A method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning

Abstract

A method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning relates to the field of simulation analysis of crowd evacuation in buildings, and solves the problem of how to conveniently and efficiently determine the true value of the steady-state speed of a certain type of people from real evacuation video data. The present invention provides the following scheme: The method includes: determining the start and end frame numbers of the effective time period in the real evacuation video; extracting the personnel movement trajectory data in the evacuation video based on the deep learning video extraction method; calculating the effective personnel movement speed data based on the obtained start and end frame numbers of the effective time period and the intercepted personnel movement trajectory; summarizing all effective personnel movement speed data, and eliminating outliers through box plots, retaining effective personnel evacuation steady-state speed data; performing statistical analysis on the remaining effective personnel evacuation steady-state speed data to determine the expected value and corresponding distribution of the steady-state speed data of various groups of people in real evacuation movements.

Description

Technical Field

[0001] The invention relates to the technical field of crowd evacuation simulation analysis in buildings. Background Art

[0002] The speed of movement of personnel during the evacuation process (hereinafter referred to as personnel evacuation speed) is one of the basic parameters required for various evacuation simulation methods when simulating the evacuation process. At the same time, the personnel evacuation speed is also an important factor affecting the results of evacuation simulation. The accuracy of the personnel evacuation speed in the simulation method directly determines the accuracy of the evacuation simulation results. The sources of personnel movement speed data during emergency evacuation in the prior art are mainly concentrated in two aspects: on the one hand, through observation or questionnaire survey methods, personnel movement data under normal conditions are extracted; on the other hand, through experimental methods, personnel movement data in the experiment is obtained.

[0003] Because people are in panic in an emergency, their movement speed is often higher than that in a normal state. Therefore, the data in the first aspect cannot reflect the actual movement characteristics of evacuated people in an emergency. In addition, people are escaping from unexpected dangerous events during the actual evacuation process, and the data obtained by the experimental method is often affected by individual subjective thinking, and it is also not in line with the evacuation characteristics in an emergency. In addition, the evacuation speed of people is not stable during the evacuation process. It will change with the changes in the surrounding environment during the movement of people, and this change is highly random and difficult to describe quantitatively. The steady-state value of the evacuation speed of people (hereinafter referred to as steady-state speed) indicates that the movement speed value of people can maintain a small range of fluctuations during the evacuation movement without external interference. The steady-state speed corresponds to a relatively stable movement state of people, that is, the movement speed of people has not changed significantly. It is not affected by the evacuation movement scene, but only related to the characteristics of the evacuated people themselves. Therefore, in the evacuation simulation method, the steady-state speed can be used to quantitatively describe the speed characteristics of the evacuation movement of people. In summary, the steady-state speed during the actual emergency evacuation process (hereinafter referred to as the true value of the steady-state speed) can be used as the basic parameter of the personnel motion characteristics of various evacuation simulation methods, thereby further improving the accuracy of the evacuation simulation results in reflecting the actual situation.

[0004] At present, the main data source channel for the statistical sample data of the true value of steady-state speed is to extract personnel movement data from real evacuation videos. In the prior art, the paper "Towards creating a combined database for earthquake pedestrians' evacuation models. Safety Science. 82 (2016) 77-94." published in 2016 specifically describes the video data extraction method based on Tracker, which is a method that can accurately extract personnel movement data from real evacuation videos. However, the video data extraction method based on Tracker requires manual assistance to determine the tracking point during its operation. When the number of people in the video is large, the manual operation workload increases sharply. In addition, its coordinate calibration process is also relatively complicated. In order to ensure the accuracy of the extracted data, it is necessary to calibrate the coordinates of people at different depths in the video scene. Therefore, it is necessary to establish a real evacuation video data extraction method that can accurately extract data and is more convenient to operate, so as to greatly reduce the workload of extracting the true value of steady-state speed from real evacuation video data.

