WO2023155482A1 - 一种人群快速聚集行为的识别方法、系统、设备及介质 - Google Patents
一种人群快速聚集行为的识别方法、系统、设备及介质 Download PDFInfo
- Publication number
- WO2023155482A1 WO2023155482A1 PCT/CN2022/128971 CN2022128971W WO2023155482A1 WO 2023155482 A1 WO2023155482 A1 WO 2023155482A1 CN 2022128971 W CN2022128971 W CN 2022128971W WO 2023155482 A1 WO2023155482 A1 WO 2023155482A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- frame
- gathering
- crowd
- rectangular
- area
- Prior art date
- Legal status (The legal status 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 status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Definitions
- the present disclosure relates to the technical field of video surveillance, and in particular to a method, system, device, and computer-readable storage medium for identifying a crowd's rapid gathering behavior.
- Traditional crowd gathering detection methods mainly include optical flow method (refer to patent document CN107330372B), inter-frame difference method (refer to patent document CN112232316B) and mathematical statistics method (refer to patent document CN105117683B); wherein, optical flow method and inter-frame difference method pass
- optical flow method and inter-frame difference method pass
- the extracted foreground features are used to estimate crowd gathering, but the optical flow method takes a long time, the frame difference method is poorly adapted to noise, and mathematical statistics is mainly designed for specific scenes.
- the above three methods are difficult to adapt to various environmental conditions.
- the technical problem to be solved in the present disclosure is that the traditional crowd gathering detection method is difficult to adapt to various environmental conditions.
- one of the purposes of the present disclosure is to provide a highly adaptable method for identifying rapid crowd gathering behavior, which is applicable to the requirements of multiple scenarios.
- the second object of the present disclosure is to provide a recognition system for rapid crowd gathering behavior.
- the third objective of the present disclosure is to provide an electronic device.
- the fourth objective of the present disclosure is to provide a computer-readable storage medium.
- a method for identifying rapid crowd gathering behavior comprising:
- the head detection model is obtained by training the neural network through an image sample set marked with pedestrian heads; and when the head detection model detects that there is a head in the image, the The location generates its corresponding rectangular box for display.
- a rectangular region growing process is performed on each preliminary aggregated region to obtain the final aggregated region.
- the method for growing the rectangular area is:
- each aggregation area calculates the target rectangle with the most adjacent rectangles in each aggregation area, obtain the center point of the target rectangle in the current aggregation area, and expand the area outward with the first fixed distance as the radius
- a set of rectangles traversing the set of rectangles to search for a rectangle whose distance is less than or equal to a second fixed distance and storing it in a second adjacent rectangle array; wherein the second fixed distance is less than the first fixed distance;
- a union operation is performed on the first adjacent rectangular array and the second adjacent rectangular array corresponding to each aggregation area to obtain an area growing result of each aggregation area.
- the first fixed distance (the square of the width value of the rectangular frame+the square of the height value of the rectangular frame)*2.5
- the second fixed distance (the square of the width value of the rectangular frame+the square of the height value )*1.5.
- the calculation method for the congestion coefficient of the aggregation area in the current frame image is:
- len[L(n)] indicates the size of the currently detected rectangular frame array
- len(G p ) indicates the sum of the number of people in all gathering areas
- len(G b ) indicates the total number of gathering areas
- the value of ⁇ is related to the size of the number of people in the largest gathering area, corresponding to ([3,9), [9,12), [12,+ ⁇ )) number of people.
- the calculation method for the congestion coefficient of the non-clustered area in the current frame image is:
- the scene coefficient ⁇ is 0.01 ⁇ 0.04, and len[L(n)] represents the size of the currently detected rectangular frame array; when Crd exceeds 1.0, the value is 1.0.
- a system for identifying rapid crowd gathering behavior performing the identification method for rapid crowd gathering behavior as described above, including:
- the head detection module is used to obtain continuous frame images, and performs head detection on each frame of images sequentially based on the head detection model;
- the crowd analysis module is used to perform cluster analysis on the aggregation area of the entire frame of images according to the head detection results, calculate the number of heads in the aggregation area after each cluster, and calculate the congestion coefficient corresponding to each frame of image;
- the judgment output module is used to judge whether there is a frame image in which both the congestion coefficient and the number of people in the maximum gathering area exceed their corresponding preset thresholds within the preset time period, and if so, output the result of rapid crowd gathering.
- An electronic device comprising a processor, a memory, and a computer program stored on the memory and operable on the processor, when the processor executes the computer program, the recognition of the above-mentioned rapid crowd gathering behavior is realized method.
- a computer-readable storage medium on which a computer program is stored, and when the computer program is executed, the above-mentioned method for identifying rapid crowd gathering behavior is realized.
- This disclosure proposes a crowd feature description method, which uses head recognition to solve the problem of feature expression difficulties caused by the serious lack of human body contours when crowds are densely distributed;
- the distribution situation divides the crowd-intensive areas in the image, and improves the accuracy of rapid crowd detection; at the same time, combined with the congestion coefficient calculation method, the disclosed method has stronger scene adaptability, can adapt to the needs of multiple scenes at the same time, and has high feasibility. sex.
- FIG. 1 is a schematic diagram of the principle of the identification method for the rapid gathering behavior of the disclosed crowd
- FIG. 2 is a schematic flow diagram of the identification method for the rapid gathering behavior of the disclosed crowd
- FIG. 3 is a schematic diagram of the neural network structure of the head detection model of the present disclosure.
- FIG. 4 is a block diagram of a recognition system for the rapid crowd gathering behavior of the present disclosure.
- This embodiment provides a method for identifying the rapid gathering behavior of crowds. This method can be applied to actual scenes with large background noise and complex and changeable environments, and can adapt to the requirements of multiple scenes at the same time, and has stronger adaptability.
- the rapid crowd gathering behavior in this embodiment specifically refers to the rapid change of the crowd from the previous sparse and uncongested state to the congested state within a specified time, that is, the sudden crowd gathering describes that within a certain period of time, the crowd changes from the initial sparse , and no one even quickly gathers in one or several locations, and the degree of congestion and the number of people in the gathering area exceed the threshold set in advance.
