CN101872414B - People flow rate statistical method and system capable of removing false targets - Google Patents

People flow rate statistical method and system capable of removing false targets Download PDF

Info

Publication number
CN101872414B
CN101872414B CN 201010118136 CN201010118136A CN101872414B CN 101872414 B CN101872414 B CN 101872414B CN 201010118136 CN201010118136 CN 201010118136 CN 201010118136 A CN201010118136 A CN 201010118136A CN 101872414 B CN101872414 B CN 101872414B
Authority
CN
China
Prior art keywords
head
classifier
target
detection
module
Prior art date
Application number
CN 201010118136
Other languages
Chinese (zh)
Other versions
CN101872414A (en
Inventor
任烨
呼志刚
朱勇
胡扬忠
蔡巍巍
贾永华
邬伟琪
Original Assignee
杭州海康威视软件有限公司
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
Application filed by 杭州海康威视软件有限公司 filed Critical 杭州海康威视软件有限公司
Priority to CN 201010118136 priority Critical patent/CN101872414B/en
Publication of CN101872414A publication Critical patent/CN101872414A/en
Application granted granted Critical
Publication of CN101872414B publication Critical patent/CN101872414B/en

Links

Abstract

The invention discloses a people flow rate statistical method and a system capable of removing false targets, wherein the method comprises the following steps: adopting a classifier to carry out human head detection on a current image, and determining all human heads in the current image; tracking all the determined human heads, and forming a motion trajectory of human head targets; carrying out smoothness analysis on the motion trajectory of the human head targets; and carrying out people flow rate counting in the direction of the motion trajectory of the human head targets after the analysis. Therefore, the method and the system can remove the false targets through the smoothness analysis of the trajectory of the human head targets, and further improve the detection accuracy.

Description

可去除虚假目标的人流量统计的方法及系统 Removable false targets people who traffic statistics methods and systems

技术领域 FIELD

[0001] 本发明涉及视频监控及图像处理与分析技术领域,尤其涉及一种可去除虚假目标的人流量统计的方法及系统。 [0001] The present invention relates to video surveillance and analysis of image processing technology and, particularly, to a system and method of removing the false targets human traffic statistics.

背景技术 Background technique

[0002] 随着社会的不断进步,视频监控系统的应用范围越来越广。 [0002] As society progresses, the scope of application of video surveillance system more widely. 在超市、商场、体育馆以及机场车站等场所的出入口常安装有监控摄像机,以便保安人员和管理者对这些场所的出入口进行监控。 Entrances and exits in supermarkets, shopping malls, stadiums and airport stations and other places often equipped with surveillance cameras, security personnel and managers in order to monitor the entrances and exits of these places. 另一方面,超市、商场、体育馆以及机场车站等场所进出的人流量对于上述场所的经营者或管理者来说有着重要的意义,其中,人流量是指按一定方向流动的人数, 本文中特指按进入/离开两个方向流动的人数。 On the other hand, supermarkets, shopping malls, stadiums and airport stations and other places in and out of these places of human traffic for the operators or managers of great significance, which refers to the flow of people in the special number of people flowing in a certain direction, this article refers to press to enter / leave the number of people flowing in both directions.

[0003] 现有的视频监控中,人流量统计主要是通过监控人员人工清点来实现。 [0003] existing video monitoring, mainly through human traffic statistics monitoring personnel manual inventory to achieve. 这种人工统计人流量的方法在监控时间短、人流量稀疏的情况下比较可靠,但由于人眼生物特性的限制,当监控时间较长,人流量密集时,统计的准确性将大大下降,而且人工统计的方式需要耗费大量的人力成本。 This method of artificial statistics to monitor the flow of people in a short time, under the sparse flow of people is more reliable, but due to the biological properties of the human eye, when monitoring a long time, the flow of people-intensive, the statistical accuracy will be greatly reduced, and statistical manual way requires a lot of labor costs. 基于视频分析的人流量统计方法可以实现人流量的自动统计,解决人工统计带来的各种问题。 It can automatically count people who traffic flow based on video analysis of statistical methods to solve the problems caused by artificial statistics. 目前,基于视频分析的流量统计方法主要有三类: Currently, video-based traffic statistical methods of analysis are mainly three types:

[0004] 一是基于特征点跟踪的方法,该方法首先跟踪一些运动的特征点,然后对特征点的轨迹进行聚类分析,从而得到人流量信息;基于特征点跟踪的方法需要跟踪一些运动的特征点,然后对特征点的轨迹进行聚类分析,从而得到人流量信息,该方法的缺点是特征点本身难以稳定地跟踪,计数精度较差。 [0004] One method of tracking the feature points based on the method first tracking feature points of some movement, and the trajectory of the feature point cluster analysis to obtain the information flow of people; feature point tracking method based on motion needs to keep track of some feature point and the feature point trajectory cluster analysis to obtain information on the flow of people, disadvantage of this process is a feature point it is hard to stably track, count accuracy is poor.

[0005] 二是基于人体分割和跟踪的方法,该方法首先需要提取出运动目标块,然后对运动目标块进行分割得到单个人体目标,最后跟踪各个人体目标实现人流量的统计;基于人体分割和跟踪的方法首先需要提取处运动目标块,然后对运动目标块进行分割得到单个人体目标,最后跟踪得到各个人体的轨迹,从而实现人流量的统计。 [0005] Second method of human segmentation and tracking based on the method first need to extract the moving target block, and then moving target block obtained by dividing a single human object, the last track individual human target achieved human traffic statistics; based on body segmentation and the first way to track moving objects need to be extracted at the block, and then moving target blocks obtained by dividing a single human object, and finally get the various tracking the trajectory of the body in order to achieve human traffic statistics. 该方法的缺点是当人体存在遮挡时,人体分割的准确性难以得到保证,影响统计精度。 The disadvantage of this method is that when there is a human shield, the body segmentation accuracy is difficult to be assured that affect the accuracy of the statistics.

[0006] 三是基于人头或头肩检测和跟踪的方法,该方法在视频中检测人头或头肩,通过对人头或头肩的跟踪进行人流量的统计。 [0006] Third head or head and shoulder detection and tracking method based on the method of detecting head or head and shoulders in the video, carried by the flow of people tracking head or head and shoulders of statistics. 基于人头检测和跟踪的方法是在视频中检测人头,通过对人头的跟踪进行人流量的统计,当摄像机角度合适时,人头出现遮挡的情况较少,因此基于人头检测的方法较前两种方法准确性有所提高,目前有公司提出了基于人头检测统计人数的方法,例如北京中星微电子在申请号200910076256. X的专利文件所提到的方法中,首先提取运动前景,然后采用haar特征训练两个串行的分类器在前景中搜索预定尺寸的人头,实现人头检测,其中,haar特征,是一种矩形特征,通过改变矩形的尺寸和组合方式可以描述目标的形状和灰度信息。 The method of detection and tracking head is detected based on the head in a video, statistics on the traffic carried by the person tracking the head, when appropriate camera angle, less frequently head sheltered, and therefore head detecting method based on the previous two methods accuracy has increased, the company currently has a method detection statistics based on the number of heads, such as Beijing Vimicro method patent application No. 200910076256. X of the documents mentioned in the first extracted moving foreground and then using haar feature two training classification serial search in the foreground of the head of a predetermined size, to achieve the detection head, wherein, Haar features, a rectangular feature to be described the shape and the gradation information of the object by changing the size and combination of rectangles. 该方法仅通过运动估计确定人头平移矢量速度, 从而统计出人头数量,这种方式准确度不高,对于虚假目标不易识别,导致人头统计不准确。 This method is only determined by motion estimation head translation vector velocity, so that the statistics of the number of the head, is not high accuracy in this way, false targets for easy recognition, leading to inaccurate statistics head.

