[go: up one dir, main page]

CN117953009A - A group personnel trajectory prediction method based on spatiotemporal characteristics - Google Patents

A group personnel trajectory prediction method based on spatiotemporal characteristics Download PDF

Info

Publication number
CN117953009A
CN117953009A CN202311864476.5A CN202311864476A CN117953009A CN 117953009 A CN117953009 A CN 117953009A CN 202311864476 A CN202311864476 A CN 202311864476A CN 117953009 A CN117953009 A CN 117953009A
Authority
CN
China
Prior art keywords
target
people
vector
spatiotemporal
trajectory
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.)
Pending
Application number
CN202311864476.5A
Other languages
Chinese (zh)
Inventor
刘宏宇
徐鑫
陈�胜
魏超凡
彭一帆
宋志雄
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Aerospace Science And Engineering Intelligent Operation Research And Information Security Research Institute Wuhan Co ltd
Original Assignee
Aerospace Science And Engineering Intelligent Operation Research And Information Security Research Institute Wuhan Co ltd
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 Aerospace Science And Engineering Intelligent Operation Research And Information Security Research Institute Wuhan Co ltd filed Critical Aerospace Science And Engineering Intelligent Operation Research And Information Security Research Institute Wuhan Co ltd
Priority to CN202311864476.5A priority Critical patent/CN117953009A/en
Publication of CN117953009A publication Critical patent/CN117953009A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20092—Interactive image processing based on input by user
    • G06T2207/20104—Interactive definition of region of interest [ROI]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • G06T2207/30201—Face
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

本发明属于视频识别技术领域，具体涉及一种基于时空特征的群体性人员轨迹预测方法。所述方法包括：步骤1：多目标轨迹识别；步骤2：时空权重向量构建；步骤3：特殊人员权重向量构建；步骤4：多目标轨迹预测分析。与现有技术相比较，本发明效果如下：(1)实现群体性人员的轨迹预测，并针对于不同时间、空间、人员进行个性化建模。(2)在公共区域与内部区域，采用不同的部署实施策略，实现群体性事件防范由被动处置向主动发现转变，提高风险预测、预警分析、人群疏导能力。(3)通过数据积累，自适应训练模型，不断提升模型的预测准确性。

The present invention belongs to the field of video recognition technology, and specifically relates to a method for predicting the trajectory of a group of people based on spatiotemporal features. The method comprises: step 1: multi-target trajectory recognition; step 2: spatiotemporal weight vector construction; step 3: special personnel weight vector construction; step 4: multi-target trajectory prediction analysis. Compared with the prior art, the effects of the present invention are as follows: (1) The trajectory prediction of a group of people is realized, and personalized modeling is performed for different times, spaces, and personnel. (2) Different deployment and implementation strategies are adopted in public areas and internal areas to realize the transformation of mass incident prevention from passive disposal to active discovery, and improve risk prediction, early warning analysis, and crowd guidance capabilities. (3) Through data accumulation and adaptive training of models, the prediction accuracy of the model is continuously improved.