- this embodiment proposes a rapid crowd gathering identification method:
- Step S1 Obtain consecutive frames of images, and perform head detection on each frame of images sequentially based on the head detection model;
- Step S2 Carry out cluster analysis on the aggregated area of the entire frame of images according to the head detection results, calculate the number of heads in each clustered aggregated area, and calculate the congestion coefficient corresponding to each frame of image;
- Step S3 Comparing the congestion coefficient with a preset value, when there is a target frame image whose congestion coefficient is greater than the preset sparse coefficient, calculate and update the target frame image as a starting point, within a preset time period The congestion coefficient and the maximum number of people in the gathering area of all frame images in the frame image; determine whether there is a frame image in which the congestion coefficient and the maximum number of people in the gathering area exceed their corresponding preset thresholds in the preset time period, if there is , then output the result of rapid crowd gathering.
- This embodiment considers that the crowd is in a dense state when they gather, and the human body outline information is seriously lacking, but the head information is roughly complete.
- the head feature is the main crowd information description feature; therefore, this embodiment uses pedestrians based on head
- the detection method is used to describe the approximate division of the crowd in a certain frame of image.
- a head detection model needs to be established in advance.
- this embodiment acquires an image sample set in advance.
- the image sample set includes a large number of images containing head features.
- Each sample image takes the scale of 1920*1080 as the reference point. If the sample image exceeds 1920*1080 pixels, you need to After cropping and normalizing the image samples, mark the target whose head width is greater than 20 pixels in each sample image, and use the image samples after the above processing as the training samples of the convolutional neural network for learning and training. Train to build a head detection model.
- the convolutional neural network of the head detection model in this embodiment includes a total of 19 convolutional layers and 5 Pooling layers from conv0/relu to conv18/relu, and the regression layer consists of conv6, conv9, conv12, and conv14 respectively.
- conv16, and conv18 layers are composed of output after relu, and each regression layer has 6 PriorBoxes.
- the size of the image input by the network is 640*640.
- each frame is input into the head detection model that has been built.
- the head detection model detects that there is a head in the image, it generates Its corresponding rectangular frame is displayed.
- a rectangular frame is used to represent the position of the pedestrian's head.
- a crowd gathers, multiple rectangular frames will appear close to each other. Therefore, the distribution of the crowd can be described only by considering the relationship between each rectangular frame. It is generally believed that when more than 3 people gather in the same area, there is a greater possibility of congestion. Therefore, this embodiment takes at least 3 people to form a gathering area as the basic point, and calculates the number of gathering areas of the entire frame image and the number of each gathering area. The number of heads in the region.
- the acquisition principle of the gathering area G(Y) in this embodiment is as follows:
- the target person with the largest number of neighbors in the crowd is distributed in the densest central area. If the target rectangle with the most adjacent rectangles can be found when calculating the aggregation area, the target rectangle is the closest to a certain aggregation area. The center of , combined with the rectangular area growing process, can roughly find out all the rectangular frames in this aggregation area.
- the method for calculating the target rectangular frame with the most adjacent rectangular frames is:
- the first fixed distance is the basis for judging whether the "adjacent rectangular frame" is satisfied, specifically: in an array composed of a group of detection rectangular frames, the distance between the center points of two rectangular frames is used as the distance between different rectangular frames.
- Distance S(x 1 , x 2 ) when it is necessary to obtain the number of other rectangular frames adjacent to a certain rectangular frame x 1 , it needs to be calculated in conjunction with the first fixed distance, if a rectangular frame x 1 is connected to another rectangular frame x 2 If the distance between them is smaller than the first fixed distance, then the two are in the adjacent relationship; if the distance between the two is greater than the first fixed distance, then the two are not in the adjacent relationship.
- the first fixed distance (of the rectangular frame
- the calculation of the adjacent rectangular frame is to calculate the sum of the squares of the width and height satisfying the distance from the center point of x 1 less than 2.5 times x 1 All rectangles of .
- the square sum of 2.5 times the width and height is used, that is, the distance between the target rectangular frame and other rectangular frames in the aggregation area is kept within the first fixed distance range, if there are still other rectangular frames distributed At the edge of a rectangular box gathering area, it should belong to this gathering area.
- calculate the number of adjacent rectangular frames of each rectangular frame judge whether the number of adjacent rectangular frames of each rectangular frame is less than the preset value, and divide the number of adjacent rectangular frames smaller than the preset value and its
- the corresponding first adjacent rectangular array is deleted to obtain a preliminary gathering area; in this embodiment, a rectangular frame whose number of adjacent rectangular frames is less than 2 and the corresponding first adjacent rectangular array are deleted.
- each aggregation area in the preliminary aggregation area set is traversed, and a rectangular area growing process is performed on each aggregation area to obtain the final aggregation area G(Y).
- the method for growing the rectangular region is:
- Step S21 Traverse each aggregation area, calculate the target rectangle with the most adjacent rectangles in each aggregation area, obtain the center point of the target rectangle in the current aggregation area, and take the first fixed distance as the radius direction A collection of rectangles in the outer extension area; for example, assuming one of the aggregation areas is Vec[L(b)], and the target rectangle with the most adjacent rectangles is L(b), calculate Vec[L(b)] The set of rectangles distributed outside the circle with the center point of the target rectangular frame L(b) as the center and the first fixed distance as the radius is denoted as Vec 0 [L(b)].
- Step S23 Calculate the rectangular frame Vec last [L(b)] formed by the union of the first adjacent rectangular array Vec[ L (b)] and the second adjacent rectangular array Vec 1 [L(b)], which is a certain Rectangular region growing results for aggregated regions.
- the result of all the aggregation areas after the rectangular area growth process is recorded as the final aggregation area G(Y), and the number of heads in the final aggregation area is Y(X), and the number of heads can be directly detected by the head It is represented by the number of head rectangles detected by the model.