发明内容[0007] 有鉴于此,本发明提供一种可去除虚假目标的人流量统计的方法及系统,以解决现有人流量统计方案统计不准确的问题。 METHOD AND SYSTEM FOR SUMMARY [0007] Accordingly, the present invention provides a removable false targets people who traffic statistics, in order to solve existing human traffic statistics program is not accurate.

[0008] 为此,本发明实施例采用如下技术方案: [0008] To this end, embodiments of the present invention employs the following technical solutions:

[0009] 一种可去除虚假目标的人流量统计的方法,包括:对图像中的检测区域进行场景标定,以获取所述检测区域内人头目标尺寸变化范围,从而根据所述人头目标尺寸变化范围将所述检测区域划分为若干个子区域;在所述子区域内采用分类器对当前图像进行人头检测,确定当前图像中的各人头;对确定出的各人头进行跟踪,形成人头目标运动轨迹;对人头目标运动轨迹进行平滑度分析;根据分析后的人头目标运动轨迹方向进行人流量计数。 [0009] A removable false targets human traffic statistics, comprising: a detection region in the image a scene calibration to obtain the target size of the head changes within the range of the detection area, so that the size range according to the target head the detection area is divided into several sub-regions; head current image detected by the classifier within the sub-region, determining current image of each head; head for each of the determined track, the head is formed target trajectory; head trajectory to the target smoothing analysis; for the number of people meter head according to the direction of the target trajectory after the analysis.

[0010] 所述对人头目标运动轨迹进行平滑度分析包括:确定人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹。 [0010] The smoothing analysis target trajectory head comprising: determining a target smoothness head trajectory, determines the smoothness meets a threshold, if yes, to retain the head target trajectory, otherwise, discarding the head moving target trajectory.

[0011] 在采用分类器对当前图像进行人头检测之后、确定当前图像中的各人头之前,还包括:对分类器检测到的人头进行边缘特征细筛选处理。 Before [0011] After the classifier using the current image detecting head, each head is determined in the current image, further comprising: a classifier for detecting the edge of the head of Fine filtering process.

[0012] 所述对分类器检测到的人头进行边缘特征细筛选处理包括:计算所述分类器判断为人头目标的矩形内边缘特征与预置的上半椭圆弧的拟合度,如果拟合度大于阈值,则将该矩形确定为人头,否则将该矩形从目标列表中去除。 [0012] the detected edge classifier head of Fine filtering process comprising: calculating the classifier is determined to fit within the rectangular head with a preset target edge feature in the upper half elliptic arc, if the fitting is greater than the threshold value, it is determined that the head is rectangular, rectangular or removed from the target list.

[0013] 所述对图像中的检测区域进行场景标定,以获取所述检测区域内人头目标尺寸变化范围,从而根据所述人头目标尺寸变化范围将所述检测区域划分为若干个子区域包括: 选择标定框;计算场景深度变化系数;计算检测区域内人头目标尺寸变化范围;根据人头目标尺寸变化范围将检测区域划分为若干个子区域。 [0013] The detection performed in the image areas of the scene calibration to obtain a certain head size range within the detection area, so that the head according to the target size range detection area is divided into several sub-areas comprising: selecting calibration block; coefficient of variation of the depth of a scene is calculated; calculating head size variation range detection target area; head according to the target size range detection area is divided into several sub-regions.

[0014] 所述分类器为并联的多类分类器。 [0014] The classifier is a parallel multi-class classifier.

[0015] 所述多类分类器对图像进行人头检测包括:设置各类分类器的检测顺序,按照检测顺序依次采用各个分类器对当前图像进行人头检测,直到确定出人头,其中,所述并联的多类分类器由至少两类分类器并联而成。 [0015] The multi-class classifier head image detection comprising: setting a detection order of the various types of classification, according to the detection order of each classifier using the current image detecting head, until it is determined to succeed, wherein the parallel the multi-class classifier from at least two parallel classifier.

[0016] 所述并联的多类分类器由深色头发通用分类器、浅色头发分类器、帽子分类器和扩展分类器中的任意两种或多种并联而成。 [0016] The parallel multi-class classifier classifier by a general dark hair, light hair, or any two classifier, and a classifier caps extended more classifiers together in parallel.

[0017] 一种可去除虚假目标的人流量统计的系统,包括:场景标定模块,用于对图像中的检测区域进行场景标定,以获取所述检测区域内人头目标尺寸变化范围,从而根据所述人头目标尺寸变化范围将所述检测区域划分为若干个子区域;人头检测模块,用于在所述子区域内采用分类器对当前图像进行人头检测,确定当前图像中的各人头;人头目标跟踪模块,用于对确定出的各人头进行跟踪,形成人头目标运动轨迹;人头目标运动轨迹分析模块,用于计算人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹;人流量计数模块,用于在分析后的人头目标运动轨迹方向进行人流量计数。 [0017] A removable false targets human traffic statistics system, comprising: a scene calibration module configured to image a scene calibration detection region, the detection region to obtain the target size range head, so that in accordance with the said head to the target size range detection area is divided into several sub-regions; head detection module configured to poll the current image detected by the classifier within subregions, each head is determined in the current image; target tracking head means for determining for each out of the head tracking, the target trajectory forming head; head target trajectory analysis module for smoothness calculating the target trajectory of the head, the flatness is determined whether the threshold is met, and if so, the reserved head trajectory target, otherwise, discarding the head target trajectory; people meter module number for the number of people meter in the direction of head motion trajectory after the target analysis.

[0018] 所述分类器为并联的多类分类器;所述人头检测模块包括粗检测子模块和细筛选子模块,所述粗检测子模块用于设置各类分类器的检测顺序,按照检测顺序依次采用各个分类器对当前图像进行人头检测,直到确定出人头,其中,所述并联的多类分类器由至少两类分类器并联而成;细筛选子模块,用于对并联的多类分类器检测到的人头进行边缘特征细筛选处理。 [0018] The classifier is a parallel multi-class classifier; detecting the head module comprises coarse and fine screening detection sub-module sub-module, sub-module for detecting the coarse detection order is provided various types of classifier, according to the detection order of each classifier using the current image detecting head, until it is determined to succeed, wherein the parallel multi-class classifier from at least two parallel classifier from; fine filter sub-module, based on a plurality of parallel classifier detected head edge of fine filtering process.

[0019] 所述并联的多类分类器由深色头发通用分类器、浅色头发分类器、帽子分类器和扩展分类器中的任意两种或多种并联而成。 [0019] The parallel multi-class classifier classifier by a general dark hair, light hair, or any two classifier, and a classifier caps extended more classifiers together in parallel.