Description

Space-time feature-based crowd personnel trajectory prediction method
Technical Field
The invention belongs to the technical field of video identification, and particularly relates to a method for predicting a group personnel track based on space-time characteristics.
Background
With the development of socioeconomic performance and the deepening of urbanization, the population scale of cities rises year by year, and people gathering and congestion are easy to occur. People gathering has contingency and burstiness, and how to timely and accurately predict the track of the crowd personnel becomes a key for preventing the crowd events. In recent years, with the rapid development of smart cities and artificial intelligence technologies, cameras are distributed over streets and alleys of the cities, and video monitoring can cover a large number of public areas such as streets and the like, so that hardware foundation and data support are provided for crowd personnel track prediction. Through a CV (Computer Vision) algorithm, a Computer can automatically analyze video content without manually observing massive videos. The CV algorithm can locate and track a plurality of targets and forecast target tracks, so that the analysis, forecast and control capacity of the relevant management departments on the group behaviors can be improved, and the urban order and safety are ensured.
At present, although the multi-target tracking algorithm can identify people in a video and draw a personnel track, the influence of specific space-time conditions (such as rush hour, bus stop and the like) on the personnel track is not considered, the influence of special people (such as public people) on the crowd track is not considered, and the crowd personnel track cannot be analyzed aiming at a specific scene.
Disclosure of Invention
First, the technical problem to be solved
The invention aims to solve the technical problems that: how to provide a method for predicting the track of population personnel based on space-time characteristics.
(II) technical scheme
In order to solve the technical problems, the invention provides a method for predicting a population personnel track based on space-time characteristics, which comprises the following steps:
step 1: multi-target track identification;
Step 2: constructing space-time weight vectors;
Step 3: constructing a special personnel weight vector;
Step 4: and (5) multi-target track prediction analysis.
In the step 1, multiple target tracks are identified; the step 1 comprises the following steps:
Step 11: target identification;
Step 12: tracking a target;
Step 13: and (5) vectorizing the representation of the target.
In the step 11, the target is identified;
In image recognition, the convolutional neural network algorithm is based on the best performance; in the method, the recognition of personnel is focused; the method comprises the steps of analyzing video streams frame by frame through a multi-layer convolutional neural network, identifying and positioning personnel in video frames, wherein for an input video V, V= { V 1,V2,……,Vt }, t is a frame at different moments;
The construction of the target recognition model is divided into two stages of model training and model testing; in the training stage, inputting video data V train with labeling information; the labeling content is a plurality of rectangular frames and is expressed as (x, y, h, w), wherein x is the left upper corner abscissa of the labeling frame, y is the left upper corner ordinate of the labeling frame, h is the height of the labeling frame, and w is the width of the labeling frame; after training and model convergence, a target recognition model is obtained; in the test stage, inputting video data V test without marking information, and testing the model effect; if the accuracy requirement is met, finishing model training; otherwise, the parameters are adjusted and retrained until the accuracy requirement is met.
Wherein, in the step 12, the target is tracked;
Arranging the target frame information of the video frame by frame according to a time sequence to obtain target track information; because the image of the adjacent frames in the video has little change, the operation cost of the frame-by-frame analysis is too high, so the video frames are sampled;
The standard video is 24 frames per second, and the sampling ratio value is in the method
Wherein, in the step 13, the target vectorization represents;
For each target O i, the state at time t 0 is represented as The displacement vector at time t thereafter isBy state variableRepresents the position of the target O i by displacement vectorThe direction and speed of movement of the object O i are indicated, describing the static and dynamic characteristics of the object.
In the step 2, a space-time weight vector is constructed;
The step 2 comprises the following steps:
step 21: constructing a time weight vector;
step 22: constructing a space weight vector;
Step 23: and (5) space-time vector fusion.
In the step 21, a time weight vector is constructed;
Considering life work and rest of people, the movement trend of people flow often has a periodic rule; for example, the mass flow of people in the morning is generally higher than at night; the difference of the human flow in different time periods leads to the difference of the reference dimension; therefore, different weight parameters need to be trained for the people stream prediction models of different time periods for such variability;
continuously sampling the people stream characteristics to obtain the total displacement vector Determining the number of t and the time interval of each sampling according to the sampling frequency; for example, if the sampling frequency is set to be abnormal for 5 minutes, the sampling time is 00:00,00:05, … …,23:55, and the sampling frequency is 288;
Time weighting Wherein ε is a small constant, prevent zero removal;
step 22: constructing a space weight vector;
the specific places can also influence the traffic flow, for example, people can be on buses such as buses, and people tend to show an aggregation trend; if such an effect is not eliminated, misjudgment of the personnel gathering event is likely to be caused;
For a specific observation point x, counting the number n of people appearing in the observation point x, and then weighting the observation point x in space Wherein ε is a small constant, prevent zero removal;
Step 23: space-time vector fusion;
For different observation points x of different time periods t, the time-space weight vector is as follows
Where w i,j is the product of the time vector and the space vector.
In the step 3, a special personnel weight vector is constructed;
Special personnel (such as public figures) can have a great influence on the track of the crowd, the crowd is easy to approach to the public figures, the density of people around the special personnel is increased, and the moving speed is reduced; the special personnel go out easily to cause a swarm event, and a great challenge is caused to security; the travel of special personnel is a scene which has to be considered, but is also a small probability event, and is applicable only in special scenes;