[0005] Therefore, in the evacuation simulation method, the steady-state speed can be used to quantitatively describe the speed characteristics of the personnel evacuation movement. In summary, obtaining the true value of the steady-state speed as the basic parameter of various evacuation simulation methods can further improve the accuracy of the evacuation simulation results in reflecting the actual situation. Summary of the invention

[0006] In order to solve the problem of how to conveniently and efficiently determine the true value of the steady-state speed of a certain type of personnel from a real evacuation video in the prior art, the present invention is implemented through the following technical solutions to solve the above technical problems:

[0007] The present invention proposes a method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning, and the method comprises:

[0008] Step 1: determine the start and end frame numbers of the effective time period in the real evacuation video;

[0009] Step 2: A video data extraction method based on deep learning is used to extract the movement trajectory data of people in the real evacuation video during the effective period in step 1;

[0010] Step 3: Calculate effective personnel movement speed data, i.e., steady-state speed data, according to the start and end frame numbers of the effective time period obtained in step 1 and the personnel movement trajectory intercepted in step 2;

[0011] Step 4: Summarize all valid personnel movement speed data, remove outliers through box plots, and retain valid personnel evacuation steady-state speed data;

[0012] Step 5: Perform statistical analysis on the remaining steady-state speed data of effective personnel evacuation to determine the expected value and corresponding distribution of the steady-state speed data of each type of personnel in the actual evacuation movement, thus completing the method for determining the steady-state speed.

[0013] Further, a preferred embodiment is provided, wherein the deep learning video data extraction method in step 2 includes an evacuation personnel multi-target tracking module and an uncalibrated video scene coordinate conversion module;

[0014] The evacuee multi-target tracking module is used to identify the persons in the real evacuation video, and continuously track the identified persons in the video, and output the pixel coordinate sequence of the motion trajectory of the evacuee in the effective period;

[0015] The non-calibrated video scene coordinate conversion module is used to obtain the pixel coordinate sequence of the motion trajectory of the evacuees within the effective period from the non-calibrated video and convert it into a real-world coordinate sequence to obtain the real motion trajectory data of the evacuees.

[0016] Furthermore, a preferred implementation is provided, in which the method for the non-calibrated video scene coordinate conversion module to obtain the real motion trajectory data of the evacuated personnel is:

[0017] S1. Select 4 reference points in the scene of the real evacuation video and obtain pixel coordinates;

[0018] S2. Obtain the real-world coordinates of four reference points to form coordinate pairs of the four reference points;

[0019] S3, based on the coordinate pairs of the four reference points, solving mapping coefficients c1 to c8; the mapping coefficients are determined by a coordinate transformation model of the evacuation video scene;

[0020] S4, input the real-time pixel coordinate sequence of the personnel movement trajectory extracted from the coordinate transformation model of the evacuation video scene into the coordinate transformation model determined in S3, and output the corresponding real-world coordinate sequence.

[0021] Furthermore, a preferred implementation is provided, in which a real-time pixel coordinate sequence of a person's motion trajectory is extracted by a multi-target tracking algorithm combining YOLOv4 and Deep SORT in S4.

[0022] Further, a preferred implementation is provided, the steady-state speed data in step 3 is obtained by formula (3), that is,

[0023]

[0024] Where i represents the row number of the displacement data of the personnel evacuation movement trajectory within the effective period, corresponding to a certain moment; v i The evacuation speed of personnel at the evacuation time corresponding to the i-th row; x i represents the x-coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system, and y i N represents the y coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system; i It represents the frame number of the frame corresponding to the i-th row of displacement data in the data source evacuation video; Δt represents the time interval between adjacent frames in the data source evacuation video, that is, Δt=1 / FPS, where FPS represents the number of frames per second of the video.

[0025] Furthermore, a preferred implementation is provided, wherein the statistical results in step five include expected value, standard deviation, maximum value, maximum value and median value, and the distribution type conforming to the steady-state speed data of personnel is determined by frequency distribution histogram, normal pp graph, skewness, kurtosis test and Kolmogorov-Smirnov test.

[0026] Solution 2: A computer device includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the method for determining the steady-state speed described in any one of Solution 1.

[0027] Solution three: a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the method for determining the steady-state speed described in any one of Solution one.

[0028] The present invention is beneficial in that:

[0029] The present invention establishes a method for determining the steady-state speed of various groups of people in real evacuation movements, which can accurately and conveniently determine the steady-state evacuation speed of personnel of a specific group of people from a real evacuation video.