- the calculation of the congestion coefficient in this example is divided into two cases, that is, there is an aggregation area and there is no aggregation area; specifically:
- this embodiment introduces a scene coefficient ⁇ , which represents The ratio of people to the full scene of the whole scene, when the scene is relatively open, the ⁇ value is relatively small, and when the scene field of vision is narrow, the ⁇ value is relatively large; after many tests in this embodiment, it is shown that 0.01 ⁇ 0.04 is the best, therefore,
- the calculation method of the congestion coefficient of the current frame image where there is no cluster area is as follows:
- the scene coefficient ⁇ is 0.01 ⁇ 0.04, and len[L(n)] represents the size of the currently detected rectangular frame array; when Crd exceeds 1.0, the value is 1.0.
- the congestion coefficient calculation method is:
- len[L(n)] indicates the size of the currently detected rectangular frame array
- len(G p ) indicates the sum of the number of people in all gathering areas
- len(G b ) indicates the total number of gathering areas
- the value of ⁇ is related to the size of the number of people in the largest gathering area, corresponding to ([3,9), [9,12), [12,+ ⁇ )) number of people.
- this embodiment finds a suitable coefficient value to express the sparseness of the crowd through experiments in a large number of scenarios. It is more reasonable for the sparse coefficient to be 0.3, that is, if it is less than this value, it will not be included in the time of rapid aggregation. Referring to Fig.
- start timing when the congestion coefficient is greater than the sparse coefficient calculate and update the congestion coefficient and the maximum number of people in the gathering area of all frames in the preset time period; If it does not exceed the pre-set preset threshold, then it is considered that the crowd does not gather rapidly during this preset period of time, and then the head detection can continue to be performed on the next frame of image to analyze the crowd gathering behavior of the next frame of image. If the congestion coefficient and the number of people in the maximum gathering area both exceed the preset threshold in a certain frame within the preset time period, it means that there is a rapid crowd gathering behavior. The crowd cannot disperse immediately.
- the rapid crowd gathering behavior it is also possible to consider whether the remaining frames meet the condition that the congestion coefficient and the maximum number of people in the gathering area exceed the preset threshold. If so, the crowd can be considered to be congested Both the coefficient and the number of people in the maximum aggregation area are rapidly aggregated within the time corresponding to the frame image whose number exceeds the preset threshold.
- this embodiment provides a system for identifying rapid crowd gathering behavior, which implements the identification method for rapid crowd gathering behavior as described in Embodiment 1, including:
- the head detection module is used to obtain continuous frame images, and performs head detection on each frame of images sequentially based on the head detection model;
- the crowd analysis module is used to perform cluster analysis on the aggregation area of the entire frame of images according to the head detection results, calculate the number of heads in the aggregation area after each cluster, and calculate the congestion coefficient corresponding to each frame of image;