[0020] 可见,本发明通过对人头目标轨迹的平滑度分析可以去除虚假目标,可进一步提高检测准确率。 [0020] visible, by the present invention, the smoothness of the analysis target track head false targets may be removed, can further improve the detection accuracy. 进一步,本发明将多个分类器并联使用,能同时检测深色头发、浅色头发以及各种颜色帽子等多类人头目标,确保统计更加全面。 Further, the present invention will be classifiers in parallel using multiple, simultaneous detection of dark hair, light hair colors and a variety of other types of hats poll objectives to ensure a more comprehensive statistics. 进一步,本发明还设置了一个扩展分类器,可以根据特殊环境的应用,采集样本训练,检测指定颜色或帽子的人头,比如工厂或仓库的工作帽等。 Further, the present invention is also provided with an extension classifier, depending on the application of the special circumstances, the training sample collection, the detection head or hat designated color, such as a factory or warehouse work caps. 进一步,在多个并联的分类器作为人头粗检测的基础上,再利用边缘特征对粗检测结果进行细筛选,最后得到真正的人头目标,使得检测更加准确。 Further, a plurality of parallel free base as a crude head detection, reuse edge feature detection result of the fine rough screening to obtain the final true target head, so that more accurate detection. 另外,本发明在检测前通过场景标定自动选择检测窗口的尺寸,使本发明能自适应各种摄像机角度,拓宽了应用范围。 Further, the present invention is calibrated prior to detection of the detection window size is automatically selected by the scene, of the present invention can adaptively various camera angles, broadens the scope of application.

附图说明 BRIEF DESCRIPTION

[0021] 图I为本发明一实施例人流量统计的方法流程图; [0021] The flowchart of FIG. I human traffic statistics method of the present embodiment of the invention;

[0022] 图2为本发明另一实施例人流量统计的方法流程图; [0022] FIG. 2 is a method flowchart of cases of human traffic statistics further embodiment of the invention;

[0023] 图3为本发明较优实施例场景标定流程图; [0023] FIG 3 a flow diagram of the preferred embodiment of the present invention, the calibration scenarios;

[0024]图4为本发明较优实施例人头检测模块结构框图; [0024] Figure 4 a block diagram of head detector module according to the present invention Jiaoyou embodiment;

[0025] 图5为本发明较优实施例各类分类器级联分类过程示意图; [0025] Figure 5 during various types of classifier cascade classifier Jiaoyou schematic embodiment of the invention;

[0026] 图6为本发明较优实施例粒子滤波跟踪的流程图; [0026] Example 6 is a flowchart Jiaoyou particle filter embodiment of the present invention;

[0027] 图7为本发明较优实施例运动轨迹平滑度分析流程图; [0027] FIG. 7 embodiment Jiaoyou flowchart trajectory analysis smoothness of the present invention;

[0028] 图8为本发明人流量统计的系统结构示意图。 [0028] FIG. 8 is a schematic structural diagram of a system inventors traffic statistics.

具体实施方式 Detailed ways

[0029] 本发明提出一种可去除虚假目标的人流量统计的方法,请参见图1,为本发明一实施例流程图,包括: [0029] The present invention provides a removable false targets human traffic statistical methods, see Figure 1, a flowchart of the present embodiment of the invention, comprising:

[0030] SlOO :采用分类器对当前图像进行人头检测,确定当前图像中的各人头; [0030] SlOO: classifier using the current image detecting head, each head is determined in the current image;

[0031] SlOl :对确定出的各人头进行跟踪,形成人头目标运动轨迹; [0031] SlOl: for each head the determined track, the head is formed target trajectory;

[0032] S102 :对人头目标运动轨迹进行平滑度分析; [0032] S102: the target trajectory of the smoothing head analysis;

[0033] S103 :根据分析后的人头目标运动轨迹方向进行人流量计数。 [0033] S103: the number of people for the flowmeter head according to the direction of the target trajectory after the analysis.

[0034] 其中,对人头目标运动轨迹进行平滑度分析的过程为:确定人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹。 [0034] wherein the process head target trajectory smoothing analysis for: determining a head target trajectory smoothness, determines the smoothness meets a threshold, if yes, to retain the head target trajectory, otherwise, discarding the head target trajectory.

[0035] 可见,本发明检测通过对人头目标轨迹的平滑度分析可以去除虚假目标,可进一步提高检测准确率。 [0035] seen that the present invention is detected by the smoothness of the analysis target track head false targets may be removed, can further improve the detection accuracy.

[0036] 为了进一步提高人流量统计的准确性,在图I所示的方案基础上,可进一步进行优化,包括,场景标定、采用并联的多类分类器进行粗检测、对粗检测结果进行边缘特征细筛选等,请参见图2,为本发明另一实施例流程图,包括: [0036] In order to further improve the accuracy of traffic statistics of people, as shown in Scheme I on the basis of FIG., May be further optimized, including, scene calibration using parallel multi-class classifier rough detection, the detection result of the rough edge wherein the fine screening, see Figure 2, a flowchart of another embodiment embodiment, the present invention comprises:

[0037] S201 :场景标定; [0037] S201: Scene calibration;

[0038] 具体地,场景标定是指对图像中的检测区域进行场景标定,从而将检测区域划分为若干个子区域。 [0038] Specifically, the scene calibration means for detecting an image scene calibration region so that the detection area is divided into several sub-regions.

[0039] S202 :人头检测; [0039] S202: detection head;

[0040] 人头检测进一步包括并联分类器粗检测以及边缘特征细筛选两个步骤,从而确定当前图像中的各人头。 [0040] The detecting head further comprises a parallel and an edge detecting rough classifier wherein two fine screening step to determine the current image of each head.

[0041] S203 :人头目标跟踪; [0041] S203: the head tracking;

[0042] 通过对确定出的各人头进行跟踪,形成人头目标运动轨迹。 [0042] determined by the head of each track, the head is formed target trajectory.

[0043] S204 :对人头目标运动轨迹进行平滑度分析; [0043] S204: the target trajectory of the smoothing head analysis;

[0044] 具体地,对人头目标运动轨迹进行平滑度分析包括:确定人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹。 [0044] Specifically, a head target trajectories smoothness analysis comprises: determining a smoothness head target trajectory, determines the smoothness meets a threshold, if yes, to retain the head target trajectory, otherwise, discarding the head target trajectory.

[0045] S205 :人流量统计:通过人头目标运动轨迹方向对人流量进行计数。 [0045] S205: human traffic statistics: count the flow of people through the direction of head movement trajectory targets.

[0046] 需要说明的是,上述场景标定、对并联分类器粗检测的人头进行边缘特征细筛选, 以及对人头目标运动轨迹进行分析的改进点可结合应用,也可单独使用。 [0046] Incidentally, the above-described calibration scene, rough detection of the parallel sorter head edge of Fine filtering, and the target trajectory of the head point analysis can be improved in conjunction with the application, also be used alone.

[0047] 下面对包含所有改进点的本发明最优实施例进行详细分析。 [0047] Hereinafter, the best embodiment of the present invention includes all modifications of the embodiments detailed analysis points.