[0030] The present invention adopts a multi-target tracking algorithm combining YOLOv4 and Deep SORT, which can automatically identify and track evacuees in real evacuation videos and output the pixel coordinate sequence corresponding to their motion displacement trajectory. This step is fully automated, and no manual assistance is required during the process. Compared with the prior art, it can greatly reduce the workload of extracting the true value of steady-state speed from real evacuation video data. In addition, YOLOv4 uses a more powerful deep learning network, which can process larger input images and improve accuracy, so that people in the video can be identified more accurately and efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the correspondence between two types of coordinates of reference points in implementation mode one.

[0032] Figure 2 The evacuation scene images in the video that summarizes the data provided in the embodiment are shown.

[0033] Figure 3 This is a box plot of sample data of steady-state speed of personnel evacuation in implementation mode one.

[0034] Figure 4 It is the frequency distribution histogram of the effective sample data of the steady-state speed of personnel evacuation in implementation method six.

[0035] Figure 5 This is the normal PP diagram of the effective sample data of the steady-state speed of personnel evacuation in implementation mode six. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the implementation methods of the present application clearer, the technical solutions in the implementation methods of the present application will be clearly and completely described below in conjunction with the drawings in the implementation methods of the present application. Obviously, the described implementation methods are only part of the implementation methods of the present application, not all of the implementation methods.

[0037] Embodiment 1: This embodiment provides a method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning, and the method includes:

[0038] Step 1: determine the start and end frame numbers of the effective time period in the real evacuation video;

[0039] Step 2: A video data extraction method based on deep learning is used to extract the movement trajectory data of people in the real evacuation video during the effective period in step 1;

[0040] Step 3: Calculate effective personnel movement speed data, i.e., steady-state speed data, according to the start and end frame numbers of the effective time period obtained in step 1 and the personnel movement trajectory intercepted in step 2;

[0041] Step 4: Summarize all valid personnel movement speed data, remove outliers through box plots, and retain valid personnel evacuation steady-state speed data;

[0042] Step 5: Perform statistical analysis on the remaining steady-state speed data of effective personnel evacuation to determine the expected value and corresponding distribution of the steady-state speed data of each type of personnel in the actual evacuation movement, thus completing the method for determining the steady-state speed.

[0043] Implementation method 2: This implementation method is a further limitation of the method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning described in implementation method 1. The deep learning video data extraction method in step 2 includes an evacuee multi-target tracking module and an uncalibrated video scene coordinate conversion module;

[0044] The evacuee multi-target tracking module is used to identify the persons in the real evacuation video, and continuously track the identified persons in the video, and output the pixel coordinate sequence of the motion trajectory of the evacuee in the effective period;

[0045] The non-calibrated video scene coordinate conversion module is used to obtain the pixel coordinate sequence of the motion trajectory of the evacuees within the effective period from the non-calibrated video and convert it into a real-world coordinate sequence to obtain the real motion trajectory data of the evacuees.

[0046] Implementation method three: This implementation method is a further limitation of the method for determining the steady-state speed of various groups of people in real evacuation motion based on deep learning described in implementation method two. The method for the non-calibrated video scene coordinate conversion module to obtain the real motion trajectory data of the evacuated personnel is:

[0047] S1. Select 4 reference points in the scene of the real evacuation video and obtain pixel coordinates;

[0048] S2. Obtain the real-world coordinates of four reference points to form coordinate pairs of the four reference points;

[0049] S3, based on the coordinate pairs of the four reference points, solving mapping coefficients c1 to c8; the mapping coefficients are determined by a coordinate transformation model of the evacuation video scene;

[0050] S4, input the real-time pixel coordinate sequence of the personnel movement trajectory extracted from the coordinate transformation model of the evacuation video scene into the coordinate transformation model determined in S3, and output the corresponding real-world coordinate sequence.

[0051] Implementation method 4. This implementation method is a further limitation of the method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning described in implementation method 3. In S4, the real-time pixel coordinate sequence of the movement trajectory of the personnel is extracted by the multi-target tracking algorithm combining YOLOv4 and Deep SORT.