- the judgment output module is used to judge whether there is a frame image in which both the congestion coefficient and the number of people in the maximum gathering area exceed their corresponding preset thresholds within the preset time period, and if so, output the result of rapid crowd gathering.
- This embodiment utilizes the Box-Gather (detection frame aggregation) method to process the head detection results to obtain the number of heads in the aggregation area and each area; Congestion degree (0-1); within a specified period of time, when the degree of congestion exceeds a certain threshold and the number of people in the gathering area exceeds the preset threshold, it is considered that the crowd gathers rapidly.
- This embodiment can adapt to scenes with large background noise and complex and changeable environments, and greatly improves the adaptability of crowd gathering behavior recognition.
- This embodiment provides an electronic device, which includes a processor, a memory, and a computer program stored on the memory and operable on the processor.
- the processor executes the computer program, the first embodiment is implemented.
- a method for identifying rapid crowd gathering behavior in addition, this embodiment also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed, the aforementioned method for identifying crowd rapid gathering behavior is realized.
- the recognition method for the rapid gathering behavior of crowds provided by the present disclosure accurately divides the crowd-dense areas in the image according to the crowd distribution by clustering each gathering area, and combined with the congestion coefficient calculation method, it not only has stronger scene adaptation It can adapt to the needs of multiple scenarios at the same time, and can also improve the accuracy of rapid crowd gathering detection, which has strong industrial applicability.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
本公开公开一种人群快速聚集行为的识别方法、系统、设备及介质，其识别方法包括：获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；当存在拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以目标帧图像为起点，在预设时间段内所有帧图像的拥堵系数及最大聚集区域人数；判断预设时间段内是否存在拥堵系数和最大聚集区域人数均超过其对应的预设阈值的帧图像，若存在，则输出人群快速聚集的结果。本公开解决了人群密集时人体轮廓缺失严重带来的特征表达困难问题，同时提高识别方法适应性。
Description
本公开要求于2022年02月17日提交中国专利局、申请号为202210147617.0、发明名称为“一种人群快速聚集行为的识别方法、系统、设备及介质”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
本公开涉及视频监控技术领域，尤其涉及一种人群快速聚集行为的识别方法、系统、设备及计算机可读存储介质。
现阶段，随着公共安全意识的提升和安防技术的快速发展，视频监控系统已逐步应用于平安城市领域。传统通过人工检索查看实时视频的监控方式已日渐不能满足城市快速发展的需求，而随着计算机性能的不断提升、计算机视觉技术的不断发展，基于图像处理的方式越来越多的被运用到智能视频监控中。
传统的人群聚集检测方法主要有光流法(参考专利文献CN107330372B)、帧间差分法(参考专利文献CN112232316B)及数学统计法(参考专利文献CN105117683B)；其中，光流法和帧间差分法通过提取的前景特征来估计人群聚集情况，但光流法耗时长，帧间差分法对噪声适应差，而数学统计主要针对特定场景设计，上述三种方法均难以适应多样的环境条件。
发明内容
(一)要解决的技术问题
本公开要解决的技术问题为传统的人群聚集检测方法难以适应多样的环境条件的问题。
(二)技术方案
为了克服现有技术的不足，本公开的目的之一在于提供一种适应性较强的人群快速聚集行为的识别方法，可适用于多个场景需求。
本公开的目的之二在于提供一种人群快速聚集行为的识别系统。
本公开的目的之三在于提供一种电子设备。
本公开的目的之四在于提供一种计算机可读存储介质。
本公开的目的之一采用如下技术方案实现：
一种人群快速聚集行为的识别方法，包括：
获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；
根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；
当存在所述拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以所述目标帧图像为起点，在预设时间段内所有帧图像的所述拥堵系数及最大聚集区域人数；判断所述预设时间段内是否存在所述拥堵系数和所述最大聚集区域人数均超过其对应的预设阈值的帧图像，若存在，则输出人群快速聚集的结果。
进一步地，所述头部检测模型是通过标注有行人头部的图像样本集对神经网络进行训练获得；并在所述头部检测模型检测到图像中存在头部时，为每个头部所在位置生成其对应的矩形框进行显示。
进一步地，所述聚类分析的方法为：