[0048] I、场景标定 [0048] I, scene calibration

[0049] 由于用于人流量统计的摄像机一般都是固定安装的,场景变化性较小,因此场景标定模块只需要在第一帧检测人头目标前启用,之后各帧检测人头时均采用第一帧标定的结果即可。 [0049] Since the traffic statistics for humans are generally fixedly mounted cameras, a change of scene is small, so the scene to enable the calibration module just prior to the first frame head detection target, are used a first time after the detection of each frame head the results of the calibration frame can be. 如果场景发生变化,则需要再次启用场景标定。 If the scene changes, you need to enable calibration scene again.

[0050] 在摄像机无旋转的情况下,场景的深度变化可以近似为沿图像y坐标成线性变化,即: [0050] In the case where no rotation of the camera, the depth of a scene change can be approximated linearly along the y coordinate of the image changes, namely:

[0051] w(x, y) = f Xy+c (I) [0051] w (x, y) = f Xy + c (I)

[0052] 其中,w(x, y)表示中心图像坐标为(x,y)的人头目标外接矩形的宽度,f为场景深度系数,c为常数。 Width [0052] wherein, w (x, y) represents the center of the image coordinates (x, y) in the target head circumscribed rectangle, f is the coefficient of scene depth, c is a constant. 场景标定的目的就是通过标定框确定f和c的值,从而通过式(I)求出图像中任意坐标处人头目标外接矩形的尺寸。 Object of the scene is to determine the calibration value c by f and the calibration block so that the size of the image at the head of any of the target coordinates of the circumscribed rectangle obtained by the formula (I).

[0053] 本发明通过选择4〜6个标定框计算式(I)中的两个未知量f和C,从而得到场景的深度变化系数,然后将检测区域外接矩形的上边缘和下边缘坐标代入式(I)中,得到检测区域内最小人头尺寸Wmin和最大人头尺寸最后,根据人头尺寸变化范围将检测区域分为若干个子区域,每个子区域对应一个变化较小的人头尺寸范围,在接下来的人头检测模块中,每个子区域用不同尺寸窗口搜索候选矩形。 [0053] 4 ~ 6 of the present invention, by selecting a calibration formula calculation block two unknowns f and C (I), thereby to obtain a depth coefficient of variation of the scene, then the upper edge of the circumscribed rectangle of the coordinate detection region and the lower edge substituting in the formula (the I), to give the maximum and minimum Wmin head size of the head size of the detection zone Finally, according to the head size of the variation range detection area is divided into several sub-regions, each sub-region corresponding to a small change in head size range, in the next the head detection modules, each sub-region candidate rectangles of different search window size.

[0054] 场景标定步骤框图如图3所示,包括: [0054] Scene calibration step diagram shown in Figure 3, comprising:

[0055] S301 :选择标定框; [0055] S301: Select the calibration block;

[0056] S302 :计算场景深度变化系数; [0056] S302: calculating the depth of a scene change coefficient;

[0057] S303 :计算检测区域内人头目标尺寸变化范围; [0057] S303: calculating a target detection head size variation within the region;

[0058] S304 :根据人头目标尺寸变化范围将检测区域划分为若干个子区域。 [0058] S304: The head size of the target range detection area is divided into several sub-regions.

[0059] 至此,场景标定结束。 [0059] Thus, the scene calibration has been completed. 接下来开始在每一帧图像中进行人头的检测、跟踪和计数。 Then it starts detecting the head in each image frame, tracking and counting.

[0060] 2、人头检测 [0060] 2, the detection head

[0061] 本发明中的人头检测分为并联分类器粗检测和边缘特征细筛选两个环节。 Head detection [0061] The present invention is divided into the rough classifier detects and parallel edges two areas wherein the fine filter.

[0062] 并联分类器粗检测环节中通过预先训练好的分类器将大部分非人头目标排除,剩下人头目标和部分误检为人头目标的非人头目标,然后再通过边缘特征细筛选环节去除大部分误检,保留真实人头目标。 [0062] The crude parallel classifier in detecting the most part non-negative target head classifier trained in advance, the remaining head portion of the target and the non-error detection head to head target object, wherein the edge then removed by the fine screening links most false detection, real head retention goals. 人头检测模块结构框图如图4所示。 Head detection module block diagram shown in Figure 4.

[0063] 本发明采用haar特征基于Adaboost算法分别训练包含正面人头和背面人头的深色头发通用分类器、深色头发正面分支分类器、深色头发背面分支分类器、浅色头发分类器、帽子分类器以及为适应特定环境专门设置的扩展分类器等多个分类器。 [0063] The present invention is based on the use of dark hair haar characterized Adaboost algorithm are generic classifier training comprises a front head and the back of the head, the front branch classifier dark hair, dark hair branch back surface classifier, classifier light hair, hat a plurality of classifiers and classifiers adapt a particular specialized environment provided extended classifiers. 多个分类器的组合方式如图4粗检测环节所示:深色头发通用分类器与正面分支分类器、背面分支分类器组合成树形结构,然后与浅色头发分类器、帽子分类器以及扩展分类器形成并联,分类器检测结果进入人头边缘细筛选环节,最后得到真实的人头目标。 Combination of a plurality of classifiers shown in FIG rough detection part 4: dark hair generic classifier and classifiers front branch back into a branched tree classifier combination, and then light hair classifier, and a classifier hat form a parallel extension classifier, classifier edge detection result of the fine filter into the head part, the head finally get the real target.

[0064] 2. I、并联分类器粗检测环节 [0064] 2. I, the rough detection classifier parallel link

[0065] 训练器需要预先用大量正样本和负样本进行训练,本发明采用人脸检测中使用的haar特征加Adaboost算法训练识别器。 [0065] The need to pre-training device for training with a large number of positive and negative samples, the present invention employs a face detection haar characterized added Adaboost algorithm used recognizer training.

[0066] Haar特征由两个或三个不同尺寸的矩形构成。 [0066] Haar feature consists of three or two rectangles of different sizes. 通过改变矩形的尺寸、组合方式和角度可以描述特定目标的形状和灰度信息。 By varying the size of the rectangle, and combinations angle may describe the shape and grayscale information for a particular target. Adaboost算法是一种能将若干弱分类器组合成强分类器的方法。 Adaboost algorithm is a plurality of weak classifiers can be combined into a strong classifier method. 每一个弱分类器选择一个或几个haar特征来对样本进行分类,若干个弱分类器通过Adaboost算法组合成一级强分类器。 Each weak classifier haar select one or several features to classify the samples, a plurality of weak classifiers are combined into a strong classifier through Adaboost algorithm. 本发明中所述的各类分类器,均由若干级强分类器级联而成。 The present invention in various types of classifier, several stages of strong classifiers by cascading.

[0067] 本发明在检测区域内,根据场景标定模块得到的人头目标尺寸,采用穷举的方式搜索人头目标候选矩形。 [0067] The present invention is in the detection area, depending on the scene nominal target size head module obtained using exhaustive search target candidates rectangular head. 将候选矩形分别输入到深色头发通用分类器、浅色头发分类器、帽子分类器以及扩展分类器中进行分类,如果被分类为人头,则该候选矩形被检测为人头目标输出,继续判断下一个候选矩形,否则,将选候选矩形丢弃,继续判断下一个候选矩形。 Candidate rectangles are input into the dark hair generic classifier, classifier classifies light hair, and a cap extension classifier classifier, as if the head is classified, the candidate rectangles are detected as the target output head, is determined to continue a candidate rectangle, otherwise, discarding selected candidate rectangles, continues to determine the next candidate rectangle.