[0052] In this implementation, in the multi-target tracking algorithm combining YOLOv4 and Deep SORT, YOLOv4 is used to detect and identify the target type personnel specified by the extracted data. It is an efficient target detection algorithm based on a deep convolutional neural network and can perform target detection in scenarios with high real-time requirements. DeepSORT is used to track the target personnel in the video screen detected and identified by YOLOv4 in a time series. DeepSORT is a target tracking algorithm based on the Kalman filter and the Hungarian algorithm. It can track the targets in the video frames in a time series and can adaptively adjust the parameters of the Kalman filter to improve the tracking accuracy and robustness.

[0053] In the actual video analysis process, the multi-target tracking algorithm that combines YOLOv4 and Deep SORT first detects and identifies the target type personnel specified by the extracted data in the video image frame through YOLOv4, and then inputs the dynamic target results detected and identified by YOLOv4 into DeepSORT. DeepSORT then assigns the same number to the same dynamic target and displays it in the upper left corner of the target box, connects the same dynamic targets that appear in multiple frames, and finally forms the motion trajectory data of the dynamic target.

[0054] Implementation method 5: This implementation method further limits the method for determining the steady-state speed of various groups of people in real evacuation movement based on deep learning described in implementation method 1. The steady-state speed data in step 3 is obtained by formula (3), that is,

[0055]

[0056] Where i represents the row number of the displacement data of the personnel evacuation movement trajectory within the effective period, corresponding to a certain moment; v i The evacuation speed of personnel at the evacuation time corresponding to the i-th row; x i represents the x-coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system, and y i N represents the y coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system; i It represents the frame number of the frame corresponding to the i-th row of displacement data in the data source evacuation video; Δt represents the time interval between adjacent frames in the data source evacuation video, that is, Δt=1 / FPS, where FPS represents the number of frames per second of the video.

[0057] Implementation method six. This implementation method is a further limitation of the method for determining the steady-state speed of various types of people in real evacuation movements based on deep learning described in implementation method one. The statistical results in step five include expected value, standard deviation, maximum value, maximum value and median value, and the distribution type of the steady-state speed data of the personnel is determined by frequency distribution histogram, pp diagram, skewness, kurtosis test and Kolmogorov-Smirnov test.

[0058] Embodiment 7. This embodiment proposes a computer device, including a memory and a processor, characterized in that a computer program is stored in the memory, and when the processor runs the computer program stored in the memory, the processor executes the method for determining the steady-state speed described in any one of embodiments 1 to 6.

[0059] Embodiment 8: This embodiment proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method for determining the steady-state speed described in any one of embodiments 1 to 6 are implemented.

[0060] Embodiment 9. This embodiment provides the following embodiments, which are used to explain the above embodiments 1 to 7. Embodiment 1 specifically includes:

[0061] Step 1: determine the start and end frame numbers of the effective time period in the real evacuation video;

[0062] According to the definition of steady-state speed, we believe that the video time period during which personnel are not affected by the surrounding environment (i.e., the influence of people around them is small and there are no obstacles) and can evacuate freely during emergency evacuation is the effective time period for extracting the steady-state speed of personnel evacuation. Therefore, we need to first determine the start and end video frame numbers corresponding to the effective time period from the video.

[0063] Step 2: Use a deep learning-based video data extraction method to extract the movement trajectory data of people in the real evacuation video during this period;

[0064] In order to accurately and efficiently extract personnel motion data from real evacuation videos, this embodiment establishes a video analysis method that can fully automatically extract motion data from uncalibrated real evacuation videos. The method consists of two functional modules, including an evacuee multi-target tracking module and an uncalibrated video scene coordinate conversion module.

[0065] For the multi-target tracking module for evacuees, after comparing the recognition effect of the more mature application-level multi-target tracking algorithm in the prior art, this implementation method selects a multi-target tracking algorithm that combines YOLOv4 and Deep SORT to automatically identify and track the specified types of evacuees in the video to obtain the pixel coordinate sequence corresponding to their motion displacement trajectory. The types of people here mainly refer to men, women, the elderly, children, adults, etc., which can be clearly distinguished only by the characteristics of the picture. YOLOv4 uses a more powerful deep learning network that can process larger input images and improve accuracy, so that people in the video can be identified more accurately and efficiently.