根据每个头部对应的矩形框的位置计算每个矩形框与其余矩形框之间的距离，将距离小于第一固定距离的所有矩形框存储在第一邻近矩形数组中；
计算每个矩形框的相邻矩形框数量，判断每个矩形框的相邻矩形框数量是否小于预设值，将相邻矩形框数量小于预设值的矩形框及其对应的所述第一邻近矩形数组删除以获得初步聚集区域；
对每个初步聚集区域进行矩形区域生长处理以获得最终的聚集区域。
进一步地，所述矩形区域生长处理的方法为：
遍历每个聚集区域，计算出每个聚集区域中拥有最多邻近矩形的目标矩形框，获取以当前聚集区域中所述目标矩形框中点为圆心，以第一固定距离为半径向外扩展区域内的矩形集合，遍历矩形合集以搜索距离小于等于第二固定距离的矩形框并将其存储在第二邻近矩形数组中；其中第二固定距离小于第一固定距离；
将每个聚集区域所对应的第一邻近矩形数组和第二邻近矩形数组进行并集运算以获得每个聚集区域的区域生长结果。
进一步地，所述第一固定距离＝(矩形框的宽度值的平方+矩形框的高度值的平方)*2.5，所述第二固定距离＝(矩形框的宽度值的平方+高度值的平方)*1.5。
进一步地，对于当前帧图像中的聚集区域的所述拥堵系数的计算方法为：
其中，len[L(n)]表示当前检测的矩形框数组大小，len(G
p)表示所有聚集区域人数总和，len(G
b)表示聚集区域总数目，β∈(0.4,0.5,0.6)，β取值与最大聚集区域内人数大小有关，分别对应([3,9),[9,12),[12,+∞))人数取值。
进一步地，对于当前帧图像中非聚集区域的所述拥堵系数的计算方法为：
Crd＝α*len[L(n)]；
其中，场景系数α为0.01≤α≤0.04，len[L(n)]表示当前检测的矩形框数组大小；当Crd超过1.0时，取值为1.0。
本公开的目的之二采用如下技术方案实现：
一种人群快速聚集行为的识别系统，执行如上述的人群快速聚集行为的识别方法，包括：
头部检测模块，用于获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；
人群分析模块，用于根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；当存在所述拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以所述目标帧图像为起点，在预设时间段内所有帧图像的所述拥堵系数及最大聚集区域人数；
判断输出模块，用于判断所述预设时间段内是否存在所述拥堵系数和所述最大聚集区域人数均超过其对应的预设阈值的帧图像，若存在，则输出人群快速聚集的结果。
本公开的目的之三采用如下技术方案实现：
一种电子设备，其包括处理器、存储器及存储于所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现上述的人群快速聚集行为的识别方法。
本公开的目的之四采用如下技术方案实现：
一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被执行时实现上述的人群快速聚集行为的识别方法。
(三)有益效果
本公开提出一种人群特征描述方法，利用头部识别方式来解决了人群分布密集时人体轮廓缺失严重带来的特征表达困难问题；本公开通过对每个聚集区域进行聚类处理，准确根据人群分布情况划分图像中的人群密集区域，提高人群快速聚集检测的准确性；同时，结合拥堵系数计算方法使得本公开方法具有更强的场景适应性，能同时适应多个场景需求，具有较高可行性。
图1为本公开人群快速聚集行为的识别方法的原理示意图；
图2为本公开人群快速聚集行为的识别方法流程示意图；
图3为本公开头部检测模型的神经网络结构示意图；
图4为本公开人群快速聚集行为的识别系统的模块示意图。
下面，结合附图以及具体实施方式，对本公开做进一步描述，需要说明的是，在不相冲突的前提下，以下描述的各实施例之间或各技术特征之间可以任意组合形成新的实施例。
实施例一
本实施例提供一种人群快速聚集行为的识别方法，本方法可应用在背景噪声较大、环境复杂多变的实际场景中，可同时适应多个场景需求，具有更强的适应性。
本实施例的人群快速聚集行为具体指的是在指定时间内人群从之前的稀疏并不拥堵的状态快速转变为拥堵状态，即人群突然聚集描述的是在某一段时间内，人群由最初的稀疏，甚至无人到迅速集结在某个或几个位置，且拥堵程度、聚集区域人员数目超过事先设定的阈值。
根据以上人群快速聚集的特征，参考图1、图2所示，本实施例提出一种人群快速聚集识别方法：
步骤S1：获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；
步骤S2：根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；
步骤S3：将所述拥堵系数与预设值进行比对，当存在所述拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以所述目标帧图像为起点，在预设时间段内所有帧图像的所述拥堵系数及最大聚集区域人数；判断所述预设时间段内是否存在所述拥堵系数和所述最大聚集区域人数均超过其对应的预设阈值的帧图像，若存在，则输出人群快速聚集的结果。
本实施例考虑人群在聚集时处于密集状态，人体轮廓信息缺失严重，而人头信息却大致完整，在密集的人群中人头特征是主要的人群信息描述特征；因此，本实施例使用基于行人头部检测方法来在描述 某帧图像中人群大致分部情况。本实施例在对当前帧图像进行分析之前，需要预先建立头部检测模型。
具体地，本实施例预先获取图像样本集，图像样本集包括有大量包含头部特征的图像，每张样本图像以1920*1080尺度为基准点，若样本图像超过1920*1080像素时，则需对图像样本进行裁剪并进行归一化处理后，再对每张样本图像中头部宽度大于20像素的目标标注出来，将经过上述处理后的图像样本作为卷积神经网络的训练样本进行学习和训练以构建出头部检测模型。
如图3所示，本实施例头部检测模型的卷积神经网络包括了conv0/relu到conv18/relu共19个卷积层和5个Pooling层，回归层分别由conv6、conv9、conv12、conv14、conv16、conv18层接relu后的输出组成，每层回归层都具有6个PriorBox。网络输入的图像大小为640*640，对输入的样本图像进行训练和学习后可获得一个包含背景和头部信息的头部检测模型
本实施例通过实际场景拍摄获得连续帧图像后，将每一帧都输入已经建好的头部检测模型中，头部检测模型检测到图像中存在头部时，为每个头部所在位置生成其对应的矩形框进行显示。
本实施例利用矩形框代表行人头部所在位置，当人群聚集时，将会呈现出多个矩形框彼此紧贴现象，因此只需考虑各个矩形框之间的关系便能描述人群分布情况。一般认为当超过3个人聚集在同一个区域时，存在拥堵的可能性较大，因此，本实施例以最少3个人组成一个聚集区域为基本点，计算整帧图像聚集区域个数及每个聚集区域内的头部数量。
本实施例中聚集区域G(Y)的获取原理如下：
一般认为人群中拥有最多邻近人数的目标人物分布在最密集的中心区域，若在计算聚集区域时，能找出拥有最多邻近矩形框的目标矩形框，该目标矩形框则最接近某个聚集区域的中心，再结合矩形区域生长处理，就能大致找出这个聚集区域内所有的矩形框。
本实施例中计算出拥有最多邻近矩形框的目标矩形框的方法为：
设由一组检测矩形框构成的数组为L(n)，当前矩形框表示为L(k)(k∈[0,n])，计算当前矩形框L(k)与L(k)以外的其余矩形框之间的距离，将距离小于第一固定距离的所有矩形框存在第一邻近矩形数组中，计算方法可表示为：
其中，|L(k)→L(k+m)|表示两个矩形框L(k)和L(k+m)中心点的距离，m∈[-k,0)∪(0,n-k]；Vec[L(k)]表示所有矩形框对应的邻近矩形框分布。
对Vec[L(k)]进行排序，就能找出拥有最多邻近矩形的目标矩形框，目标矩形框代表着聚集区域的中心。
所述第一固定距离是判断是否满足“邻近矩形框”的依据，具体地：在由一组检测矩形框组成的数组中，以两个矩形框中心点间的距离作为不同矩形框之间的距离S(x
1,x
2)，当需要获取与某个矩形框x
1邻近的其他矩形框数目时，需要结合第一固定距离进行计算，若一矩形框x
1与另一矩形框x
2之间的距离小于第一固定距离，则二者为邻近关系；若二者距离大于第一固定距离，则二者并非邻近关系。本实施例考虑到图像中因距离相机远近存在透视失真的因素，将第一固定距离固定为矩形框x
1的宽、高平方和的2.5倍，即所述第一固定距离＝(矩形框的宽度值的平方+矩形框的高度值的平方)*2.5，记为dis(L)，邻近矩形框的计算即是计算出满足与x
1中心点距离小于2.5倍x
1的宽、高平方和的所有矩形框。
在邻近矩形框的计算过程中使用了2.5倍宽、高平方和，即聚集区 域内目标矩形框到其他矩形框之间的距离均保持在第一固定距离范围内，若仍有其它矩形框分布在某个矩形框聚集区域的边缘，它应从属于这个聚集区域。为了进一步优化聚集区域，计算每个矩形框的相邻矩形框数量，判断每个矩形框的相邻矩形框数量是否小于预设值，将相邻矩形框数量小于预设值的矩形框及其对应的所述第一邻近矩形数组删除以获得初步聚集区域；本实施例中删除邻近矩形框数量少于2的矩形框及对应的第一邻近矩形数组。
其后，遍历初步聚集区域集合中的每一个聚集区域，再对每个聚集区域进行矩形区域生长处理以获得最终的聚集区域G(Y)。具体地，所述矩形区域生长处理的方法为：
步骤S21：遍历每个聚集区域，计算出每个聚集区域中拥有最多邻近矩形的目标矩形框后，获取以当前聚集区域中所述目标矩形框中点为圆心，以第一固定距离为半径向外扩展区域内的矩形集合；举个例子，假设其中一个聚集区域为Vec[L(b)]，具有最多邻近矩形框的目标矩形框是L(b)，计算Vec[L(b)]中以目标矩形框L(b)中点为圆心，以第一固定距离为半径的圆外侧分布的矩形集合，记为Vec
0[L(b)]。
步骤S22：遍历矩形合集Vec
0[L(b)]每一个元素，并从L(n)中搜索距离小于等于第二固定距离的矩形框，并将其存储在第二邻近矩形数组Vec
1[L(b)]中；其中第二固定距离小于第一固定距离，本实施例中第二固定距离为矩形框的宽、高平方和的1.5倍，即所述第二固定距离＝(矩形框的宽度值的平方+高度值的平方)*1.5。
步骤S23：求出第一邻近矩形数组Vec[L(b)]和第二邻近矩形数组Vec
1[L(b)]并集构成的矩形框Vec
last[L(b)]，即为某个聚集区域的矩形区域生长结果。
把所有聚集区域经过矩形区域生长处理后的结果记为最终的聚集区域G(Y)，而最终的聚集区域中的头部数量则为Y(X)，而头部数量可直接通过头部检测模型检测到的头部矩形框数量来体现。
本实例中拥堵系数的计算分为两种情况进行计算，即存在聚集区域和不存在聚集区域；具体地：
当人群分布离散，不存在聚集区域时，人群在环境中每个位置出现的概率近似相同，即人群符合均匀分布；基于该特征，本实施例引入场景系数α，其表示场景中某个位置出现人员与整个场景满员的比率，当场景较为开阔时，α值偏小，当场景视野较窄时，α值偏大；本实施例经过多次试验后表明0.01≤α≤0.04最佳，因此，当前帧图像不存在聚集区域的拥堵系数计算方法如下：
Crd＝α*len[L(n)]；
其中，场景系数α为0.01≤α≤0.04，len[L(n)]表示当前检测的矩形框数组大小；当Crd超过1.0时，取值为1.0。
当场景中出现明显的人群聚集区域时，那么该环境渐渐地表现为拥挤，拥挤的位置相比离散人群分布的位置将变得更为重要，因此，本实施例当前帧图像存在聚集区域时的拥堵系数计算方法为：
其中，len[L(n)]表示当前检测的矩形框数组大小，len(G
p)表示所有聚集区域人数总和，len(G
b)表示聚集区域总数目，β∈(0.4,0.5,0.6)，β取值与最大聚集区域内人数大小有关，分别对应([3,9),[9,12),[12,+∞))人数取值。
本实施例计算出拥堵系数后，若在预设时间段内拥堵系数增长较大，且聚集区域内人数超过预设阈值，则可认为这是一种人群快速聚集行为。因此，本实施例经过大量场景的试验，找到一个合适的表达人群稀疏的系数值，其稀疏系数取0.3较合理，即小于这个数值时，不计入快速聚集的时间之内。参考图2所示，当拥堵系数大于稀疏系数时开始计时，计算并更新预设时间段内所有帧的拥堵系数及最大聚集区域人数；若在预设时间段内拥堵系数和最大聚集区域人数均未超过事先设定的预设阈值，那么就认为这段预设时间段内人群没有快速聚集，则可以继续对下一帧图像进行头部检测以对下一帧图像的人群聚集行为进行分析。若在预设时间段内拥堵系数和最大聚集区域人数在某一帧均超过预设阈值，则代表着存在人群快速聚集的行为，由于当某一帧图像存在人群聚集行为后下一帧图像中人群无法马上散去，因此，在进行人群快速聚集行为判断时，还可考虑剩余的帧内是否满足拥堵系数和最大聚集区域人数都超过预设阈值的条件，若有，则可认为人群在拥堵系数和最大聚集区域人数都超过预设阈值的帧图像所对应的时间内快速聚集。
实施例二
如图4所示，本实施例提供一种人群快速聚集行为的识别系统，执行如实施例一所述的人群快速聚集行为的识别方法，包括：
头部检测模块，用于获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；
人群分析模块，用于根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；当存在所述拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以所述目标帧图像为起点，在预设时间段内所有帧图像的所述拥堵系数及最大聚集区域人数；
判断输出模块，用于判断所述预设时间段内是否存在所述拥堵系数和所述最大聚集区域人数均超过其对应的预设阈值的帧图像，若存 在，则输出人群快速聚集的结果。
本实施例利用Box-Gather(检测框聚集)方法对人头检测结果处理，获取聚集区域及每个区域内的人头数目；并采用拥堵系数计算方法，利用检测所得的头部数量计算整张图像的拥挤程度(0～1)；在一段指定时间内，当拥挤程度超过某个阈值，且聚集区域内的人数超过事先设定的阈值时，则认为人群快速聚集。本实施例可适应背景噪声较大、环境复杂多变的场景，大幅度提高人群聚集行为识别的适应性。
本实施例提供一种电子设备，其包括处理器、存储器及存储于所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现实施例一中的人群快速聚集行为的识别方法；另外，本实施例还提供一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被执行时实现上述的人群快速聚集行为的识别方法。
本实施例中的系统、设备及存储介质与前述实施例中的方法是基于同一发明构思下的多个方面，在前面已经对方法实施过程作了详细的描述，所以本领域技术人员可根据前述描述清楚地了解本实施例中的系统、设备及存储介质的结构及实施过程，为了说明书的简洁，在此就不再赘述。
上述实施方式仅为本公开的优选实施方式，不能以此来限定本公开保护的范围，本领域的技术人员在本公开的基础上所做的任何非实质性的变化及替换均属于本公开所要求保护的范围。
本公开提供的人群快速聚集行为的识别方法，通过对每个聚集区域进行聚类处理，准确根据人群分布情况划分图像中的人群密集区域，且结合拥堵系数计算方法使得不仅具有更强的场景适应性，能同时适应多个场景需求，也可以提高人群快速聚集检测的准确性，具有很强的工业实用性。
Claims (10)
- 一种人群快速聚集行为的识别方法，其特征在于，包括：获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；当存在所述拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以所述目标帧图像为起点，在预设时间段内所有帧图像的所述拥堵系数及最大聚集区域人数；判断所述预设时间段内是否存在所述拥堵系数和所述最大聚集区域人数均超过其对应的预设阈值的帧图像，若存在，则输出人群快速聚集的结果。
- 根据权利要求1所述的人群快速聚集行为的识别方法，其特征在于，所述头部检测模型是通过标注有行人头部的图像样本集对神经网络进行训练获得；并在所述头部检测模型检测到图像中存在头部时，为每个头部所在位置生成其对应的矩形框进行显示。