[0068] 在上述过程中,一个候选矩形被分类器分类为人头目标需要逐级通过级联分类器的各级强分类器,否则被分类为非人头目标,其过程示意图如图5所示。 [0068] In the above process, is classified as a candidate rectangle classifier head requires certain levels step by step by a strong classifier cascade classifier, or the head is classified as a target, the process is schematically shown in Fig.

[0069] 另外,上述分类器检测过程中,优先选择的分类器可以根据实际应用调整。 [0069] Further, the above-described detection process classifier, the classifier may be preferred practice adjusted. 一般的应用场景中深色头发的概率最大,因此优先选择深色头发分类器检测,在特定场景,比如检测仓库门口,可优先选择工作帽样本训练得到的扩展分类器检测,以加快检测速度。 General application scenarios the maximum probability dark-haired, dark-haired therefore preferred classifier to detect, in certain scenarios, such as detecting the warehouse door, may prefer the extended classifier is trained to detect cap work samples, in order to speed up the detection speed.

[0070] 2. 2、边缘特征细筛选环节 [0070] 2.2, wherein the edge part of fine filter

[0071] 通过并联分类器粗检测环节,大部分非人头矩形被排除了,只留下真实人头矩形和被分类器误检为人头的矩形。 [0071] By detecting part of the parallel sorter thick, most non-rectangular head is excluded, leaving only the true rectangular head and a classifier for the head of a rectangular false detection. 边缘特征细筛选环节则能通过提取矩形内的边缘特征去除大部分误检矩形,保留真实人头目标。 Wherein the edge part of the fine filter in the rectangular edge feature extraction to remove most of erroneous detection by rectangular head retention real target.

[0072] 本发明采用椭圆上半圆弧作为人头模型,边缘特征细筛选就是计算被分类器判断为人头目标的矩形内边缘特征与椭圆上半圆弧的拟合度。 [0072] The present invention uses a semi-circular elliptical head model, wherein the edge is calculated by determining the fine screening classifier fit the semicircular arc on the inner edge of the target rectangle head ellipse feature. 如果拟合度大于判断阈值,则该矩形为真实人头矩形,否则为误检人头矩形,将该矩形从目标列表中去除。 If the fit is greater than the determination threshold value, the rectangular head is true rectangle, otherwise false detection head rectangular, the rectangle is removed from the target list.

[0073] 3、人头跟踪 [0073] 3, the tracking head

[0074] 人头目标检测出来后需要进行跟踪,形成目标运动轨迹,以避免同一个目标重复计数。 After the required [0074] head detected tracking target, the target trajectory is formed, with a target in order to avoid double counting. 本发明的目标跟踪模块采用粒子滤波算法对人头目标进行跟踪。 Target tracking module of the present invention using particle filtering algorithm for target tracking head.

[0075] 粒子滤波跟踪的流程如图6所示,具体过程如下: [0075] The particle filter tracking process shown in Figure 6, the specific process is as follows:

[0076] 步骤601 :粒子初始化; [0076] Step 601: Initialization particles;

[0077] 当新检测到的人头目标没有已有的粒子对应时,则新生成一个粒子跟踪器,并用新检测到的目标初始化跟踪器中各个粒子的位置和尺寸,并赋给各粒子相等的权重值。 [0077] When a new target is not detected in the head corresponding to the existing particles, a newly generated particle tracker, and with the newly detected object initialization tracker location and size of individual particles, each particle and assigns equal Weights.

[0078] 步骤602 :粒子重采样;[0079] 在跟踪过程中,粒子经过几次权重更新后会出现“退化现象”,即接近真实人头矩形的少数粒子的权重会变得较大,而远离人头矩形的大部分粒子的权重变得很小,大量的计算会浪费在这些权重很小的粒子上。 [0078] Step 602: particle resampling; [0079] in the tracking process, the particles after several weight update will be "degradation" that right of access to a small number of particles in real heavy head rectangle becomes larger, and far away from most right rectangular heavy particles become very small head, large number of calculations will be wasted on these weights very small particles. 为了解决“退化现象”,每次粒子权重更新后应该对粒子进行重采样。 In order to solve the "degradation", each particle should update right after heavy particle resampling.

[0080] 粒子重采样就是保留和复制权重较大的粒子,剔除权重较小的粒子,使原来带权重的粒子映射为等权重的粒子继续预测跟踪。 [0080] particles is retained and reproduced resampling weights larger particles, excluding the weight of smaller particles, so that the original particles mapped with weight weight weight like particle is forecast to continue tracking. 跟踪器新生成时,跟踪器中各粒子的权重相等,因此,不需要在再进行重采样。 Generating a new tracker, tracker weight equal weights of particles, and therefore, does not need further resampling.

[0081] 步骤603 :粒子的传播; 603 [0081] Step: propagation of particles;

[0082] 粒子的传播,也即粒子的状态转移,是指粒子的状态随时间的更新过程。 [0082] propagation particles, i.e. the transition state of the particles, refers to the state of the particles of the update process over time. 本发明中,粒子的状态是指粒子所代表的目标矩形的位置和尺寸。 In the present invention, the state and the particle size refers to the location of the destination rectangle represented particles. 粒子的传播采用一种随机运动过程实现,即粒子的当前状态由上一个状态加上一个随机量得到。 Using a stochastic propagation particle movement is achieved in that the current state of the particles plus a random amount obtained by the previous state. 这样,当前的每一个粒子都代表着人头目标在当前帧中的一个可能位置和尺寸。 Thus, each of the current particle represents the head position and the size may be a target in the current frame.

[0083] 步骤604 :根据观测值更新粒子权重; [0083] Step 604: The observations to update the particle weight;

[0084] 粒子通过传播方式只是得到了人头目标在当前帧中的可能位置和尺寸,还需要利用当前图像的观测值来确定哪些粒子最有可能是人头矩形。 [0084] The particles obtained by the propagation mode only possible location and size of the target head in the current frame, but also to determine which particles are most likely to be observed with the current value of the rectangular head image. 本发明中提取粒子对应图像矩形的haar特征和边缘特征作为观测值更新粒子的权重。 Particles of the present invention extracts features corresponding to the rectangular image haar edge feature as an observation value and weight update particle weight. 粒子的观测值与真实人头越接近, 则该粒子对应的矩形越可能是人头矩形,粒子的权重增大;否则,粒子的权重减小。 Particles is observed and the head closer to the true, the rectangle corresponding to the particle may be more rectangular head, increasing the weight of the particle; otherwise, the weight of the particle decreases.