The step 3 comprises the following steps:
Step 31: constructing a special personnel library;
According to actual conditions, a face library of special personnel is established at different observation points, and a global face library can also be established; establishing a face library, and storing a unique ID (identity) of a person, a face picture, face characteristics and weight vectors;
Step 32: constructing a special personnel weight vector;
Special personnel have a certain attraction effect on surrounding people, and attraction weights are constructed according to the attraction effect intensity; estimating attractive weight of the special personnel by classifying the special personnel; because the number of special personnel is relatively small, the workload of constructing the special personnel weight vector is low.
In the step 4, multi-target track prediction analysis is performed;
The whole flow of the step 4 is as follows:
Firstly, acquiring a state variable of a target O i through a multi-target track recognition algorithm based on a convolutional neural network Displacement vectorSequence of state variables And a sequence of displacement vectorsMultiplying the space-time weight vector, inputting a sequence prediction model, and predicting the state variable sequence/>, in a next period of timeAnd displacement vector sequence
Then, recognizing a face in the video by adopting a face recognition algorithm, and checking whether special personnel appear; according to the detection result, two conditions are classified;
case one: special personnel are present; according to the position of the special person at the moment t A special person appeal weight w; correcting the original output result, wherein the predicted position of the target O i at the time t isCorrection vector isThe predicted result is the state variable sequence And a sequence of displacement vectors
And a second case: no special personnel exist; the predicted result is the state variable sequence And displacement vector sequence
In the step 4, model iteration is performed;
Because the actual crowd is dynamically changed, three model components in the method need to be iterated continuously according to the actual data so as to adapt to the change;
a. Track prediction model
At intervals, data sampling is carried out, new data are added into an original training set for retraining, the scale of the training set is continuously enlarged, and the accuracy of the model is improved;
b. Space-time vector weights
Sampling data at intervals; calculating space-time vectors according to the new data to ensure that the model can adapt to changing conditions;
c. Special personnel store
According to actual conditions, a special personnel library is continuously expanded, and the level is adjusted;
The deployment of the step 4 is implemented as follows:
The method has two deployment application scenarios;
a. Public place deployment
The method is oriented to public areas such as streets, has high mobility, and does not need to establish a special personnel library; the data of the space-time vector mainly comes from the morning and evening peaks;
b. Interior area deployment
The method is oriented to internal areas such as a park, people are relatively fixed, a relatively perfect internal personnel library can be established, and key personnel are extracted to be added into a special personnel library; and, space-time vector modeling can be performed according to a specific place.
(III) beneficial effects
The invention provides a method for predicting a group personnel track based on space-time characteristics, which comprises the steps of constructing a space-time characteristic weight vector and a special personnel weight vector, reallocating the weights of a multi-target track vector, introducing the influence of specific time, place and personnel into a multi-target track prediction algorithm, analyzing the influence of individuals and environments on individual tracks, and realizing the timely prediction of the group events and timely finding and removing risks.
Compared with the prior art, the invention has the following effects:
(1) The track prediction of the crowd personnel is realized, and personalized modeling is carried out aiming at different time, space and personnel.
(2) Different deployment implementation strategies are adopted in the public area and the internal area, so that the transition from passive treatment to active discovery of crowd prevention is realized, and the risk prediction, early warning analysis and crowd dispersion capacity are improved.
(3) And through data accumulation, the model is self-adaptively trained, and the prediction accuracy of the model is continuously improved.
Drawings
FIG. 1 is a block diagram of a multi-objective personnel trajectory prediction method.
Detailed Description
To make the objects, contents and advantages of the present invention more apparent, the following detailed description of the present invention will be given with reference to the accompanying drawings and examples.
In order to solve the technical problems, the invention provides a method for predicting a population personnel track based on space-time characteristics, which comprises the following steps:
step 1: multi-target track identification;
Step 2: constructing space-time weight vectors;
Step 3: constructing a special personnel weight vector;
Step 4: and (5) multi-target track prediction analysis.
In the step 1, multiple target tracks are identified; the step 1 comprises the following steps:
Step 11: target identification;
Step 12: tracking a target;
Step 13: and (5) vectorizing the representation of the target.
In the step 11, the target is identified;
In image recognition, the convolutional neural network algorithm is based on the best performance; in the method, the recognition of personnel is focused; the method comprises the steps of analyzing video streams frame by frame through a multi-layer convolutional neural network, identifying and positioning personnel in video frames, wherein for an input video V, V= { V 1,V2,……,Vt }, t is a frame at different moments;
The construction of the target recognition model is divided into two stages of model training and model testing; in the training stage, inputting video data V train with labeling information; the labeling content is a plurality of rectangular frames and is expressed as (x, y, h, w), wherein x is the left upper corner abscissa of the labeling frame, y is the left upper corner ordinate of the labeling frame, h is the height of the labeling frame, and w is the width of the labeling frame; after training and model convergence, a target recognition model is obtained; in the test stage, inputting video data V test without marking information, and testing the model effect; if the accuracy requirement is met, finishing model training; otherwise, the parameters are adjusted and retrained until the accuracy requirement is met.