- 根据权利要求2所述的人群快速聚集行为的识别方法，其特征在于，所述聚类分析的方法为：根据每个头部对应的矩形框的位置计算每个矩形框与其余矩形框之间的距离，将距离小于第一固定距离的所有矩形框存储在第一邻近矩形数组中；计算每个矩形框的相邻矩形框数量，判断每个矩形框的相邻矩形框数量是否小于预设值，将相邻矩形框数量小于预设值的矩形框及其对应的所述第一邻近矩形数组删除以获得初步聚集区域；对每个初步聚集区域进行矩形区域生长处理以获得最终的聚集区域。
- 根据权利要求3所述的人群快速聚集行为的识别方法，其特征在 于，所述矩形区域生长处理的方法为：遍历每个所述聚集区域，计算出每个所述聚集区域中拥有最多邻近矩形的目标矩形框，获取以当前聚集区域中所述目标矩形框中点为圆心，以所述第一固定距离为半径向外扩展区域内的矩形集合，遍历矩形合集以搜索距离小于等于第二固定距离的矩形框并将其存储在第二邻近矩形数组中；其中所述第二固定距离小于所述第一固定距离；将每个所述聚集区域所对应的所述第一邻近矩形数组和所述第二邻近矩形数组进行并集运算以获得每个所述聚集区域的区域生长结果。
- 根据权利要求4所述的人群快速聚集行为的识别方法，其特征在于，所述第一固定距离＝(矩形框的宽度值的平方+矩形框的高度值的平方)*2.5，所述第二固定距离＝(矩形框的宽度值的平方+高度值的平方)*1.5。
- 根据权利要求1所述的人群快速聚集行为的识别方法，其特征在于，对于当前帧图像中的所述聚集区域的所述拥堵系数的计算方法为：
其中，len[L(n)]表示当前检测的矩形框数组大小，len(G p)表示所有聚集区域人数总和，len(G b)表示聚集区域总数目，β∈(0.4,0.5,0.6)，β取值与最大聚集区域内人数大小有关，分别对应[3,9),[9,12),[12,+∞)人数取值。 - 根据权利要求5所述的人群快速聚集行为的识别方法，其特征在于，对于当前帧图像中非聚集区域的所述拥堵系数的计算方法为：Crd＝α*len[L(n)]；其中，场景系数α为0.01≤α≤0.04，len[L(n)]表示当前检测的矩形框 数组大小；当Crd超过1.0时，取值为1.0。
- 一种人群快速聚集行为的识别系统，其特征在于，执行如权利要求1～7任意一项所述的人群快速聚集行为的识别方法，包括：头部检测模块，用于获取连续帧图像，基于头部检测模型依次对每一帧图像进行头部检测；人群分析模块，用于根据头部检测结果对整帧图像的聚集区域进行聚类分析，计算每个聚类后的聚集区域内的人头数量，并对每一帧图像对应的拥堵系数进行计算；当存在所述拥堵系数大于预设稀疏系数的目标帧图像时，计算并更新以所述目标帧图像为起点，在预设时间段内所有帧图像的所述拥堵系数及最大聚集区域人数；判断输出模块，用于判断所述预设时间段内是否存在所述拥堵系数和所述最大聚集区域人数均超过其对应的预设阈值的帧图像，若存在，则输出人群快速聚集的结果。
- 一种电子设备，其特征在于，其包括处理器、存储器及存储于所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现权利要求1～7任一所述的人群快速聚集行为的识别方法。
- 一种计算机可读存储介质，其特征在于，其上存储有计算机程序，所述计算机程序被执行时实现权利要求1～7任一所述的人群快速聚集行为的识别方法。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210147617.0A CN114627406B (zh) | 2022-02-17 | 2022-02-17 | 一种人群快速聚集行为的识别方法、系统、设备及介质 |
| CN202210147617.0 | 2022-02-17 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2023155482A1 true WO2023155482A1 (zh) | 2023-08-24 |
Family
ID=81900375
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2022/128971 Ceased WO2023155482A1 (zh) | 2022-02-17 | 2022-11-01 | 一种人群快速聚集行为的识别方法、系统、设备及介质 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN114627406B (zh) |
| WO (1) | WO2023155482A1 (zh) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117156259A (zh) * | 2023-10-30 | 2023-12-01 | 海信集团控股股份有限公司 | 一种视频流获取方法及电子设备 |
| CN117197734A (zh) * | 2023-08-30 | 2023-12-08 | 西北工业大学 | 一种基于图结构的密集人群对应关系判读方法 |
| CN118552588A (zh) * | 2024-07-24 | 2024-08-27 | 西安高商智能科技有限责任公司 | 一种电阻阵列红外成像的目标模型构建方法及系统 |
| WO2025148631A1 (zh) * | 2024-01-08 | 2025-07-17 | 中电科大数据研究院有限公司 | 一种识别人群聚集的方法、装置及计算机可读存储介质 |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114627406B (zh) * | 2022-02-17 | 2025-07-08 | 广电运通集团股份有限公司 | 一种人群快速聚集行为的识别方法、系统、设备及介质 |
| CN117523653A (zh) * | 2022-07-29 | 2024-02-06 | 青岛海信电子技术服务有限公司 | 异常聚集检测方法、装置、设备和介质 |
| CN116189073A (zh) * | 2022-12-20 | 2023-05-30 | 北京无线电计量测试研究所 | 一种聚集行为识别的方法及装置 |
| CN116129522A (zh) * | 2022-12-30 | 2023-05-16 | 成都云天励飞技术有限公司 | 一种行为识别方法、装置、电子设备及计算机存储介质 |
| CN116206421B (zh) * | 2023-01-04 | 2024-06-04 | 歌尔股份有限公司 | 人群聚集的风险预警方法、装置、电子设备及介质 |
| CN116129361B (zh) * | 2023-03-24 | 2023-08-08 | 武汉中科通达高新技术股份有限公司 | 一种基于距离度量的人群聚集识别方法及装置 |
| CN119541860A (zh) * | 2024-11-21 | 2025-02-28 | 北京一石数智健康科技有限公司 | 智能化体育场馆健康数据采集设备 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2011076316A (ja) * | 2009-09-30 | 2011-04-14 | Fujifilm Corp | 群衆監視装置および方法ならびにプログラム |
| CN107341480A (zh) * | 2017-07-12 | 2017-11-10 | 中国电子科技集团公司第二十八研究所 | 一种改进型pccnn神经网络模型的人群聚集检测方法 |
| CN111242096A (zh) * | 2020-02-26 | 2020-06-05 | 贵州安防工程技术研究中心有限公司 | 基于人数梯度的人群聚集判别方法与系统 |
| CN112560807A (zh) * | 2021-02-07 | 2021-03-26 | 南京云创大数据科技股份有限公司 | 一种基于人头检测的人群聚集检测方法 |