[0085] 步骤605 :更新目标运动轨迹; [0085] Step 605: Update the target trajectory;

[0086] 将粒子按权重大小排序,取出权重最大的粒子,计算权重最大的粒子对应的矩形与检测得到的所有人头目标矩形的重叠面积,重叠面积最大,且大于设定阈值的人头目标即是该粒子所在跟踪器代表的人头目标在当前帧中对应的人头,则用该人头目标的位置更新跟踪器的目标运动轨迹,并用该人头目标代替权重最大的粒子,进入下一帧跟踪;如果权重最大的粒子与当前帧中检测出来的所有人头目标均不重叠或重叠面积小于阈值,则认为该粒子所在跟踪器代表的人头目标在当前帧中没有找到对应的人头,则用该粒子的位置更新跟踪器的目标运动轨迹,并进入下一帧跟踪。 [0086] particles by weight is large small sort, removed the largest weight particle, calculating the maximum weight of the overlap area of ​​all the head of the destination rectangle rectangle detection obtained particles corresponding to the overlapping area of ​​the largest, and larger than the set threshold value of the head target that is head objective of the particle where the tracker represented corresponding to the current frame head, then using the head position of the target update tracker target trajectory, and the heaviest particles with the head target instead of the right, the next frame tracking; If the weight All head target maximum particle detected in the current frame do not overlap or the overlap area is less than the threshold value, it is considered the head objective of the particle where the tracker representative does not find the corresponding heads in the current frame, the location update of the particles is used target trajectory tracker, track and into the next frame. 如果权重最大的粒子连续N(N> 2)帧找不到对应人头目标,则说明该粒子所在的跟踪器代表的人头目标以及消失,剔除该跟踪器。 If the weight of the largest particle successive N (N> 2) corresponding to the head frame can not find the target, then the head of the particle where the target tracker representatives and disappears, excluding the tracker.

[0087] 经过上述五个步骤,帧与帧之间的人头目标便关联起来形成了人头目标的运动轨迹。 [0087] After above five steps, between the head frame and the target frame will be associated together to form the head of the target trajectory.

[0088] 4、轨迹平滑度分析模块 [0088] 4, trajectory analysis module smoothness

[0089] 一般来说,真实人头目标的运动比较平滑,而误检目标则可能会呈现出杂乱的运动,因此,本发明通过对目标运动轨迹的平滑度分析去除误检,进一步提高检测准确性。 [0089] In general, the real head moving object is smooth, and the target false detection might exhibit chaotic motion, therefore, of the present invention by the smoothing target trajectory analysis removing erroneous detection, to further improve the detection accuracy .

[0090] 对跟踪模块生成的目标运动轨迹进行分析,计算目标轨迹的平滑系数,如果平滑系数大于设定的平滑阈值,则保留该轨迹;否则,剔除该轨迹。 [0090] The target tracking module generates the motion trajectory analysis, smoothing coefficient calculation target track, if the smoothing factor is greater than the set threshold value smoothing, the track is retained; otherwise, excluding the trajectory. 轨迹平滑度分析模块流程如图7所示,包括: Track smoothness flow analysis module shown in Figure 7, comprising:

[0091] S701 :获取目标运动轨迹; [0091] S701: acquiring a target trajectory;

[0092] S702 :确定人头目标运动轨迹的平滑度; [0092] S702: determining the smoothness of the head of the target trajectory;

[0093] S703 :判断平滑度是否满足预置的平滑度阈值要求,若是,执行S704,否则,执行S705 ; [0093] S703: determining whether or not to meet smoothness requirements of smoothness preset threshold, if yes, S704, otherwise performs S705;

[0094] S704 :保留该目标运动轨迹;[0095] S705 :丢弃该目标运动轨迹; [0094] S704: retention of the target trajectory; [0095] S705: discarding the target trajectory;

[0096] S706 :输出目标运动轨迹。 [0096] S706: Output the target trajectory.

[0097] 5、人流量计数模块 [0097] 5, the number of people meter module

[0098] 本发明通过人头目标运动轨迹方向对人流量进行计数。 [0098] The present invention for the flow of people moving in the track direction by counting the target head. 本发明在检测区域内判断该目标轨迹的方向与设定的“人流进入”方向是否一致,如果一致,则“进入人数”计数加一, 否则“离开人数”计数加一。 Analyzing the target track of the present invention in the detection area coincides with a direction of setting the "flow into the", if yes, "enter number" count is incremented by one, otherwise, "the number of leaving" count by one. 计数完成后将该目标标记为“已计数”,使轨迹处于无效状态, 避免同一个目标重复计数。 After completion of the count target flag to "count" the track in an invalid state, to avoid double counting the same goal.

[0099] 至此,通过场景标定、人头检测、人头目标跟踪、人头目标运动轨迹分析和人流量统计这五大步骤,即完成了对人流量的全面、准确统计。 [0099] Thus, by scaling the scene, head detection, target tracking poll, the head target trajectory analysis and human traffic statistics these five steps to complete a comprehensive, accurate statistics on the flow of people.

[0100] 与上述方法相对应,本发明还提供一种人流量统计的系统,该系统可通过软件、硬件或软硬件结合实现。 [0100] correspond to the above method, the present invention also provides a human traffic statistics system, the system can be implemented in conjunction with software, hardware, or hardware.

[0101] 参考图8,该系统包括: [0101] Referring to Figure 8, the system comprising:

[0102] 人头检测模块801,用于采用分类器对当前图像进行人头检测,确定当前图像中的各人头; [0102] head detection module 801, classifier for employing image current detection head, each head is determined in the current image;

[0103] 人头目标跟踪模块802,用于对人头检测模块801确定出的各人头进行跟踪,形成人头目标运动轨迹; [0103] head tracking module 802, for each head detector module 801 determines the head of the track, the target trajectory forming head;

[0104] 人流量计数模块803,用于在人头目标跟踪模块802确定的人头目标运动轨迹方向进行人流量计数; [0104] the number of people meter module 803, a flow meter for the number of people moving in the direction of the head target track head tracking module 802 determines the target;

[0105] 特别地,该系统还包括人头目标运动轨迹分析模块804,用于计算人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹。 [0105] In particular, the system further comprising a head moving target trajectory analysis module 804 for smoothing head calculating the target trajectory, determines the smoothness meets a threshold, if yes, to retain the head target trajectory, otherwise, discards the head target trajectory. 此时,人流量计数模块803是在人头目标运动轨迹分析模块804的基础上,根据运动轨迹方向的人头进行统计的。 In this case, the number of people meter head module 803 is the target trajectory on the basis of analysis module 804, according to statistics head trajectory direction.

[0106] 优选地,分类器采用并联的多类分类器实现,例如,由深色头发通用分类器、浅色头发分类器、帽子分类器和扩展分类器中的任意两种或多种并联而成,此时,人头检测模块801包括粗检测子模块和细筛选子模块,其中,粗检测子模块用于设置各类分类器的检测顺序,按照检测顺序依次采用各个分类器对当前图像进行人头检测,直到确定出人头;筛选子模块用于对并联的多类分类器检测到的人头进行边缘特征细筛选处理。 [0106] Preferably, the classifier using parallel multi-class classifier implemented, for example, a general-purpose classifier dark hair, light hair classifier, classifier and caps of any two spreading classifiers or more parallel and into this case, the head detection module 801 includes a coarse and a fine filter sub-module detecting sub-module, wherein the crude sub-module for detecting various types of setting the detection order classifier, according to the detection order of each classifier using the current image head detection, until it is determined to succeed; submodule for screening the head of the parallel multi-class classifier to detect the edge of fine filtering process.

[0107] 优选地,该系统还包括: [0107] Preferably, the system further comprising:

[0108] 场景标定模块805,用于对图像中的检测区域进行场景标定,从而将检测区域划分为若干个子区域。 [0108] Scene calibration module 805, configured to image a scene calibration detection region, so that the detection area is divided into several sub-regions. 其中,场景标定模块805的目的是获得场景的深度系数,根据场景深度系数可以计算出图像中各个位置的人头目标的大小,为人头目标检测模块提供检测尺寸。 Wherein the calibration module object scene 805 is a scene depth coefficients obtained can calculate the size of the head position of each object image according to the scene depth coefficient, a size of the detection head to provide target detection module. 此时,人头检测模块801根据场景标定模块805提供的尺寸,在指定的若干个子区域内搜索人头目标。 In this case, the head detection module 801 calibration module 805 provided in accordance with the size of the scene, the target search head in several sub-area specified.

[0109] 上述系统的具体实现请参见方法实施例,不作赘述。 Specific [0109] The system implemented see Example, will not be described herein.

[0110] 可见,本发明通过对人头目标轨迹的平滑度分析可以去除虚假目标,可提高检测准确率。 [0110] visible, by the present invention, the smoothness of the head target track analysis may be removed false targets, detection accuracy can be improved. 进一步,本发明采用haar特征基于Adaboost算法训练多个并联的分类器作为人头粗检测,再利用边缘特征对粗检测结果进行细筛选,最后得到真正的人头目标。 Further, the present invention employs a plurality of parallel features haar Adaboost algorithm trained classifiers as the rough detection head reuse edge feature detection results of the coarse to fine filtering based on the last head get real target. 本发明中将多个分类器并联使用,能同时检测深色头发、浅色头发以及各种颜色帽子等多类人头目标, 本发明还设置了一个扩展分类器,可以根据特殊环境的应用,采集样本训练,检测指定颜色或帽子的人头,比如工厂或仓库的工作帽等。 In the present invention, a plurality of classifiers in parallel, simultaneous detection of dark hair, light hair colors and various other types of hats head objective, the present invention is also provided with an extension classifier, depending on the application of the special environment, gathering sample training, a hat or head specified color detection, such as a factory or warehouse work caps. 另外,本发明在检测前通过场景标定自动选择检测窗口的尺寸,使本发明能自适应各种摄像机角度,拓宽了应用范围。 Further, the present invention is calibrated prior to detection of the detection window size is automatically selected by the scene, of the present invention can adaptively various camera angles, broadens the scope of application.

[0111] 以上所述仅是本发明的优选实施方式,应当指出,对于本技术领域的普通技术人员来说,在不脱离本发明原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本发明的保护范围。 [0111] The above are only preferred embodiments of the present invention, it should be noted that those of ordinary skill in the art, in the present invention without departing from the principles of the premise, can make various improvements and modifications, such modifications and modifications should also be regarded as the protection scope of the present invention.

Claims (11)

1. 一种可去除虚假目标的人流量统计的方法,其特征在于,包括:对图像中的检测区域进行场景标定,以获取所述检测区域内人头目标尺寸变化范围, 从而根据所述人头目标尺寸变化范围将所述检测区域划分为若干个子区域;在所述子区域内采用分类器对当前图像进行人头检测,确定当前图像中的各人头; 对确定出的各人头进行跟踪,形成人头目标运动轨迹;对人头目标运动轨迹进行平滑度分析;根据分析后的人头目标运动轨迹方向进行人流量计数。 A removable false targets human traffic statistics, characterized in that, comprising: a detection region in the image scene calibration to obtain the target size of the head region detection range, so that the head according to the target the dimensional change range detection area is divided into several sub-regions; head to each target tracking head is determined, is formed; employed in the sub-region on the current image classifier detects the head, each head is determined in the current image trajectory; on the smoothing head target trajectory analysis; for the number of people meter head according to the direction of the target trajectory after the analysis.
2.根据权利要求I所述方法,其特征在于,所述对人头目标运动轨迹进行平滑度分析包括:确定人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹。 2. The method according to claim I, wherein the analysis includes smoothing the head target trajectory: determining a smoothness head target trajectory, determines the smoothness meets a threshold, if yes, to retain the head target trajectory, otherwise, discard the head target trajectory.
3.根据权利要求I所述方法,其特征在于,在采用分类器对当前图像进行人头检测之后、确定当前图像中的各人头之前,还包括:对分类器检测到的人头进行边缘特征细筛选处理。 3. The method according to claim I, wherein, after the classifier using the current image detecting head, before the current image of each head is determined, further comprising: a classifier for detecting the edge of the head of Fine filtering deal with.
4.根据权利要求3所述方法,其特征在于,所述对分类器检测到的人头进行边缘特征细筛选处理包括:计算所述分类器判断为人头目标的矩形内边缘特征与预置的上半椭圆弧的拟合度,如果拟合度大于阈值,则将该矩形确定为人头,否则将该矩形从目标列表中去除。 4. The method according to claim 3, characterized in that the detected head classifier edge of Fine filtering process comprising: calculating the classifier determines rectangle on the inner head edge feature and the target preset semi-elliptic arc fit, if the fitting is greater than the threshold value, then the rectangle is determined to be the head, a rectangular or removed from the target list.
5.根据权利要求I所述方法,其特征在于,所述对图像中的检测区域进行场景标定,以获取所述检测区域内人头目标尺寸变化范围,从而根据所述人头目标尺寸变化范围将所述检测区域划分为若干个子区域包括:选择标定框;计算场景深度变化系数;计算检测区域内人头目标尺寸变化范围;根据人头目标尺寸变化范围将检测区域划分为若干个子区域。 5. The method according to claim I, wherein the calibration of a scene image detection region, the detection region to obtain the target size range head, whereby the head according to the target size range The said detection area is divided into several sub-areas comprising: selecting a calibration block; scene depth variation coefficient calculated; calculating head size variation range detection target area; head according to the target size range detection area is divided into several sub-regions.
6.根据权利要求I至5任一项所述方法,其特征在于,所述分类器为并联的多类分类器。 6. The method according to any one claims I to 5, wherein the classifier is a parallel multi-class classifier.
7.根据权利要求6所述方法,其特征在于,所述多类分类器对图像进行人头检测包括: 设置各类分类器的检测顺序,按照检测顺序依次采用各个分类器对当前图像进行人头检测,直到确定出人头,其中,所述并联的多类分类器由至少两类分类器并联而成。 7. The method according to claim 6, wherein said multi-class classifier image detecting head comprising: setting various types of classifiers detection order, according to the detection order of each classifier using the current image detecting head , until it is determined to succeed, wherein the parallel multi-class classifier from at least two parallel classifier.
8.根据权利要求6所述方法,其特征在于,所述并联的多类分类器由深色头发通用分类器、浅色头发分类器、帽子分类器和扩展分类器中的任意两种或多种并联而成。 8. The method according to claim 6, wherein any two of the parallel multi-class classifier dark hair by the general classifier, classifier light hair, and the extension caps classifier or classifiers species parallel connection.
9. 一种可去除虚假目标的人流量统计的系统,其特征在于,包括:场景标定模块,用于对图像中的检测区域进行场景标定,以获取所述检测区域内人头目标尺寸变化范围,从而根据所述人头目标尺寸变化范围将所述检测区域划分为若干个子区域;人头检测模块,用于在所述子区域内采用分类器对当前图像进行人头检测,确定当前图像中的各人头;人头目标跟踪模块,用于对确定出的各人头进行跟踪,形成人头目标运动轨迹;人头目标运动轨迹分析模块,用于计算人头目标运动轨迹的平滑度,判断所述平滑度是否满足阈值,若是,保留该人头目标运动轨迹,否则,丢弃该人头目标运动轨迹;人流量计数模块,用于在分析后的人头目标运动轨迹方向进行人流量计数。 A removable false targets human traffic statistics system, characterized by comprising: a scene calibration module configured to image a scene calibration detection region, the detection region to obtain the target size range head, the head such that the variation range of the size of the target detection area is divided into several sub-regions; head detection module configured to poll the current image detected by the classifier within subregions, each head is determined in the current image; head tracking module for each of the determined head tracking, the target trajectory forming head; head target trajectory analysis module for smoothness calculating the target trajectory of the head, the flatness is determined whether the threshold is met, if retain the head target trajectory, otherwise, discarding the head target trajectory; people meter module number for the number of people meter in the direction of head motion trajectory after the target analysis.
10.根据权利要求9所述系统,其特征在于,所述分类器为并联的多类分类器;所述人头检测模块包括粗检测子模块和细筛选子模块,所述粗检测子模块用于设置各类分类器的检测顺序,按照检测顺序依次采用各个分类器对当前图像进行人头检测,直到确定出人头,其中,所述并联的多类分类器由至少两类分类器并联而成;细筛选子模块,用于对并联的多类分类器检测到的人头进行边缘特征细筛选处理。 10. The system of claim 9, wherein the classifier is a parallel multi-class classifier; detecting the head module comprises coarse and fine screening detection sub-module sub-module, sub-module for detecting the crude provided various types of classifier detection order, according to the detection order of each classifier using the current image detecting head, until it is determined to succeed, wherein the parallel multi-class classifier from at least two parallel classifier together; fine filtering sub-module, a head pair of parallel multi-class classifier to detect the edge of fine filtering process.
11.根据权利要求10所述系统,其特征在于,所述并联的多类分类器由深色头发通用分类器、浅色头发分类器、帽子分类器和扩展分类器中的任意两种或多种并联而成。 11. The system of claim 10, wherein the parallel multi-class classifier of any two generic classifier dark hair, light hair classifier, classifier and cap extension or classifiers species parallel connection.
CN 201010118136 2010-02-10 2010-02-10 People flow rate statistical method and system capable of removing false targets CN101872414B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201010118136 CN101872414B (en) 2010-02-10 2010-02-10 People flow rate statistical method and system capable of removing false targets