Wherein, in the step 12, the target is tracked;
Arranging the target frame information of the video frame by frame according to a time sequence to obtain target track information; because the image of the adjacent frames in the video has little change, the operation cost of the frame-by-frame analysis is too high, so the video frames are sampled;
The standard video is 24 frames per second, and the sampling ratio value is in the method
Wherein, in the step 13, the target vectorization represents;
For each target O i, the state at time t 0 is represented as The displacement vector at time t thereafter isBy state variableRepresents the position of the target O i by displacement vectorThe direction and speed of movement of the object O i are indicated, describing the static and dynamic characteristics of the object.
In the step 2, a space-time weight vector is constructed;
The step 2 comprises the following steps:
step 21: constructing a time weight vector;
step 22: constructing a space weight vector;
Step 23: and (5) space-time vector fusion.
In the step 21, a time weight vector is constructed;
Considering life work and rest of people, the movement trend of people flow often has a periodic rule; for example, the mass flow of people in the morning is generally higher than at night; the difference of the human flow in different time periods leads to the difference of the reference dimension; therefore, different weight parameters need to be trained for the people stream prediction models of different time periods for such variability;
continuously sampling the people stream characteristics to obtain the total displacement vector Determining the number of t and the time interval of each sampling according to the sampling frequency; for example, if the sampling frequency is set to be abnormal for 5 minutes, the sampling time is 00:00,00:05, … …,23:55, and the sampling frequency is 288;
Time weighting Wherein ε is a small constant, prevent zero removal;
step 22: constructing a space weight vector;
the specific places can also influence the traffic flow, for example, people can be on buses such as buses, and people tend to show an aggregation trend; if such an effect is not eliminated, misjudgment of the personnel gathering event is likely to be caused;
For a specific observation point x, counting the number n of people appearing in the observation point x, and then weighting the observation point x in space Wherein ε is a small constant, prevent zero removal;
Step 23: space-time vector fusion;
For different observation points x of different time periods t, the time-space weight vector is as follows
Where w i,j is the product of the time vector and the space vector.
In the step 3, a special personnel weight vector is constructed;
Special personnel (such as public figures) can have a great influence on the track of the crowd, the crowd is easy to approach to the public figures, the density of people around the special personnel is increased, and the moving speed is reduced; the special personnel go out easily to cause a swarm event, and a great challenge is caused to security; the travel of special personnel is a scene which has to be considered, but is also a small probability event, and is applicable only in special scenes;
The step 3 comprises the following steps:
Step 31: constructing a special personnel library;
According to actual conditions, a face library of special personnel is established at different observation points, and a global face library can also be established; establishing a face library, and storing a unique ID (identity) of a person, a face picture, face characteristics and weight vectors;
Step 32: constructing a special personnel weight vector;
Special personnel have a certain attraction effect on surrounding people, and attraction weights are constructed according to the attraction effect intensity; estimating attractive weight of the special personnel by classifying the special personnel; because the number of special personnel is relatively small, the workload of constructing the special personnel weight vector is low.
In the step 4, multi-target track prediction analysis is performed;
The whole flow of the step 4 is as follows:
Firstly, acquiring a state variable of a target O i through a multi-target track recognition algorithm based on a convolutional neural network Displacement vectorSequence of state variables And a sequence of displacement vectorsMultiplying the space-time weight vector, inputting a sequence prediction model, and predicting the state variable sequence/>, in a next period of timeAnd displacement vector sequence
Then, recognizing a face in the video by adopting a face recognition algorithm, and checking whether special personnel appear; according to the detection result, two conditions are classified;
case one: special personnel are present; according to the position of the special person at the moment t A special person appeal weight w; correcting the original output result, wherein the predicted position of the target O i at the time t isCorrection vector isThe predicted result is the state variable sequence And a sequence of displacement vectors
And a second case: no special personnel exist; the predicted result is the state variable sequence And displacement vector sequence
In the step 4, model iteration is performed;
Because the actual crowd is dynamically changed, three model components in the method need to be iterated continuously according to the actual data so as to adapt to the change;
a. Track prediction model
At intervals, data sampling is carried out, new data are added into an original training set for retraining, the scale of the training set is continuously enlarged, and the accuracy of the model is improved;
b. Space-time vector weights
Sampling data at intervals; calculating space-time vectors according to the new data to ensure that the model can adapt to changing conditions;
c. Special personnel store
According to actual conditions, a special personnel library is continuously expanded, and the level is adjusted;
The deployment of the step 4 is implemented as follows:
The method has two deployment application scenarios;
a. Public place deployment
The method is oriented to public areas such as streets, has high mobility, and does not need to establish a special personnel library; the data of the space-time vector mainly comes from the morning and evening peaks;
b. Interior area deployment
The method is oriented to internal areas such as a park, people are relatively fixed, a relatively perfect internal personnel library can be established, and key personnel are extracted to be added into a special personnel library; and, space-time vector modeling can be performed according to a specific place.
The foregoing is merely a preferred embodiment of the present invention, and it should be noted that modifications and variations could be made by those skilled in the art without departing from the technical principles of the present invention, and such modifications and variations should also be regarded as being within the scope of the invention.

