file stored in the server by the client, concentrated to obtain the processed segments, frame by frame rendering concentration section, forming a video sequence, and concentrated to a video playback dynamic adjustment target density. 本发明跟踪目标连续性好,轮廓区域完整,检测率高,误检率低,浓缩视频各时间点的目标密度基本一致。 The present invention is a good target tracking continuity complete outline region, detection rate, false detection rate is low, the target density for each time point was concentrated video consistent.

Description

一种快速视频浓缩摘要方法 A quick video summary concentration method

技术领域 FIELD

[0001] 本发明属于视频检索和视频摘要领域,尤其是一种快速视频浓缩摘要方法。 [0001] The present invention belongs to the field of video retrieval and digest video, video in particular, is a fast and concentrated digest method.

背景技术 Background technique

[0002] [I]用于视频索引和视频概要的方法和系统-200780050610.0 [0002] [I] A method and system for video indexing and video summary -200780050610.0

[0003] [2]用于产生视频概要的方法和系统-200680048754.8 [0003] [2] for generating a video summary of the method and system -200680048754.8

[0004] [3]基于时空融合的智能提取视频摘要方法-201110170308.7 [0004] [3] Video summary method of extracting time and space-based intelligence fusion -201110170308.7

[0005] [4]视频摘要系统-201020660533.X [0005] [4] The video summarization system -201020660533.X

[0006] [5]基于视频监控网络的视频自动浓缩方法-201110208090.X [0006] [5] The network-based video automatic video surveillance method of concentrating -201110208090.X

[0007] [6] PRITCH Y, RAV-ACHA A, PELEG S.Nonchronological video synopsisand indexing [J].1EEE Transactions on Pattern Analysis and MachineIntelligence, 2008,30(11): 1971-1984 [0007] [6] PRITCH Y, RAV-ACHA A, PELEG S.Nonchronological video synopsisand indexing [J] .1EEE Transactions on Pattern Analysis and MachineIntelligence, 2008,30 (11): 1971-1984

[0008] [7] Y.Pritchj S.Ratovitchj A.Hendelj and S.Pelegj ClusteredSynopsis of Surveillance Video, 6th IEEE Int.Conf.0n Advanced Video andSignal Based Surveillance (AVSSi 09), Genoa, Italy, Sept.2-4,2009 [0008] [7] Y.Pritchj S.Ratovitchj A.Hendelj and S.Pelegj ClusteredSynopsis of Surveillance Video, 6th IEEE Int.Conf.0n Advanced Video andSignal Based Surveillance (AVSSi 09), Genoa, Italy, Sept.2-4 2009

[0009] 随着视频监控的普及和视频监控技术的发展,每天都有海量的监控视频数据产生并被记录在设备上。 [0009] With the development and popularization of video surveillance video surveillance, it is mass produced and daily monitoring video data recorded on the device. 如何对这些海量的数据进行有效的浏览和分析已经成为该领域备受关注的一个问题。 How to effectively view and analyze these vast amounts of data has become a problem in this area of ​​concern. 通常人们只对视频中的某些目标(主要是运动目标)和内容感兴趣,希望能快速浏览视频一段长时间视频中出现的感兴趣内容。 Usually people only interested in some of the objectives and content of the video (mainly moving target), hoping to quickly browse video content of interest for a long time appearing in the video. 视频浓缩技术通过对视频内容分析,分割运动目标,并对它们的出现时间重排,使得在最短的时间里能向用户有效呈现所有的目标。 Video enrichment technology through video content analysis, segmentation of moving targets, and their time of occurrence of the rearrangement, so that in the shortest possible time to effectively present all of the target user.

[0010] [I] [2] [6] [7]中提出了基于运动目标分割、背景建模和碰撞检测的视频浓缩方案。 [0010] [I] [2] [6] [7] proposed object segmentation based on background modeling and collision detection video concentrated solutions. 该方案能获得比较理想的浓缩效果。 The program can obtain more ideal concentration effect. 但其中碰撞检测方案需要计算多个不同的碰撞代价项,计算量较大,不利于对高清视频的实时快速处理。 But collision detection scheme need to calculate the cost of a number of different collision term, large amount of calculation is not conducive to rapid real-time processing of HD video. [3]提出了一种基于时空融合的智能提取视频摘要方法。 [3] proposed a method for extracting a video summary based on intelligent integration of time and space. 该方法依靠帧差法获得目标轮廓,并根据矩形轮廓对目标进行跟踪。 The method relies on frame difference method to obtain the target profile and the target track in accordance with a rectangular profile. 对于跟踪到的目标序列,重排其在时间轴上出现的位置,以形成新的浓缩视频。 For tracking the target sequence, it appears rearrangement position on the time axis, to form a new video concentrated. 不同目标若有叠加,则进行透明化处理。 If different target superimposed, the clearing treatment. 该方法主要缺点在于:没有对目标的碰撞进行检测以获得较好的视觉效果;对于长轨迹未进行分段,因此若出现长时间徘徊目标,会影响浓缩视频的压缩长度。 The main drawback of this method is that: no collision object is detected to obtain a better visual effect; segmentation is not performed for long track, so if the target appears wandering time, can affect the length of the compressed video concentrated.

[0011 ] [4]提出了一种视频浓缩系统方案,该系统包含输入模块,分析模块,数据库模块和输出模块。 [0011] [4] proposes a video system solution was concentrated, the system comprises an input module, an analysis module, a database module and an output module. 输入模块获取视频后,送入分析模块进行目标检测和跟踪,将跟踪到的目标轮廓切取出来,保存在数据库中。 After obtaining the video input module, into the analysis target detection and tracking module, the tracking target contour cut out and saved in the database. 输出模块将不同帧出现的目标在同一个视频帧里呈现。 Output modules will target different frames appear showing the same video frame inside. 该系统未提及如何支持对部分处理完成的视频进行输出,也未给出如何避免目标碰撞,以及如何确保切取出来的目标长度适中,以获得较好的视觉效果。 The system does not mention how the support for some processed video output, nor show how to avoid the collision target, and to ensure moderate target length was cut out to obtain a better visual effect.

[0012] [5]提出一种基于视频监控网络的视频自动浓缩方法。 [0012] [5] proposed a method based on automatic video surveillance network video concentrated. 该方法处理来自两个有重叠区域摄像机所拍摄的视频源,提出基于图匹配和随机游走思想,对不同相机投影轨迹进行匹配,实现跨摄像机的目标跟踪。 The method for processing two video sources from overlapping area captured by the camera, and map matching is proposed based on random walk thoughts, different trajectory matching the projection camera, the camera cross-target tracking. 在跨摄像机匹配的全景图上进行视频的浓缩,可以获得大场景的浓缩视频。 Concentrated in the video camera panorama across the matching can be obtained concentrated large video scenes. 浓缩时,对重排定义了5个能量损耗,并定义了压缩率,用模拟退化来优化轨迹重排。 When concentrated rearrangement defined five energy losses, and defines a compression ratio, analog to optimize the trajectory rearrangement degradation. 该方法没有提及如何对轨迹进行分段,因此若出现长时间徘徊目标,会影响浓缩视频的压缩长度。 This method does not mention how trajectory segment, so if the target appears wandering time, can affect the length of the compressed video concentrated. 而能量项设计和模拟退火优化的计算量较大,不利于对高清视频的实时快速处理。 The calculation of the amount of energy the project design and simulated annealing optimization of large, is not conducive to rapid real-time processing of HD video.

发明内容 SUMMARY

[0013] 本发明针对现有技术存在的缺陷,提供一种快速视频浓缩摘要方法,有效提高运动目标的检测率和跟踪率,有效浓缩视频长度,实现有效密度控制。 [0013] The present invention is directed to drawbacks of the prior art, provides a quick summary video concentrating method, improve the detection and tracking of moving target rate, the effective length of the video and concentrated, to achieve an effective density control.

[0014] 为此,本发明采取如下技术方案:一种快速视频浓缩摘要方法,包括服务器端,其特征在于由服务器端对预处理视频中的运动目标进行检测跟踪,根据视频的长度或视频中检测目标的数量进行判断,将视频切为多个浓缩段,对每个浓缩段内的目标轨迹进行碰撞检测和重排,之后记录浓缩段信息进入索引文件中;还包括客户端,所述的客户端对存入服务器端内的索引文件进行分析,获取已处理的浓缩段,逐帧渲染浓缩段,形成视频序列,并对播放中的浓缩视频动态调整目标密度。 [0014] To this end, the present invention adopts the following technical solutions: A fast video summary concentrating method, including server, wherein the pre-track detecting moving objects in a video server, according to the length of a video or a video the number of the detection target judgment, cut into a plurality of video segments and concentrated, the collision of the target track in each segment detection and concentrated under rearrangements, then concentrated recorded into the index file, segment information; further comprises a client, according to the client for the index file stored in the server for analysis, obtaining concentrated segment has been processed, frame by frame rendering concentrated zone to form a video sequence, and dynamically adjusts video playback concentrated in the target density.

[0015] 所述的运动目标检测是采用自适应阈值的混合高斯方法对场景进行背景建模,结合帧间变化提取前景,在提取前景区域时利用多尺度信息对区域轮廓进行精细化,利用密度估计方法对随机区域进行定位,最后采用随机区域采样的方法对背景模型进行更新,有效检测出低对比度的目标;所述的目标跟踪是采用多假设的方法对多帧的运动检测区域进行关联,对目标轮廓进行预测,并基于边缘信息定位当前帧的轮廓位置,当目标发生分裂、碰撞、丢检时根据该位置产生假设,最后利用匈牙利算法给出最优的假设,并对历史假设进行裁剪,得到目标的跟踪轨迹。 [0015] The moving object detection method is the use of Gaussian mixture adaptive threshold background modeling a scene, a change in conjunction with an inter extracting the foreground, region contour fine multi-scale information when extracting the foreground area, using the density the method of estimating the random positioning area, the final area using random sampling of the background model update, effectively detect the low contrast target; the target tracking method is to use multi-hypothesis motion-detection area for associating a plurality of frames, the target profile to predict, based on the contour edge position information of the positioning of the current frame, when the target were to split the collision, generate hypotheses based on the detected position lost. Finally, Hungary algorithm gives the optimal assumptions, and assumptions cropping history , tracked the trajectory of the target.

[0016] 所述浓缩段的生成是采用以下方法实现:当一段视频累积时间长度超过Tmax或目标数量超过Nmax (与轨迹长度最大允许值Lmax和预设密度d正相关)时,则产生一个新的浓缩段。 [0016] The generated concentration section is achieved by the following method: when the accumulated time exceeds the length of a video or the target quantity exceeds Nmax of Tmax (track length and Lmax and a preset maximum allowable value related to the density d n), generating a new concentrated segment.

[0017] 通过运动目标检测和跟踪可以获得每个目标在视频中出现的轨迹信息,包括帧、区域、包围盒,根据切分目标轨迹各帧的包围盒位置,对视频中的长轨迹进行切分,确保每个轨迹长度大于Lmin小于Lmax。 [0017] The locus information may be obtained for each object appearing in the video by detecting and tracking moving objects, comprising a frame, an area, bounding box, according to the segmentation target position trajectory bounding box of each frame of video cut long trajectory minutes, to ensure that each track length greater than Lmin less than Lmax.

[0018] 判断目标间的碰撞,定义能量项对碰撞进行惩罚,再采用变步长迭代的贪婪法,确保每次迭代能量都有下降,且迭代收敛速度快,并用随机化法避免陷入局部最优解,完成目标碰撞的检测和重排。 [0018] Analyzing the collision between the target, the definition of the term collision energy penalty, and then using a variable step size greedy iterative method, each iteration to ensure that energy has decreased, and the iterative fast convergence speed, and to avoid local randomization method by most An optimum solution, the completed target collision detection and reordering.

[0019] 所述变步长迭代的贪婪法的优化步骤如下: [0019] The variable step iterative method greedy optimization step as follows:

[0020] a.初始化:设定初始迭代步长SI,最终迭代步长S2,其中S2〈S1。 [0020] a initialization: setting the initial iterative step SI, the final iteration step S2, where S2 <S1. 设定步长变化数ds,每个步长迭代次数N。 Set step size change in the number ds, each iteration step N. 设定当前步长S=Sl。 Setting the current step size S = Sl.

[0021] b.以当前步长S迭代N次: . [0021] b the current iteration step S N times:

[0022] a)计算当前碰撞代价EI。 [0022] a) calculating a current cost of the collision EI.

[0023] b)随机选择一条轨迹。 [0023] b) a randomly selected track.

[0024] c)以步长S为间隔,在浓缩段内所有可能位置,重新放置轨迹出现时间。 [0024] c) In step S intervals, all possible positions, repositioning time trajectory occurs within the concentration section.

[0025] d)计算所有位置中最小碰撞代价E2。 [0025] d) calculating the position of all the minimum cost of a collision E2. [0026] e)若碰撞代价E2〈El,则将轨迹放置在最小碰撞代价处。 [0026] e) If the collision consideration E2 <El, then placed in the tracks at the minimum cost of a collision.

[0027] c.设定S = S - ds。 . [0027] c set to S = S - ds. 若S>=S2,重复步骤2,否则结束。 If S> = S2, repeat step 2, else end.

[0028] 客户端在对视频逐帧渲染时,先根据当前帧ID在索引文件中查找该时刻对应的背景图,并查找所有该时刻出现的目标对应的区域像素值,将目标区域叠加到背景图上,若一个位置有多个目标出现,则该位置的像素值是多个目标像素值的平均。 [0028] When the client rendering the video frame by frame, to find the corresponding time index file in the background image based on the current frame ID, and find all the target region of the pixel values ​​corresponding to the occurrence time of the target area superimposed background figure, if a plurality of target present position, the position of the pixel value of target pixel value is the average of the plurality.

[0029] 背景图通过对多帧图像累积平均获得,先设定一个累积区间,若相邻累积区间的背景图发生变化超过门限Tl,则记录一张新的背景图;若变化超过门限T2(T2>T1),则标记为新的浓缩段。 [0029] FIG background image by a plurality of frames obtained cumulative average, a cumulative interval previously set, if the accumulated interval of adjacent background change exceeds a threshold Tl, the recording of a new background image; if the change exceeds the threshold T2 ( T2> T1), is marked for the new segment was concentrated.

[0030] 浓缩段生成时,使用默认的浓缩密度d,视频在客户端播放时,能根据希望的播放密度进行动态调整,当设置新的播放密度dn时,重新排列每个目标的出现时间,定义T。 [0030] When the concentration section to generate, using the default concentrated density, d, in the client video playback, playback can be dynamically adjusted based on the desired density, when setting a new playback density DN, rearranging time of occurrence of each object, the definition of T. 为目标的原始出现时间,则新的时间为Tn=I^cVdntj As the goal of the original appearance time, the new time Tn = I ^ cVdntj

[0031] 本发明具有以下优点:[0032] 1.跟踪目标连续性好,轮廓区域完整,检测率高,误检率低; [0031] The present invention has the following advantages: [0032] 1. The target tracking good continuity, complete contour region, detection rate, false positive rate is low;

[0033] 2.浓缩视频各时间点的目标密度基本一致; [0033] 2. The target density is concentrated at each time point are basically the same video;

[0034] 3.能对过长目标切分成小段播放,视频压缩效率高,播放视觉效果好; [0034] 3. The target can be cut into small pieces for long play, high-efficiency video compression, playback visual effects;

[0035] 4.碰撞检测和重排速度快; [0035] 4. The collision detection and fast rearrangement;

[0036] 5.对于需要长时间处理的视频可以支持边处理边播放。 [0036] The time required for video processing margin processing can be supported while playing.

[0037] 6.播放时根据需要可调整密度。 The need to adjust the density when [0037] 6. playback.

附图说明 BRIEF DESCRIPTION

[0038] 图1为本发明的流程图。 [0038] FIG. 1 is a flowchart of the present invention.

具体实施方式 detailed description

[0039] 下面通过实施例,对本发明的技术方案作进一步具体的说明。 [0039] The following Examples, the technical solution of the present invention will be further specifically described.

[0040] 如图1所示的快速视频浓缩摘要方法,包括服务器端和客户端,处理的具体步骤如下:服务器端先检测并分割该视频中出现的运动目标检测,采用自适应阈值的混合高斯方法对场景进行背景建模,结合帧间变化提取前景,在提取前景区域时利用多尺度信息对区域轮廓进行精细化。 Fast video summarization method illustrated concentrated [0040] As shown in FIG 1, comprises a server and client, specific process steps are as follows: the server to detect the moving object detection and segmentation appearing in the video, mixed Gaussian adaptive threshold A method for modeling the scene background, foreground extraction binding inter-change, the contour of the region to fine multi-scale information when extracting the foreground area. 结合纹理特征和帧间一致性变化,有效抑制了光照变化的干扰,并能有效检测出低对比度的目标;利用密度估计方法对树叶晃动、水波流动等随机区域进行定位;最后采用随机区域采样的方法对背景模型进行更新,增强了背景模型的鲁棒性。 Inter consistency and texture features binding changes, effectively suppress interference illumination changes, and can effectively detect low contrast target; density estimation method using the shaking leaves, random waves flow like region is positioned; final random sampling region methods background model update improves the robustness of the background model.

[0041] 采用多假设的方法对多帧的运动检测区域进行关联,对目标轮廓进行预测,并基于边缘信息定位当前帧的轮廓位置,当目标发生分裂、碰撞、丢检时根据该位置产生假设。 [0041] A method of multi-hypothesis of the motion detection area of ​​the multi-frame is associated, the target contour prediction, based on the edge information of the positioning contour position of the current frame, when the target split in collision, generate hypotheses based on the position lost subject . 最后利用匈牙利算法给出最优的假设,并对历史假设进行裁剪,得到目标的跟踪轨迹 Finally, Hungary algorithm gives the optimal assumptions, assumptions and historical crop, tracked the trajectory of the target

[0042] 输入视频一般有两种形式:视频文件和实时视频流。 [0042] There are two general forms of input video: video files and live video streaming. 对于视频文件,其时间长度和帧率是确定的,而实时视频流的时间长度不确定。 For video files, the time length and frame rate are determined, and the uncertain length of time live video streams. 视频中不同时间段的目标密度可能也是不同的,例如街道监控视频白天人流比较密集,而夜晚的人流比较稀疏。 Video target density for different time periods may also be different, such as street surveillance video intensive flow of people during the day, while at night the sparse crowd. 而且随着时间的推移,光照的变化或视野中物体的增减,会导致场景发生改变。 And over time, the increase or decrease or change in the object field of view of the light, will cause a scene change. 为确保浓缩视频在不同时间段有相似的密度,而且背景随时间推移发生变化,浓缩采用分段处理的方式。 To ensure that the video has a similar density is concentrated at different times, and the background change over time, the concentrated manner using the segmentation process. 当以下任意条件满足时,则产生一个新的浓缩段:累积视频时间长度超过Tmax ;目标数量超过Nmax(与轨迹长度最大允许值Lmax和预设密度d正相关)时;当场景的背景发生显著变化时,结合轨迹的重排可保证浓缩后各时间点目标密度基本一致。 When the number of the target exceeds Nmax of (the track length maximum allowable value Lmax and a preset density d a positive correlation);; cumulative length of video time exceeds Tmax when a background scene occurs significant: when any of the following conditions are satisfied, generate a new concentration section when changes, rearrangements bound track to ensure consistent and concentrated each time point target density.

[0043] 运动目标检测和跟踪可以获得每个目标在视频中出现的轨迹信息,包括帧(或绝对时间)、区域、包围盒。 [0043] The moving object detection and tracking track information can be obtained for each object appearing in the video, including a frame (or absolute time), area, bounding box. 对于视频中出现的长轨迹进行切分,以确保每个轨迹长度不超过最大允许值Lmax。 For long track appearing in the video be segmented in order to ensure that each track length does not exceed the maximum allowable value Lmax. 由于过短的目标轨迹在浏览时有闪烁感,为确保人眼视觉效果,切分后轨迹不能短于预定义的最短可视长度Lmin。 Because short of the target trajectory light flashes while browsing, to ensure that the human visual effects, after slicing trajectory can not be shorter than a predefined minimum visual length Lmin.

[0044] 根据切分目标轨迹各帧的包围盒位置,可以判断目标间的碰撞。 [0044] The position of the bounding box of each frame segmentation target locus, the collision between the target can be determined. 定义目标轨迹中第i条轨迹为Ti,第j条轨迹为Tj。 Define the target track is the i-th track Ti, as the j-th track Tj. 整个浓缩段的总碰撞代价与各轨迹之间发生重叠的总面积以及时间上错位的代价之和:E=Eo+Et The total cost of the entire overlap collision occurs between the concentration section and the total area of ​​the tracks and the time offset and Consideration: E = Eo + Et

[0045] 两条轨迹的重叠代价定义为用视频图像尺寸归一化的包围盒重叠区域面积: [0045] The cost of overlapping to two tracks define a normalized bounding box overlapping area of ​​video image size:

Figure CN102708182BD00071

[0046] 由于视频切分为浓缩段,因此时间上错位代价可以忽略不计,近似的有E~Eo0这种近似使得碰撞代价计算量大大减少。 [0046] Since the video is divided into cut sections was concentrated, and therefore the cost of the time offset may be negligible, with a similar approximation E ~ Eo0 cost calculation such collisions is greatly reduced. 为最小化碰撞总能量,采用变步长迭代的贪婪法。 In order to minimize the total energy of the collision, using the iteration variable step size greedy method. 优化步骤如下: Optimization steps are as follows:

[0047] 1.初始化:设定初始迭代步长SI,最终迭代步长S2,其中S2〈S1。 [0047] 1. Initialization: setting an initial iteration step SI, the final iteration step S2, where S2 <S1. 设定步长变化数ds,每个步长迭代次数N。 Set step size change in the number ds, each iteration step N. 设定当前步长S=Sl。 Setting the current step size S = Sl.

[0048] 2.以当前步长S迭代N次: [0048] 2. In the current iteration step S N times:

[0049] a.计算当前碰撞代价El。 [0049] a. Collision to calculate the current cost of El.

[0050] b.随机选择一条轨迹。 [0050] b. A randomly selected track.

[0051] c.以步长S为间隔,在浓缩段内所有可能位置,重新放置轨迹出现时间。 [0051] c. In step S intervals, all possible positions, repositioning time trajectory occurs within the concentration section.

[0052] d.计算所有位置中最小碰撞代价E2。 [0052] d. Calculating the minimum collision consideration all position E2.

[0053] e.若碰撞代价E2〈E1,则将轨迹放置在最小碰撞代价处。 [0053] e. If the cost of a collision E2 <E1, the track will be placed at the minimum cost of a collision.

[0054] 3.设定S = S - ds。 [0054] 3. The set S = S - ds. 若S>=S2,重复步骤2。 If S> = S2, repeat step 2. 否则结束。 Otherwise it ends.

[0055] 以上优化步骤可以确保迭代过程中能量逐步下降。 [0055] above optimization steps to ensure that the iterative process of energy gradually decline. 为加快计算速度,步骤2中计算最小碰撞代价E2可以替换为直接计算碰撞代价是否下降。 To speed up the calculation, step 2 to calculate the minimum cost of collision whether the collision E2 may alternatively decrease the cost for the direct calculation. 即每次迭代后,选中轨迹的放置位置为所有可能位置上与其他轨迹碰撞代价最小处。 That is, after each iteration, the placement of the selected track is all possible positions on collision with the other tracks at minimum cost. 由于对轨迹选择进行了随机化,可以避免能量优化的过程陷入局部最优。 Because of the track selection was randomized to avoid energy optimization process into local optimization. 变步长的寻优是由粗到精的搜索思想,比直接搜索最细的步长效率更高。 Variable step optimization search is coarse to fine thinking, higher than the finest step size direct search efficiency. 以上能量项的定义和寻优方式确保了快速的目标碰撞检测和重排。 Definition and optimization of the energy terms described above ensures rapid collision detection target and rearrangements.

[0056] 背景图通过对多帧图像累积平均获得。 [0056] FIG background by obtaining the average multi-frame image accumulation. 设定一个累积区间,若相邻累积区间的背景图发生变化超过门限Tl,则记录一张新的背景图;若变化超过门限T2(T2>T1),则标记为新的浓缩段。 A cumulative interval is set, if the accumulated interval of adjacent background change exceeds a threshold Tl, the recording of a new background image; if the change exceeds the threshold T2 (T2> T1), is marked for the new segment was concentrated.

[0057] 本发明还提出了一种C\S结构的视频浓缩系统架构,并支持边处理边播放浓缩视频。 [0057] The present invention further provides a C \ S video concentrator system architecture structures, and support while processing the video while playing concentrated. 服务端处理时,对视频进行动态分段,以实现并行处理。 When the server process, dynamic video segment, in order to achieve parallel processing. 对于每个并行单元,视频段按前述方法处理,并自适应切分为浓缩段。 For each parallel unit, a video processing section according to the aforementioned method, and the adaptive cut into sections and concentrated. 每个浓缩段存储的信息包括:在该视频段出现的目标轨迹;每个轨迹出现和消失的时间;该浓缩段累积的背景图;每张背景图起始和结束时间。 Each concentration section information stored comprising: a target track of the video segment appears; each track appear and disappear in time; the accumulation of background concentration section; each start and end times background.

[0058] 将视频中所有浓缩段的信息存储到一个索引文件中。 [0058] The information of all segments concentrated to a video index file is stored. 索引文件头记录已保存浓缩段个数,每个浓缩段对应原始视频的起始结束时间,以及在索引文件中保存的位置。 An index number of the saved header record concentration section, corresponding to the start and end of each segment and concentrated time of the original video, and the position is stored in the index file.

[0059] 客户端获取索引文件后,分析索引文件头,可获得已完成的浓缩段并进行播放,实现边处理边播放,提供较好的用户体验。 After the [0059] client obtains the index file, the index file header analysis obtained finished segments concentrate and play, while playing implement edge treatment, to provide a better user experience. 客户端对视频逐帧渲染时,先根据当前帧ID在索引文件中查找该时刻对应的背景图,并查找所有该时刻出现的目标对应的区域像素值,将目标区域叠加到背景图上。 Client rendering the video frame by frame, to find the ID of the current frame in the time index file corresponding to the background, and find the areas of the pixel values ​​of all the target time corresponding to occur, the superimposition target region to the background. 若一个位置有多个目标出现,则该位置的像素值是多个目标像素值的平均(即透明叠加)。 If a plurality of target present position, the position of the pixel value of the target pixel value is the average of a plurality of (i.e., a transparent overlay).

[0060] 前述浓缩段生成时,会使用默认的浓缩密度d。 When [0060] the generating segment was concentrated, the concentrate will use the default density d. 视频在客户端播放时,可能希望播放密度可以动态调整。 When playing video client, you may want to play density can be dynamically adjusted. 客户设置新的播放密度为dn时,重新排列每个目标的出现时间。 When the customer set up a new player density dn, rearrange the time of occurrence of each target. 定义T。 The definition of T. 为目标的原始出现时间。 As the goal of the original epoch. 则新的出现时间为Tn=T0*d/dn。 The appearance of new time Tn = T0 * d / dn.

[0061] 这种重排方式可以保证浓缩密度降低时,碰撞能量降低,提高视觉效果。 When [0061] This rearrangement concentrated manner can be guaranteed to reduce the density, the collision energy is reduced, improving the visual effect. 由于只需要直接计算每个目标轨迹的出现时间,而不用计算碰撞能量等,因此可以实现实时的密度调整。 Since only direct calculation appears each time the target trajectory, without computing the collision energy, etc., and therefore can achieve real-time density adjustment.

[0062] 需要特别指出的是,上述实施例的方式仅限于描述实施例,但本发明不止局限于上述方式,且本领域的技术人员据此可在不脱离本发明的范围内方便的进行修饰,因此本发明的范围应当包括本发明所揭示的原理和新特征的最大范围。 [0062] Of particular note, the above described embodiments are only examples, but the present invention is not only limited to the above embodiment, and those skilled in the art can easily be modified accordingly without departing from the scope of the invention Therefore the maximum range shall include the scope of the invention disclosed in the present invention the principles and novel features.

Claims (5)

  1. 1.一种快速视频浓缩摘要方法,其特征在于先由服务器端对预处理视频中的运动目标进行检测跟踪,根据视频的长度或视频中检测目标的数量进行判断,将视频切为多个浓缩段,对每个浓缩段内的目标轨迹进行碰撞检测和重排,之后记录浓缩段信息进入索引文件中;再由客户端对存入服务器端内的索引文件进行分析,获取已处理的浓缩段,逐帧渲染浓缩段,形成视频序列,并对播放中的浓缩视频动态调整目标密度; 所述的运动目标检测是采用自适应阈值的混合高斯方法对场景进行背景建模,结合帧间变化提取前景,在提取前景区域时利用多尺度信息对区域轮廓进行精细化,利用密度估计方法对随机区域进行定位,最后采用随机区域采样的方法对背景模型进行更新,有效检测出低对比度的目标;所述的目标跟踪是采用多假设的方法对多帧的运动检测 A fast video summary concentrating method, wherein the server is detected first by the pre-video tracking moving objects, the length is judged according to the number of video or video in the detection target, cut into a plurality of video and concentrated segment, a target track within each segment collision detection and concentrated under rearrangements, then concentrated recorded into the index file, segment information; then analyzing the index file stored in the server by the client, concentrated to obtain the processed segments , frame by frame rendering concentration section, form a video sequence, the video player and concentrated to dynamically adjust the target density; the moving object detection method is the use of Gaussian mixture adaptive threshold background modeling of a scene, a change in conjunction with inter-extraction foreground, region contour refinement utilized when extracting the foreground area of ​​the multi-scale information, the random region is positioned using the density estimation methods, and finally adopt the random region sampling the background model update, effectively detect the low contrast target; the said target tracking method is to use multi-hypothesis motion of a plurality of frames detected 域进行关联,对目标轮廓进行预测,并基于边缘信息定位当前帧的轮廓位置,当目标发生分裂、碰撞、丢检时根据该位置产生假设,最后利用匈牙利算法给出最优的假设,并对历史假设进行裁剪,得到目标的跟踪轨迹; 所述浓缩段的生成是采用以下方法实现:当一段视频累积时间长度超过Tmax或目标数量超过Nmax时,则产生一个新的浓缩段; 通过运动目标检测和跟踪可以获得每个目标在视频中出现的轨迹信息,对视频中的长轨迹进行切分,确保每个轨迹长度大于Lmin小于Lmax ; 通过运动目标检测和跟踪可以获得每个目标在视频中出现的轨迹信息,包括帧、区域、包围盒,根据切分目标轨迹各帧的包围盒位置,判断目标间的碰撞,定义能量项对碰撞进行惩罚,再采用变步长迭代的贪婪法,确保每次迭代能量都有下降,且迭代收敛速度快,并用随机化法 Associating domain, the predicted contour of the target, and position of the edge based on the contour information of the current frame is positioned, when the target split occurs, the collision is generated according to the position assumed lost object, Finally Hungarian algorithm gives the optimal hypothesis, and Suppose cropping history, track tracked object; generating segment of the concentrate is implemented by the following method: when the cumulative length of a video or exceeds Tmax exceeds the number Nmax of the target, generated a new segment was concentrated; moving target detection via and tracking track information can be obtained for each object appearing in the video, the video slicing the long trajectory, to ensure that each track is greater than the length Lmax of less than Lmin; each target can be obtained by detecting and tracking moving objects appear in the video trajectory information, including a frame, an area, bounding box, according to the bounding box of the position of each frame segmentation target track, judging the collision between the target, defines the energy term collision punishment, then using variable step iterative greedy method to ensure that each iterations have decreased energy, and the iterative fast convergence speed, and the method by randomizing 免陷入局部最优解,完成目标碰撞的检测和重排。 Avoid falling into local optimal solution, to complete the target of collision detection and rearrangements.
  2. 2.根据权利要求1所述的一种快速视频浓缩摘要方法,其特征在于所述变步长迭代的贪婪法的优化步骤如下: (0.初始化:设定初始迭代步长SI,最终迭代步长S2,其中S2〈S1。设定步长变化数ds,每个步长迭代次数N,设定当前步长S=Sl ; (2).以当前步长S迭代N次: a)计算当前碰撞代价El ; b)随机选择一条轨迹; c)以步长S为间隔,在浓缩段内所有可能位置,重新放置轨迹出现时间; d)计算所有位置中最小碰撞代价E2 ; e)若碰撞代价E2〈E1,则将轨迹放置在最小碰撞代价处; (3).设定S=S - ds,若S>=S2,重复步骤(2),否则结束。 The concentrate Fast video summarization method according to claim 1, wherein said variable step iterative method greedy optimization step as follows: (0. Initialization: setting an initial iteration step the SI, the final iteration step long S2, where S2 <S1 setting step changes the number of ds, each iteration step N, the current step size set S = Sl; (2) the current iteration step S N times:.. a) calculating a current collision consideration El; b) randomly selecting a track; c) to step S intervals, all possible positions, repositioning track time of occurrence within said concentrating zone; D) is calculated for all positions in a minimum impact the cost E2; e) if a collision consideration E2 <E1, the track will be placed at a minimum cost of a collision; (3) set to S = S - ds, if S.> = S2, repeat steps (2) otherwise end.
  3. 3.根据权利要求1或2所述的一种快速视频浓缩摘要方法,其特征在于客户端在对视频逐帧渲染时,先根据当前帧ID在索引文件中查找该时刻对应的背景图,并查找所有该时刻出现的目标对应的区域像素值,将目标区域叠加到背景图上,若一个位置有多个目标出现,则该位置的像素值是多个目标像素值的平均。 3. A fast video 1 or the method as claimed in claim concentrated Abstract, wherein when the client rendering the video frame by frame, to find the corresponding time index file in the background image based on the current frame ID, and Find the target region of the pixel values ​​of all time corresponding to occur, the superimposition target region to the background, if a plurality of target present position, the position of the pixel value of target pixel value is the average of the plurality.
  4. 4.根据权利要求3所述的一种快速视频浓缩摘要方法,其特征在于背景图通过对多帧图像累积平均获得,先设定一个累积区间,若相邻累积区间的背景图发生变化超过门限Tl,则记录一张新的背景图;若变化超过门限T2,且Τ2>t1,则标记为新的浓缩段。 The concentrate Fast video summarization method according to claim 3, characterized in that the background image by a cumulative average of a plurality of frames is obtained, a cumulative interval previously set, if the accumulated variation occurs adjacent background section exceeds a threshold Tl, then recording a new background image; if the change exceeds the threshold T2, and Τ2> t1, is marked for the new segment was concentrated.
  5. 5.根据权利要求4所述的一种快速视频浓缩摘要方法,其特征在于浓缩段生成时,使用默认的浓缩密度d,视频在客户端播放时,能根据希望的播放密度进行动态调整,当设置新的播放密度dn时,重新排列每个目标的出现时间,定义T0为目标的原始出现时间,则新的时间为Tn=T0*d/dn。 4 A fast video summary of the concentration method according to claim, characterized in that the concentration section generation, the default concentrated density, d, in the client video playback, playback can be dynamically adjusted based on the density desired, when when setting a new playback density DN, rearranged the time of occurrence of each target, T0 is defined as the time of occurrence of the original target, then the new time is Tn = T0 * d / dn.
CN 201210142026 2012-05-08 2012-05-08 Rapid video concentration abstracting method CN102708182B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201210142026 CN102708182B (en) 2012-05-08 2012-05-08 Rapid video concentration abstracting method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201210142026 CN102708182B (en) 2012-05-08 2012-05-08 Rapid video concentration abstracting method

Publications (2)

Publication Number Publication Date
CN102708182A true CN102708182A (en) 2012-10-03
CN102708182B true CN102708182B (en) 2014-07-02

Family

ID=46900948

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201210142026 CN102708182B (en) 2012-05-08 2012-05-08 Rapid video concentration abstracting method

Country Status (1)

Country Link
CN (1) CN102708182B (en)

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102930061B (en) * 2012-11-28 2016-01-06 安徽水天信息科技有限公司 A video summarization method based on the motion target detection
CN103079117B (en) * 2012-12-30 2016-05-25 信帧电子技术(北京)有限公司 Video digest generation method and a video summary generation means
CN103096185B (en) * 2012-12-30 2016-04-20 信帧电子技术(北京)有限公司 A video digest generation method and apparatus
CN103226586B (en) * 2013-04-10 2016-06-22 中国科学院自动化研究所 Based on the energy distribution of the optimal strategy video digest method
CN104284057B (en) * 2013-07-05 2016-08-10 浙江大华技术股份有限公司 A video processing method and apparatus
CN103345764B (en) * 2013-07-12 2016-02-10 西安电子科技大学 Double target content based on the video monitor digest generation method
CN104301699B (en) * 2013-07-16 2016-04-06 浙江大华技术股份有限公司 An image processing method and apparatus
CN103455625B (en) * 2013-09-18 2016-07-06 武汉烽火众智数字技术有限责任公司 A method for objective video summary rapid rearrangement method
CN104618681A (en) * 2013-11-01 2015-05-13 南京中兴力维软件有限公司 Method and device for multi-channel video condensation
CN103617234B (en) * 2013-11-26 2017-10-24 公安部第三研究所 Concentrating the active video apparatus and method
CN103686095B (en) * 2014-01-02 2017-05-17 中安消技术有限公司 A video system and method for concentrating
CN103793477B (en) * 2014-01-10 2017-02-08 同观科技(深圳)有限公司 A system and method for generating a video summary
CN103778237B (en) * 2014-01-27 2017-02-15 北京邮电大学 A video summary generation, based on space-time restructuring activity events
CN103957472B (en) * 2014-04-10 2017-01-18 华中科技大学 One kind of guarantee optimal timing of events reorganization digest video generation method and system based on
CN104284158B (en) * 2014-10-23 2018-09-14 南京信必达智能技术有限公司 Used in event-oriented method of intelligent monitoring camera
CN105007433A (en) * 2015-06-03 2015-10-28 南京邮电大学 Target-based moving object arrangement method enabling energy constraint minimization
CN105262932B (en) * 2015-10-20 2018-06-29 深圳市华尊科技股份有限公司 Kinds of video processing method and terminal
CN105357594B (en) * 2015-11-19 2018-08-31 南京云创大数据科技股份有限公司 Abstract Generation Based on massive video and video concentrated cluster of algorithms h264

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101262568A (en) 2008-04-21 2008-09-10 中国科学院计算技术研究所 A method and system for generating video outline
CN101689394A (en) 2007-02-01 2010-03-31 耶路撒冷希伯来大学伊森姆研究发展有限公司 Method and system for video indexing and video synopsis
CN102375816A (en) 2010-08-10 2012-03-14 中国科学院自动化研究所 Online video concentration device, system and method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7881505B2 (en) * 2006-09-29 2011-02-01 Pittsburgh Pattern Recognition, Inc. Video retrieval system for human face content

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101689394A (en) 2007-02-01 2010-03-31 耶路撒冷希伯来大学伊森姆研究发展有限公司 Method and system for video indexing and video synopsis
CN101262568A (en) 2008-04-21 2008-09-10 中国科学院计算技术研究所 A method and system for generating video outline
CN102375816A (en) 2010-08-10 2012-03-14 中国科学院自动化研究所 Online video concentration device, system and method

Also Published As

Publication number Publication date Type
CN102708182A (en) 2012-10-03 application

Similar Documents

Publication Publication Date Title
Li et al. Video object cut and paste
Pan et al. Automatic detection of replay segments in broadcast sports programs by detection of logos in scene transitions
US20080019661A1 (en) Producing output video from multiple media sources including multiple video sources
US20040194129A1 (en) Method and apparatus for intelligent and automatic sensor control using multimedia database system
Ngo et al. Automatic video summarization by graph modeling
US20100104261A1 (en) Brief and high-interest video summary generation
US7424175B2 (en) Video segmentation using statistical pixel modeling
US20050182503A1 (en) System and method for the automatic and semi-automatic media editing
Shih et al. Exemplar-based video inpainting without ghost shadow artifacts by maintaining temporal continuity
US8311277B2 (en) Method and system for video indexing and video synopsis
CN101616264A (en) Method and system for cataloging news video
CN101287142A (en) Method for converting flat video to tridimensional video based on bidirectional tracing and characteristic points correction
CN101719144A (en) Method for segmenting and indexing scenes by combining captions and video image information
Pritch et al. Webcam synopsis: Peeking around the world
US20040264939A1 (en) Content-based dynamic photo-to-video methods and apparatuses
Li et al. A general framework for sports video summarization with its application to soccer
CN102324019A (en) Method and system for automatically extracting gesture candidate region in video sequence
Antić et al. K-means based segmentation for real-time zenithal people counting
CN103236160A (en) Road network traffic condition monitoring system based on video image processing technology
Mohamed et al. Background modelling and background subtraction performance for object detection
Silva et al. An extended center-symmetric local binary pattern for background modeling and subtraction in videos
Poignant et al. From text detection in videos to person identification
CN101236657A (en) Single movement target track tracking and recording method
Wang et al. Semantic event detection using conditional random fields
CN101957997A (en) Regional average value kernel density estimation-based moving target detecting method in dynamic scene

Legal Events

Date Code Title Description
C06 Publication
C10 Request of examination as to substance
C53 Change or modification
C14 Granted
C56 Change in the name or address of the patentee

Owner name: ZHEJIANG ICARE VISION TECHNOLOGY CO., LTD.

Free format text: FORMER NAME: HANGZHOU ICARE VISION TECHNOLOGY CO., LTD.