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201010118136 CN101872414B (en) 2010-02-10 2010-02-10 People flow rate statistical method and system capable of removing false targets

Publications (2)

Publication Number Publication Date
CN101872414A CN101872414A (en) 2010-10-27
CN101872414B true CN101872414B (en) 2012-07-25

Family

ID=42997269

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201010118136 CN101872414B (en) 2010-02-10 2010-02-10 People flow rate statistical method and system capable of removing false targets

Country Status (1)

Country Link
CN (1) CN101872414B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130101159A1 (en) * 2011-10-21 2013-04-25 Qualcomm Incorporated Image and video based pedestrian traffic estimation
CN103577832B (en) * 2012-07-30 2016-05-25 华中科技大学 A kind of based on the contextual people flow rate statistical method of space-time
CN105704434A (en) * 2014-11-28 2016-06-22 上海新联纬讯科技发展有限公司 Stadium population monitoring method and system based on intelligent video identification
CN104778474B (en) * 2015-03-23 2019-06-07 四川九洲电器集团有限责任公司 A kind of classifier construction method and object detection method for target detection

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101416512A (en) 2006-04-07 2009-04-22 微软公司 Quantization adjustment based on texture level
CN101464946A (en) 2009-01-08 2009-06-24 上海交通大学 Detection method based on head identification and tracking characteristics
CN101477641A (en) 2009-01-07 2009-07-08 北京中星微电子有限公司 Demographic method and system based on video monitoring

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2431717A (en) * 2005-10-31 2007-05-02 Sony Uk Ltd Scene analysis

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101416512A (en) 2006-04-07 2009-04-22 微软公司 Quantization adjustment based on texture level
CN101477641A (en) 2009-01-07 2009-07-08 北京中星微电子有限公司 Demographic method and system based on video monitoring
CN101464946A (en) 2009-01-08 2009-06-24 上海交通大学 Detection method based on head identification and tracking characteristics

Also Published As

Publication number Publication date
CN101872414A (en) 2010-10-27

Similar Documents

Publication Publication Date Title
US8611604B2 (en) Object detection device
US8472672B2 (en) System and process for detecting, tracking and counting human objects of interest
Xing et al. Multi-object tracking through occlusions by local tracklets filtering and global tracklets association with detection responses
US8582816B2 (en) Method and apparatus for video analytics based object counting
US8983133B2 (en) Multi-view object detection using appearance model transfer from similar scenes
Li et al. Rapid and robust human detection and tracking based on omega-shape features
Yang et al. Online learned discriminative part-based appearance models for multi-human tracking
EP2801078B1 (en) Context aware moving object detection
Stalder et al. Cascaded confidence filtering for improved tracking-by-detection
US20070098222A1 (en) Scene analysis
CN101231755A (en) Moving target tracking and quantity statistics method
CN104303193B (en) Target classification based on cluster
CN101739551A (en) Method and system for identifying moving objects
CN1794264A (en) Method and system of real time detecting and continuous tracing human face in video frequency sequence
CN1897015A (en) Method and system for inspecting and tracting vehicle based on machine vision
CN102214291A (en) Method for quickly and accurately detecting and tracking human face based on video sequence
CN101389004A (en) Moving target classification method based on on-line study
CN102799863B (en) Method for detecting group crowd abnormal behaviors in video monitoring
CN100589561C (en) Dubious static object detecting method based on video content analysis
CN101577812A (en) Method and system for post monitoring
CN101477641A (en) Demographic method and system based on video monitoring
Zhao et al. A people counting system based on face detection and tracking in a video
CN101729872B (en) Video monitoring image based method for automatically distinguishing traffic states of roads
CN101464946A (en) Detection method based on head identification and tracking characteristics
US9008365B2 (en) Systems and methods for pedestrian detection in images

Legal Events

Date Code Title Description
C06 Publication
C10 Request of examination as to substance
C14 Granted
ASS Succession or assignment of patent right

Owner name: HANGZHOU HIKVISION DIGITAL TECHNOLOGY CO., LTD.

Free format text: FORMER OWNER: HANGZHOU HAIKANG WEISHI SOFTWARE CO., LTD.

Effective date: 20121025

C41 Transfer of the right of patent application or the patent right
COR Bibliographic change or correction in the description

Free format text: CORRECT: ADDRESS; FROM: 310012 HANGZHOU, ZHEJIANG PROVINCE TO: 310051 HANGZHOU, ZHEJIANG PROVINCE