Claims (10)

1.一种基于时空特征的群体性人员轨迹预测方法，其特征在于，所述方法包括：1. A method for predicting the trajectory of a group of people based on spatiotemporal features, characterized in that the method includes: 步骤1：多目标轨迹识别；Step 1: Multi-target trajectory recognition; 步骤2：时空权重向量构建；Step 2: Constructing the spatiotemporal weight vector; 步骤3：特殊人员权重向量构建；Step 3: Constructing weight vectors for special personnel; 步骤4：多目标轨迹预测分析。Step 4: Multi-target trajectory prediction analysis. 2.如权利要求1所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤1中，多目标轨迹识别；所述步骤1包括：2. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 1, characterized in that, in step 1, multi-target trajectory identification is performed; step 1 includes: 步骤11：目标识别；Step 11: Target identification; 步骤12：目标跟踪；Step 12: Target tracking; 步骤13：目标向量化表示。Step 13: Vectorize the target representation. 3.如权利要求2所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤11中，目标识别；3. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 2, characterized in that, in step 11, target identification; 在图像识别中，基于卷积神经网络算法表现最好；在本方法中，重点关注于人员的识别；本方法通过多层卷积神经网络对视频流进行逐帧分析，识别并定位视频帧中的人员，对于输入视频V,V＝{V1,V2,……,Vt}，t为不同时刻的帧；In image recognition, convolutional neural network-based algorithms perform best. In this method, the focus is on the identification of people. This method uses a multi-layer convolutional neural network to analyze the video stream frame by frame, identify and locate people in the video frames. For the input video V, V = { V1 , V2 , ..., Vt }, t is the frame at different times. 目标识别模型构建分为模型训练和模型测试两个阶段；在训练阶段，输入带有标注信息的视频数据Vtrain；标注内容为若干个矩形边框，表示为(x,y,h,w)，其中x为标注边框的左上角横坐标，y为标注边框的左上角纵坐标，h为标注边框的高度，w为标注边框的宽度；通过训练，模型收敛后，得到目标识别模型；在测试阶段，输入无标注信息的视频数据Vtest，测试模型效果；若满足准确性需求，模型训练完毕；否则，调整参数，重新训练，直到符合准确率需求为止。The target recognition model construction consists of two phases: model training and model testing. In the training phase, the input is video data V<sub> train </sub> with labeled information. The labels consist of several rectangular borders, represented as (x, y, h, w), where x is the x-coordinate of the top-left corner of the border, y is the y-coordinate of the top-left corner, h is the height of the border, and w is the width of the border. After training, the model converges, resulting in the target recognition model. In the testing phase, the input is video data V <sub>test </sub> without labeled information to test the model's performance. If the accuracy requirement is met, the model training is complete; otherwise, the parameters are adjusted, and the model is retrained until the accuracy requirement is met. 4.如权利要求3所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤12中，目标跟踪；4. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 3, characterized in that, in step 12, target tracking; 将视频逐帧的目标边框信息按照时间顺序排列，得到目标轨迹信息；由于视频中相邻帧图像变化不大，逐帧分析的运算成本过高，所以对视频帧进行采样；The target bounding box information of each frame of the video is arranged in chronological order to obtain the target trajectory information. Since the image changes little between adjacent frames in the video, the computational cost of frame-by-frame analysis is too high, so the video frames are sampled. 标准视频为每秒24帧，在本方法中采样比取值为 Standard video is 24 frames per second; in this method, the sampling ratio is taken as... 5.如权利要求4所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤13中，目标向量化表示；5. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 4, characterized in that, in step 13, the target is represented by a vectorization. 对于每个目标Oi，在时刻t0的状态表示为之后的时刻t的位移向量为通过状态变量表示目标Oi的位置，通过位移向量表示目标Oi的移动方向和速度，描述目标的静态与动态特征。For each objective O <sub>i</sub> , the state at time t <sub>0 </sub> is represented as: The displacement vector at the subsequent time t is Through state variables The position of target Oi is represented by the displacement vector. It represents the direction and speed of movement of target Oi , and describes the static and dynamic characteristics of the target. 6.如权利要求5所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤2中，时空权重向量构建；6. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 5, characterized in that, in step 2, a spatiotemporal weight vector is constructed; 所述步骤2包括：Step 2 includes: 步骤21：时间权重向量构建；Step 21: Constructing the time weight vector; 步骤22：空间权重向量构建；Step 22: Constructing the spatial weight vector; 步骤23：时空向量融合。Step 23: Spatiotemporal vector fusion. 7.如权利要求6所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤21中，时间权重向量构建；7. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 6, characterized in that, in step 21, a time weight vector is constructed; 考虑到人们的生活作息，人流的运动趋势往往具有周期性规律；早晨的人流规模一般高于夜间；不同时间段人流量的差异导致基准量纲的不同；所以，需要针对这种差异性对不同时间段的人流预测模型训练不同的权重参数；Considering people's daily routines, the movement of people often follows a periodic pattern; the scale of people in the morning is generally higher than at night; the difference in people flow at different times leads to different baseline dimensions; therefore, it is necessary to train different weight parameters for the people flow prediction model at different times to take into account this difference. 对其人流特征进行连续采样，获得其总位移向量根据采样频率确定t的数量与每次采样的时间间隔；设定采样频率为5分钟异常，则采样时刻为00:00,00:05，……,23:55，采样次数为288；By continuously sampling the characteristics of the pedestrian flow, the total displacement vector can be obtained. The number of samples t and the time interval between each sample are determined based on the sampling frequency; if the sampling frequency is set to 5 minutes, then the sampling times are 00:00, 00:05, ..., 23:55, and the number of samples is 288. 时间权重其中ε为小常量，防止除零；Time weight Where ε is a small constant to prevent division by zero; 步骤22：空间权重向量构建；Step 22: Constructing the spatial weight vector; 特定的地点也会对人流量产生影响；如果不消除这种影响，很有可能导致人员聚集事件误判；Specific locations can also affect the flow of people; if this effect is not eliminated, it could lead to misjudgments of gatherings of people. 对于特定观察点x，统计其出现人数n，则空间权重其中ε为小常量，防止除零；For a specific observation point x, if the number of people appearing at that point n is counted, then the spatial weight is... Where ε is a small constant to prevent division by zero; 步骤23：时空向量融合；Step 23: Spatiotemporal vector fusion; 对于不同时间段t的不同观察点x，其时空权重向量为For different observation points x at different time periods t, the spatiotemporal weight vector is: 其中，wi,j为时间向量与空间向量的乘积。Where w <sub>i,j </sub> is the product of the time vector and the space vector. 8.如权利要求7所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤3中，特殊人员权重向量构建；8. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 7, characterized in that, in step 3, a special personnel weight vector is constructed; 特殊人员会对人群轨迹造成较大影响，表现形式为人群易向公众人物位置靠近，特殊人员周围的人员密度增大且移动速度降低；特殊人员出行很容易造成群体性事件，对安防造成较大挑战；特殊人员出行是一个不得不考虑的场景，但同时也是一个小概率事件，仅在特殊场景下适用；Special individuals can significantly impact crowd movement, manifesting as a tendency for crowds to gravitate towards public figures, increased crowd density around them, and slower movement speed. Their travels can easily trigger mass incidents, posing a significant challenge to security. While such travel is a scenario that must be considered, it is also a low-probability event and should only be applied in specific situations. 所述步骤3包括：Step 3 includes: 步骤31：特殊人员库构建；Step 31: Constructing a special personnel database; 根据实际情况，在不同的观察点建立特殊人员人脸库，也可以建立全局人脸库；建立人脸库，保存人员唯一ID、人脸图片、人脸特征、权重向量；Depending on the actual situation, a special personnel face database can be established at different observation points, or a global face database can be established; the face database should store the unique ID of each person, face image, face features, and weight vector; 步骤32：特殊人员权重向量构建；Step 32: Constructing the weight vector for special personnel; 特殊人员对周围人群具有一定吸引效果，根据吸引效果强弱，构建吸引权重；通过对特殊人员划分级别，估计特殊人员的吸引力权重；由于特殊人员数量相对较少，特殊人员权重向量构建的工作量较低。Special individuals have a certain attraction effect on the surrounding population. Based on the strength of the attraction effect, attraction weights are constructed. By classifying special individuals into different levels, the attraction weights of special individuals are estimated. Since the number of special individuals is relatively small, the workload of constructing the special individual weight vector is low. 9.如权利要求8所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤4中，进行多目标轨迹预测分析；9. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 8, characterized in that, in step 4, multi-target trajectory prediction analysis is performed; 所述步骤4的整体流程为：The overall process of step 4 is as follows: 首先，通过基于卷积神经网络的多目标轨迹识别算法，获取目标Oi的状态变量以及位移向量将状态变量序列 和位移向量序列乘以时空权重向量，并输入序列预测模型，预测接下来一段时间内的状态变量序列和位移向量序列 First, the state variables of target Oi are obtained through a multi-target trajectory recognition algorithm based on a convolutional neural network. and displacement vector sequence of state variables and displacement vector sequence Multiply by the spatiotemporal weight vector and input into the sequence prediction model to predict the sequence of state variables over the next period of time. and displacement vector sequence 然后，采用人脸识别算法识别视频中的人脸，检查是否有特殊人员出现；根据检测结果，分为两种情况；Then, a facial recognition algorithm is used to identify faces in the video to check for any suspicious individuals; based on the detection results, two scenarios are identified. 情况一：存在特殊人员；根据特殊人员t时刻的位置以及特殊人员吸引力权重w；对原始输出结果进行修正，目标Oi在t时刻的预测位置为修正向量为预测结果即为状态变量序列 和位移向量序列 Scenario 1: Special personnel exist; based on the special personnel's position at time t. And the special personnel attraction weight w; the original output results are corrected, and the predicted position of target Oi at time t is... The correction vector is The prediction result is the sequence of state variables. and displacement vector sequence 情况二：不存在特殊人员；预测结果即为状态变量序列 和位移向量序列 Scenario 2: No special personnel are present; the prediction result is the sequence of state variables. and displacement vector sequence 10.如权利要求9所述的基于时空特征的群体性人员轨迹预测方法，其特征在于，所述步骤4中，进行模型迭代；10. The method for predicting the trajectory of a group of people based on spatiotemporal features as described in claim 9, characterized in that, in step 4, model iteration is performed; 由于实际人群是动态变化的，本方法中有三个模型组件需要根据实际数据不断迭代，以适应变化；Since the actual population is dynamic, three model components in this method need to be continuously iterated based on actual data to adapt to changes; a.轨迹预测模型a. Trajectory prediction model 每隔一段时间，进行数据采样，将新数据加入原始训练集重新训练，训练集规模不断扩大，提高模型的准确性；Every so often, data is sampled, and new data is added to the original training set for retraining. The size of the training set is continuously expanded, thereby improving the accuracy of the model. b.时空向量权重b. Spatiotemporal vector weights 每隔一段时间，进行数据采样；根据新数据计算时空向量，以确保模型能够适应变化的情况；Data is sampled periodically; spatiotemporal vectors are calculated based on the new data to ensure that the model can adapt to changing conditions. c.特殊人员库c. Special Personnel Database 根据实际情况，不断扩充特殊人员库，并调整级别；Based on actual circumstances, the pool of special personnel will be continuously expanded and their levels adjusted. 所述步骤4的部署实施如下：The deployment and implementation of step 4 are as follows: 本方法具有两种部署应用场景；This method has two deployment scenarios; a.公共场所部署a. Deployment in public places 面向街道等公共区域，人员流动性大，不需要建立特殊人员库；时空向量的数据主要来源于早晚高峰；For public areas such as streets, where there is high population mobility, there is no need to establish a special personnel database; the spatiotemporal vector data mainly comes from morning and evening rush hours. b.内部区域部署b. Internal area deployment 面向园区等内部区域，人员相对固定，可以建立相对完善的内部人员库，并提取关键人员加入特殊人员库；并且，可以根据特定场所进行时空向量建模。For internal areas such as industrial parks, where personnel are relatively fixed, a relatively complete internal personnel database can be established, and key personnel can be extracted and added to a special personnel database; furthermore, spatiotemporal vector modeling can be performed based on specific locations.
CN202311864476.5A 2023-12-28 2023-12-28 A group personnel trajectory prediction method based on spatiotemporal characteristics Pending CN117953009A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311864476.5A CN117953009A (en) 2023-12-28 2023-12-28 A group personnel trajectory prediction method based on spatiotemporal characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311864476.5A CN117953009A (en) 2023-12-28 2023-12-28 A group personnel trajectory prediction method based on spatiotemporal characteristics