| CN114627406A (zh) * | 2022-02-17 | 2022-06-14 | 广州广电运通金融电子股份有限公司 | 一种人群快速聚集行为的识别方法、系统、设备及介质 |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101325690A (zh) * | 2007-06-12 | 2008-12-17 | 上海正电科技发展有限公司 | 监控视频流中人流分析与人群聚集过程的检测方法及系统 |
| CN103839065B (zh) * | 2014-02-14 | 2017-02-01 | 南京航空航天大学 | 人群动态聚集特征提取方法 |
| CN107729799A (zh) * | 2017-06-13 | 2018-02-23 | 银江股份有限公司 | 基于深度卷积神经网络的人群异常行为视觉检测及分析预警系统 |
| CN107862437B (zh) * | 2017-10-16 | 2022-02-22 | 中国人民公安大学 | 基于风险概率评估的公共区域人群聚集预警方法及系统 |
| CN109446989A (zh) * | 2018-10-29 | 2019-03-08 | 上海七牛信息技术有限公司 | 人群聚集检测方法、装置及存储介质 |
| CN111291587A (zh) * | 2018-12-06 | 2020-06-16 | 深圳光启空间技术有限公司 | 一种基于密集人群的行人检测方法、存储介质及处理器 |
| CN111738225B (zh) * | 2020-07-29 | 2020-12-11 | 中电科新型智慧城市研究院有限公司 | 人群聚集检测方法、装置、设备及存储介质 |
-
2022
- 2022-02-17 CN CN202210147617.0A patent/CN114627406B/zh active Active
- 2022-11-01 WO PCT/CN2022/128971 patent/WO2023155482A1/zh not_active Ceased
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2011076316A (ja) * | 2009-09-30 | 2011-04-14 | Fujifilm Corp | 群衆監視装置および方法ならびにプログラム |
| CN107341480A (zh) * | 2017-07-12 | 2017-11-10 | 中国电子科技集团公司第二十八研究所 | 一种改进型pccnn神经网络模型的人群聚集检测方法 |
| CN111242096A (zh) * | 2020-02-26 | 2020-06-05 | 贵州安防工程技术研究中心有限公司 | 基于人数梯度的人群聚集判别方法与系统 |
| CN112560807A (zh) * | 2021-02-07 | 2021-03-26 | 南京云创大数据科技股份有限公司 | 一种基于人头检测的人群聚集检测方法 |
| CN114627406A (zh) * | 2022-02-17 | 2022-06-14 | 广州广电运通金融电子股份有限公司 | 一种人群快速聚集行为的识别方法、系统、设备及介质 |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117197734A (zh) * | 2023-08-30 | 2023-12-08 | 西北工业大学 | 一种基于图结构的密集人群对应关系判读方法 |
| CN117156259A (zh) * | 2023-10-30 | 2023-12-01 | 海信集团控股股份有限公司 | 一种视频流获取方法及电子设备 |
| CN117156259B (zh) * | 2023-10-30 | 2024-03-22 | 海信集团控股股份有限公司 | 一种视频流获取方法及电子设备 |
| WO2025148631A1 (zh) * | 2024-01-08 | 2025-07-17 | 中电科大数据研究院有限公司 | 一种识别人群聚集的方法、装置及计算机可读存储介质 |
| CN118552588A (zh) * | 2024-07-24 | 2024-08-27 | 西安高商智能科技有限责任公司 | 一种电阻阵列红外成像的目标模型构建方法及系统 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114627406A (zh) | 2022-06-14 |
| CN114627406B (zh) | 2025-07-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2023155482A1 (zh) | 一种人群快速聚集行为的识别方法、系统、设备及介质 | |
| CN104751491B (zh) | 一种人群跟踪及人流量统计方法及装置 | |
| CN110532970B (zh) | 人脸2d图像的年龄性别属性分析方法、系统、设备和介质 | |
| CN107622258B (zh) | 一种结合静态底层特征和运动信息的快速行人检测方法 | |
| CN109345547B (zh) | 基于深度学习多任务网络的交通车道线检测方法及装置 | |
| CN104978567B (zh) | 基于场景分类的车辆检测方法 | |
| CN106570486A (zh) | 基于特征融合和贝叶斯分类的核相关滤波目标跟踪方法 | |
| CN110309765B (zh) | 一种视频运动目标高效检测方法 | |
| WO2020094088A1 (zh) | 一种图像抓拍方法、监控相机及监控系统 | |
| CN105335716A (zh) | 一种基于改进udn提取联合特征的行人检测方法 | |
| CN111310662A (zh) | 一种基于集成深度网络的火焰检测识别方法及系统 | |
| CN101196991A (zh) | 密集客流计数和行人步行速度自动检测方法及系统 | |
| CN111027370A (zh) | 一种多目标跟踪及行为分析检测方法 | |
| CN116030396B (zh) | 一种用于视频结构化提取的精确分割方法 | |
| CN107590486B (zh) | 运动目标识别方法和系统、自行车流量统计方法和设备 | |
| CN112084928A (zh) | 基于视觉注意力机制和ConvLSTM网络的道路交通事故检测方法 | |
| CN104599291B (zh) | 基于结构相似度和显著性分析的红外运动目标检测方法 | |
| CN113706481A (zh) | 精子质量检测方法、装置、计算机设备和存储介质 | |
| CN105957356A (zh) | 一种基于行人数量的交通控制系统及方法 | |
| CN109063630B (zh) | 一种基于可分离卷积技术和帧差补偿策略的快速车辆检测方法 | |
| CN112465854A (zh) | 基于无锚点检测算法的无人机跟踪方法 | |
| CN117911965B (zh) | 一种基于航拍图像的高速公路交通事故识别方法及设备 | |
| CN115512263B (zh) | 一种面向高空坠物的动态视觉监测方法及装置 | |
| CN110807430B (zh) | 一种实况全景交通标志图片预处理方法 | |
| CN108960181B (zh) | 基于多尺度分块lbp和隐马尔科夫模型的黑烟车检测方法 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22926786 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 22926786 Country of ref document: EP Kind code of ref document: A1 |