[0066] For the non-calibrated video scene coordinate conversion module, since the real earthquake evacuation video in the video set cannot perform camera calibration on the recording scene, it is impossible to use the traditional coordinate conversion method based on camera calibration to perform coordinate conversion between pixel coordinates and real-world coordinates. After analyzing the spatial movement characteristics of a large number of evacuation movements, this embodiment will obtain the motion trajectory of the evacuees on the evacuation plane, which is equivalent to obtaining the motion trajectory of the bottom center point of the bounding box that identifies the target person. On the basis of this equivalent simplification, this embodiment uses the homography change theory between corresponding planes to perform coordinate conversion. The coordinate conversion formula of the homography change theory is shown in formula (1), where it represents the real (x gi ,y gi ) world coordinates, (x i ,y i ) represents the pixel coordinates of the video image, and c1 to c8 are the mapping coefficients of the plane transformation. The formula for calculating the real world coordinates of the point can be derived from formula (1), as shown in formula (2).

[0067]

[0068]

[0069] It has been verified that the uncalibrated video scene coordinate conversion module in this embodiment can relatively accurately convert the pixel coordinate sequence of the evacuee motion trajectory obtained from the uncalibrated video into a real-world coordinate sequence, thereby obtaining the real motion trajectory of the evacuee.

[0070] The processing flow of the uncalibrated video scene coordinate conversion module can be divided into four steps, as follows:

[0071] S1. Select 4 reference points and obtain pixel coordinates.

[0072] See also Figure 1To illustrate this implementation, four non-collinear and easily identifiable reference points {A, B, C, D} are selected in the video scene according to the plane information of the evacuation scene in the video image. The four reference points are all on the ground plane in the escape scene. The pixel coordinates (x i ,y i ).

[0073] S2. Obtain the real-world coordinates of the four reference points.

[0074] First, use SketchUp to model the 3D coordinates of the video scene image, and determine the scale of the 3D coordinate system by matching the actual size of the real objects in the background. Then make guide lines for the reference points relative to the origin of the coordinate system, and then obtain the real-world coordinates of the four reference points. Finally, the coordinate pairs of four reference points are formed, and the following corresponding relationships exist in the coordinate pairs: Where i∈{A,B,C,D}.

[0075] S3. Solve the mapping coefficients c1 to c8.

[0076] Directly combine the four pairs of coordinates Substituting into equation (2) can determine the mapping coefficients c1 to c8. At this point, the unique coordinate transformation model for the current evacuation video background is determined.

[0077] S4. Batch processing of coordinate transformation.

[0078] By inputting a large number of real-time pixel coordinate sequences of personnel movement trajectories extracted by the multi-target tracking program from the evacuation video into a program with a determined coordinate transformation model, batch coordinate transformation can be automatically realized and the corresponding real-world coordinate sequence can be output.

[0079] After completing the above four process steps, all operations in step 2 are completed, and the real displacement coordinate data of the evacuation motion trajectory of all specific types of personnel in an evacuation video can be obtained. The storage structure of the real displacement coordinate data of the evacuation motion trajectory of a single person is shown in Table 1.

[0080]

[0081] Step 3: intercept the movement trajectory data of the person according to the start and end frame numbers of the effective time period, and further calculate the corresponding movement speed.

[0082] For the real displacement coordinate data of the evacuation motion trajectory of all specific types of personnel in the evacuation video obtained in 1.3., the start and end frame numbers of the effective time period are intercepted to obtain the displacement data of the personnel evacuation motion trajectory in the effective time period. Since the time interval for calculating the movement speed of personnel is short (0.1s or 0.2s), we can approximate the evacuation movement of personnel in this time interval as uniformly accelerated motion. Based on the principle that the average speed of uniformly accelerated motion is equal to the instantaneous speed at the intermediate moment, this embodiment uses formula (3) to calculate the motion speed value corresponding to a certain moment (frame number), that is,

[0083]

[0084] In formula (3), i represents the row number of the displacement data of the personnel evacuation movement trajectory within the effective period, corresponding to a certain moment. i The evacuation speed of personnel at the evacuation time corresponding to the i-th row. i represents the x-coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system, and y i Represents the y coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system. N i Indicates the frame number of the frame corresponding to the i-th row of displacement data in the data source evacuation video. Δt indicates the time interval between adjacent frames in the data source evacuation video, that is, Δt = 1 / FPS. FPS indicates the number of frames per second of the video.

[0085] According to formula (3), the corresponding evacuation movement speed data of the personnel in the effective period is calculated from the displacement data of the personnel evacuation movement trajectory in the effective period, thus completing the data processing of step 3.

[0086] Step 4: Eliminate outliers in the movement speed data within the effective period through a box plot;

[0087] Furthermore, the movement speed data of all people of the same type in the valid period are summarized together. Considering that the time interval of 0.1s data is short, it is easy to amplify the error when calculating the speed value. Therefore, for the speed data, the box plot is used to remove the outlier samples, and the retained valid sample data is used for subsequent statistical analysis.

[0088] Step 5: Perform statistical analysis on the valid sample data to determine the expected value and corresponding distribution of the steady-state speed of this type of personnel in the actual evacuation movement.

[0089] The physical meaning of steady-state speed is the speed of movement with small fluctuations that personnel can maintain during the evacuation process without external interference. The corresponding personnel movement state is relatively stable, that is, the personnel movement speed has not changed significantly. Therefore, the movement speed of personnel in the effective period is the steady-state speed. By statistically analyzing the effective sample data of the movement speed of all personnel of the same type in the effective period, the basic statistical results of the steady-state speed of this type of personnel are obtained, including the expected value (Arithmetic mean), standard deviation (St.deviation), minimum (Minimum), maximum (Maximum) and median (Maximum).

[0090] Furthermore, we can use frequency distribution histogram, normal pp plot, skewness and kurtosis test, and Kolmogorov-Smirnov (KS) test to examine the probability distribution type of the steady-state speed of this type of personnel.

[0091] Frequency distribution histogram is a common data visualization method that can be used to show the distribution of data. If the data conforms to the normal distribution, then its histogram should present a bell-shaped curve. The PP graph is a graph drawn based on the relationship between the cumulative proportion of the variable and the cumulative proportion of the specified distribution. The PP graph can be used to test whether the data conforms to the specified distribution. When the data conforms to the specified distribution, the points in the PP graph are approximately a straight line.

[0092] Skewness is a measure of the direction and degree of skewness of the statistical data distribution. It is a numerical feature that reflects the degree of asymmetry of the statistical data distribution. It is used to measure the symmetry of the data and is denoted by S. Kurtosis is a characteristic number that characterizes the height of the peak of the data probability density distribution curve at the mean value and is denoted by K. Intuitively, kurtosis reflects the degree of sharpness or flatness at the top of the frequency distribution curve. The skewness and kurtosis test formulas are shown in equations (4) and (5).

[0093]

[0094]

[0095] Among them, S S is the standard error of the S coefficient, S K is the standard error of the K coefficient. When α=0.05 and the absolute value of the Z value is greater than 1.96, it can be considered that the S coefficient or the K coefficient is significantly not equal to 0, that is, the sample data is non-normal.

[0096] Example 2: A portion of videos are selected from the real earthquake evacuation video collection collected in the prior art as the data source for determining the steady-state speed of personnel evacuation. The evacuation scenes in the video providing the data are as follows: Figure 2 shown. Figure 2 The number in the upper right corner of each scene picture is the number of the corresponding video of the scene in the real earthquake evacuation video set. The specific scene information in each scene can be obtained by referring to the prior art. The specific scene information in each scene is obtained through the paper "Extraction and modelling application of evacuation movement characteristic parameters in real earthquake evacuation video based on deep learning" (S. Li, L. Tong, C. Zhai, Extraction and modelling application of evacuation movement characteristic parameters inreal earthquake evacuation video based on deep learning, Int. J. Disaster Risk Reduct. 80 (2022) 10321) published in 2022. In the prior art, a part of videos are selected from the real earthquake evacuation video set as the data source for determining the steady-state speed of personnel evacuation.

[0097] The specific processing steps of this embodiment are as follows:

[0098] Step 1: determine the start and end frame numbers of the effective time period in the real evacuation video;

[0099] right Figure 2 For each video in the video, determine the start and end frame numbers of the effective time period one by one. Ensure that the evacuation movement state of personnel in each effective time period is relatively stable and is not disturbed by the evacuation scene environment, that is, the movement speed of personnel does not change significantly.

[0100] Step 2: Use the deep learning-based video data extraction method to extract the movement trajectory data of people in the real evacuation video of this period.

[0101] Multi-target identification and tracking of evacuees are performed on each video one by one, and the pixel coordinate sequence corresponding to the motion displacement trajectory of all personnel in the video is output.

[0102] In this embodiment, the steady-state speed of all types of people is counted, and the types of people are not subdivided. When it is necessary to determine the steady-state speed of a specific type of people (men, women, elderly, children, adults) later, the trajectory data of the corresponding people can be screened out according to the characteristics of the type of people.

[0103] Step 3: intercept the movement trajectory data of the person according to the start and end frame numbers of the effective time period, and further calculate the corresponding movement speed;

[0104] Determine the FPS of each video, extract the movement trajectory data of the personnel in each effective time period, and calculate the movement speed data, that is, the steady-state speed data, according to formula (3).

[0105] Step 4: Gather all valid personnel movement speed data together and remove outliers through box plots;

[0106] All valid personnel movement speed data are aggregated together and the box plot (see Figure 3 As shown in the figure, 17 outlier samples were removed and 633 valid sample data were finally retained.

[0107] Step 5: Perform statistical analysis on the remaining valid speed data to determine its expected value and corresponding distribution to be used as the basic parameters of the evacuation simulation method.

[0108] By statistically analyzing the remaining valid samples of steady-state speed data of personnel evacuation, the basic statistical results shown in Table 2 are obtained.

[0109]

[0110] Next, we examine the probability distribution type of the steady-state speed of personnel evacuation from the frequency distribution histogram, pp plot, skewness and kurtosis test, and Kolmogorov-Smirnov (KS) test.

[0111] The frequency distribution histogram of the steady-state speed data of personnel evacuation is as follows: Figure 4 As shown, its shape is closer to an inverted "bell shape" with symmetry on the left and right sides, so the data can be considered as normally distributed data.

[0112] The PP diagram of the steady-state speed data of personnel evacuation about the normal distribution is as follows: Figure 5 As shown. Figure 5 It can be seen that most of the points in the steady-state speed data of personnel evacuation can be distributed on a straight line, and the straight line trend is obvious. It can be considered that the data obeys the normal distribution.

[0113] The skewness and kurtosis statistical results of the steady-state speed data of personnel evacuation are shown in Table 3. Calculated by formula (4), the skewness coefficient ZS =1.659<1.96, calculated by formula (5), the kurtosis coefficient Z K =1.345<1.96. From the above results, we can see that the absolute values ​​of the skewness coefficient and the kurtosis coefficient are both less than 1.96, and it can be considered that this group of sample data conforms to the normal distribution.

[0114]

[0115] In the non-parametric test method, the Kolmogorov-Smirnov (KS) test is applicable to data with a data volume of >50. The number of valid samples in the steady-state speed data set of personnel evacuation obtained in this embodiment is 633, so this study uses the Kolmogorov-Smirnov (KS) test to test the normality of the steady-state speed data of personnel evacuation. Our original hypothesis is that "there is no significant difference between the population from which the sample comes and the normal distribution", and only P>0.05 can the original hypothesis be accepted, that is, the data conforms to the normal distribution. We use the software SPSS to perform a KS test on the steady-state speed data of personnel evacuation, and the test results are shown in Table 3. It can be seen from Table 4 that in the KS test, the lower limit of the true significance P value of the steady-state speed data of personnel evacuation is 0.200, which is much greater than 0.05. The original hypothesis is accepted, and it is believed that the distribution of the steady-state speed data of personnel evacuation is consistent with the normal distribution, that is, the steady-state speed data of personnel evacuation obeys the normal distribution.

[0116]

[0117] In summary, the four test methods all obtain the result that the steady-state speed data of personnel evacuation in this embodiment obeys the normal distribution. Therefore, in the micro-evacuation simulation method that can randomly assign the steady-state speed of personnel evacuation, the data statistical results according to Table 2 can be used to set the steady-state speed of personnel evacuation in the form of normal distribution.

[0118] It will be appreciated by those skilled in the art that the features described in the various embodiments and / or claims of the present disclosure may be combined or coupled in a variety of ways, even if such combinations or couplings are not explicitly described in the present disclosure. In particular, the various embodiments and of the present disclosure may be combined in a variety of ways without departing from the spirit and teachings of the present disclosure. All of these combinations fall within the scope of the present disclosure.

[0119] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning, characterized in that: The method comprises: Step 1: determine the start and end frame numbers of the effective time period in the real evacuation video; Step 2: A video data extraction method based on deep learning is used to extract the movement trajectory data of people in the real evacuation video during the effective period described in step 1; Step 3: Calculate effective personnel movement speed data, i.e., steady-state speed data, according to the start and end frame numbers of the effective time period obtained in step 1 and the personnel movement trajectory intercepted in step 2; Step 4: Summarize all valid personnel movement speed data, remove outliers through box plots, and retain valid personnel evacuation steady-state speed data; Step 5: Statistically analyze the remaining steady-state speed data of effective personnel evacuation to determine the expected value and corresponding distribution of the steady-state speed data of each type of personnel in the actual evacuation movement, that is, complete the method for determining the steady-state speed; The deep learning video data extraction method in step 2 includes an evacuee multi-target tracking module and an uncalibrated video scene coordinate conversion module; The evacuee multi-target tracking module is used to identify the persons in the real evacuation video, and continuously track the identified persons in the video, and output the pixel coordinate sequence of the motion trajectory of the evacuee in the effective period; The non-calibrated video scene coordinate conversion module is used to obtain the pixel coordinate sequence of the motion trajectory of the evacuees in the effective period from the non-calibrated video and convert it into a real-world coordinate sequence to obtain the real motion trajectory data of the evacuees; The method for the non-calibrated video scene coordinate conversion module to obtain the real motion trajectory data of the evacuated personnel is: S1. Select 4 reference points in the scene of the real evacuation video and obtain pixel coordinates; S2. Obtain the real-world coordinates of four reference points to form coordinate pairs of the four reference points; S3, based on the coordinate pairs of the four reference points, solving mapping coefficients c1 to c8; the mapping coefficients are determined by a coordinate transformation model of the evacuation video scene; S4, inputting the real-time pixel coordinate sequence of the personnel movement trajectory extracted from the coordinate transformation model of the evacuation video scene into the coordinate transformation model determined in S3, and outputting the corresponding real-world coordinate sequence; The real-time pixel coordinate sequence of the movement trajectory of the person extracted by the multi-target tracking algorithm combining YOLOv4 and Deep SORT in S4.

2. The method for determining the steady-state speed of various groups of people in real evacuation movements based on deep learning according to claim 1 is characterized in that: The steady-state speed data in step 3 is obtained by formula (3), namely (3) Where i represents the row number of the displacement data of the personnel evacuation movement trajectory within the effective period, corresponding to a certain moment; v i The evacuation speed of personnel at the evacuation time corresponding to the i-th row; x i represents the x-coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system, and y i N represents the y coordinate of the person at the evacuation time corresponding to the i-th row in the evacuation plane coordinate system; i It indicates the frame number of the frame corresponding to the i-th row of displacement data in the data source evacuation video; ∆t indicates the time interval between adjacent frames in the data source evacuation video, that is, ∆t=1 / FPS, where FPS indicates the number of frames per second of the video.

3. The method for determining steady-state speed of various groups of people in real evacuation movement based on deep learning according to claim 1 is characterized in that: The statistical results in step five include expected value, standard deviation, maximum value, median value, and the frequency distribution histogram, normal pp plot, skewness, kurtosis test and Kolmogorov-Smirnov test are used to determine the distribution type that the steady-state speed data of personnel conform to.

4. A computer device comprising a memory and a processor, characterized in that A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method for determining the steady-state speed according to any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for determining the steady-state speed according to any one of claims 1 to 3 are implemented.

Images