Publications (1)

Publication Number Publication Date
CN117953009A true CN117953009A (en) 2024-04-30

Family

ID=90797217

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311864476.5A Pending CN117953009A (en) 2023-12-28 2023-12-28 A group personnel trajectory prediction method based on spatiotemporal characteristics

Country Status (1)

Country Link
CN (1) CN117953009A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118396390A (en) * 2024-06-26 2024-07-26 中国民用航空总局第二研究所 Method, device, equipment and medium for determining the evolution state of mass incidents at airports
CN119962786A (en) * 2025-01-14 2025-05-09 哈尔滨工业大学 Crowd trajectory prediction method and system for shopping center traffic space based on artificial jellyfish search algorithm
CN120126220A (en) * 2025-05-09 2025-06-10 北京酷鲨科技有限公司 Method, system and equipment for predicting mine personnel trajectory based on artificial intelligence

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118396390A (en) * 2024-06-26 2024-07-26 中国民用航空总局第二研究所 Method, device, equipment and medium for determining the evolution state of mass incidents at airports
CN119962786A (en) * 2025-01-14 2025-05-09 哈尔滨工业大学 Crowd trajectory prediction method and system for shopping center traffic space based on artificial jellyfish search algorithm
CN120126220A (en) * 2025-05-09 2025-06-10 北京酷鲨科技有限公司 Method, system and equipment for predicting mine personnel trajectory based on artificial intelligence
CN120126220B (en) * 2025-05-09 2025-09-05 北京酷鲨科技有限公司 Mine personnel trajectory prediction method, system and equipment based on artificial intelligence

Similar Documents

Publication Publication Date Title
Zhang et al. Combining motion and appearance cues for anomaly detection
CN117953009A (en) A group personnel trajectory prediction method based on spatiotemporal characteristics
Makhmutova et al. Object tracking method for videomonitoring in intelligent transport systems
CN110728252B (en) Face detection method applied to regional personnel motion trail monitoring
CN115359568B (en) Simulation method for pedestrian intelligent body movement and emergency evacuation and computer equipment
CN111402298A (en) Grain Depot Video Data Compression Method Based on Target Detection and Trajectory Analysis
CN119359359B (en) An Improved Customer Flow Statistics Method for 4S Stores Based on Pedestrian Re-identification
Cao et al. Learning spatial-temporal representation for smoke vehicle detection
Alashban et al. Single convolutional neural network with three layers model for crowd density estimation
Ji et al. A hybrid model of convolutional neural networks and deep regression forests for crowd counting
CN120259946B (en) A rail transit video intelligent analysis method, medium and system
CN119540876B (en) Deep learning-based personnel abnormal behavior analysis method
Nyajowi et al. CNN real-time detection of vandalism using a hybrid-LSTM deep learning neural networks
Mantini et al. Camera Tampering Detection using Generative Reference Model and Deep Learned Features.
Mostafa et al. Abnormal event detection in crowded scenarios
Yang et al. Crowd gathering and commotion detection based on the stillness and motion model
CN119539250A (en) BIM-based trajectory analysis method, system, terminal and storage medium
CN118334699A (en) Pedestrian behavior track prediction method based on multi-feature fusion strategy
Prezioso et al. Integrating object detection and advanced analytics for smart city crowd management
CN115527270A (en) A Specific Behavior Recognition Method in Dense Crowd Environment
Jain et al. Comparative analysis YOLOv5 and YOLOv8 for vehicle classification
Giriprasad et al. Anomalies detection from video surveillance using support vector trained deep neural network classifier
Ge et al. [Retracted] Theory and Method of Data Collection for Mixed Traffic Flow Based on Image Processing Technology
CN114120240B (en) Individual behavior identification method and system in complex environment
Kwak et al. Detection of Abnormal Human Behavior Using Unsupervised Learning in Video Surveillance Systems

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination