A method for recognizing panic pedestrian accelerated running behavior based on a dynamic centroid model

Abstract

The application relates to a panic pedestrian accelerated running behavior recognition method based on a dynamic centroid model, and the method comprises the following steps: acquiring panic pedestrian images frame by frame from a video; establishing a coordinate system according to the panic pedestrian images, and constructing a dynamic centroid model of the panic pedestrian; and judging whether the panic pedestrian performs accelerated running or not based on the dynamic centroid model, by using a geometric discrimination formula and a dynamics discrimination formula. Compared with the prior art, the discrimination method has the advantages of convenient calculation and strong robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crowd flow stability analysis, in particular to a panic pedestrian accelerated running behavior recognition method based on a dynamic centroid model. BACKGROUND

[0002] A sudden safety accident in a public place is easy to cause panic behavior of panic pedestrians, and the panic behavior of panic pedestrians is usually accompanied by accelerated running behavior. The recognition of this behavior is of great significance to public safety, and has gradually become a hot topic of research in the field of public safety in recent years. At present, most of the researches related to panic pedestrians list the panic behavior of panic pedestrians as abnormal behavior, and identify multiple abnormal behaviors such as panic behavior, while there are few researches on the accelerated running behavior of panic pedestrians. In the prior art, there is a method for extracting a region of interest (ROI) from a video frame by using an optical flow feature, and then using a convolutional neural network as a classifier to detect a person running, but this method is easily disturbed by factors such as light and background. There is also a method for extracting a foot image of a runner, and identifying a running posture by analyzing the spatial, temporal and frequency characteristics of the two feet when running, but this method does not involve the movement mode of the upper limbs in the running posture.

[0003] At present, there are several deficiencies in the research on the recognition of the accelerated running behavior of panic pedestrians:

[0004] 1. Lack of recognition of the accelerated running behavior of panic pedestrians alone;

[0005] 2. Lack of modeling and analysis of the posture of panic pedestrians when accelerating and running. SUMMARY

[0006] The purpose of the present application is to provide a panic pedestrian accelerated running behavior recognition method based on a dynamic centroid model, which uses a dynamic centroid to describe the characteristics of the accelerated running behavior, and judges the accelerated running behavior of panic pedestrians through a motion discriminant, having the advantages of easy calculation and strong robustness.

[0007] The purpose of the present application can be achieved by the following technical solutions:

[0008] A panic pedestrian accelerated running behavior recognition method based on a dynamic centroid model, the method comprising:

[0009] obtaining panic pedestrian images frame by frame from a video;

[0010] establishing a coordinate system according to the panic pedestrian images, and constructing a dynamic centroid model of the panic pedestrians;

[0011] based on the dynamic centroid model, using a geometric discriminant formula and a dynamics discriminant formula to judge whether the panic pedestrians are accelerating and running;

[0012] wherein the geometric discrimination formula is,

[0013]

[0014] β1 is the included angle between the segment connecting the left hip and the left knee and the segment connecting the left knee and the left ankle, and β2 is the included angle between the segment connecting the right hip and the right knee and the segment connecting the right knee and the right ankle;

[0015] The dynamic discrimination formula is,

[0016]

[0017] x(D.C) represents the dynamic center of mass horizontal coordinate value of the panic pedestrian, x(S.C) represents the horizontal coordinate value of the static center of gravity of the panic pedestrian, dir(D.C) represents the dynamic center of mass direction of the panic pedestrian, represents the acceleration direction of the panic pedestrian, represents the speed direction of the panic pedestrian.

[0018] Further, the construction process of the dynamic center of mass model of the panic pedestrian comprises:

[0019] extracting 2D skeleton key points of the panic pedestrian from the panic pedestrian image by using a pose estimation algorithm;

[0020] connecting the 2D skeleton key points of the panic pedestrian according to the limb connection relationship by bipartite graph matching;

[0021] calculating the dynamic center of mass coordinates of the panic pedestrian by using the coordinates of the 2D skeleton key points of the panic pedestrian.

[0022] Further, the 2D skeleton key points comprise: top of head, left ear, right ear, left eye, right eye, nose, left corner of mouth, right corner of mouth, neck, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle and right ankle.

[0023] Further, the calculation process of the dynamic center of mass coordinates of the panic pedestrian comprises:

[0024] all 2D skeleton key points of the same panic pedestrian face in the panic pedestrian image are processed into one head node by calculation, marked as 0;

[0025] respectively calculating the dynamic center of mass coordinates of different segments;

[0026] calculating the dynamic center of mass coordinates of the panic pedestrian by using the segment dynamic center of mass coordinates.

[0027] Further, the calculation processing formula of the head node is:

[0028]

[0029] wherein i is the sequence number of the 2D skeleton key point of the face, (x i , y i ) is the coordinate of the 2D skeleton key point.

[0030] Further, the calculation formula of the dynamic centroid of the different segments is:

[0031]

[0032] wherein k is the number of segments, p is the number of proximal key points of the segment, d is the number of distal key points of the segment, (x k , y k ) is the coordinate of the dynamic centroid of the segment, (x p , y p ) is the coordinate of the proximal key point of the segment, (x d , y d ) is the coordinate of the distal key point of the segment, l p is the proximal segment dynamic centroid coefficient, and l d is the distal segment dynamic centroid coefficient.

[0033] Further, the proximal key point of the segment is the 2D skeleton key point closer to the human torso among the two 2D skeleton key points connected by a segment, the distal key point of the segment is the 2D skeleton key point farther away from the human torso among the two 2D skeleton key points connected by a segment, and the proximal segment dynamic centroid coefficient and the distal segment dynamic centroid coefficient are calculated based on anatomical knowledge and statistical data to reflect the contribution degree of different body parts to the dynamic centroid of the panic pedestrian.

[0034] Further, the calculation formula of the dynamic centroid of the panic pedestrian is:

[0035]

[0036] wherein (x k , y k ) is the coordinate of the dynamic centroid of the segment, m k is the panic pedestrian body mass of the kth segment, and M is the total panic pedestrian body mass.

[0037] An electronic identification device, comprising:

[0038] one or more processors;

[0039] a memory;

[0040] One or more programs stored in the memory, the one or more programs comprising instructions for performing the method for recognizing panic pedestrian accelerated running behavior based on dynamic centroid model according to any one of claims 1-9.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] 1. The present application realizes the recognition of panic pedestrian accelerated running behavior by extracting panic pedestrian image from video image and establishing dynamic centroid model; when judging whether the panic pedestrian is running, the present application uses geometric discrimination formula and dynamics discrimination formula; the geometric discrimination formula judges the panic pedestrian posture according to the angle between the left thigh and the left shank and the angle between the right thigh and the right shank, and the dynamics discrimination formula further analyzes whether the panic pedestrian is accelerated running by comparing the dynamic centroid direction of the panic pedestrian with the speed direction, acceleration direction and gravity center direction of the pedestrian, and finally comprehensively judges the accelerated running behavior of the panic pedestrian; by using this method, the calculation is convenient and the robustness is strong.

[0043] 2. The present application connects the segmented 2D skeleton key points when establishing the dynamic centroid model, and models and analyzes the posture of the panic pedestrian when accelerated running. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 Method flowchart of the present application

[0045] Figure 2 Key point example diagram of human skeleton extracted from panic pedestrian image in the embodiment of the present application;

[0046] Figure 3 Comparison diagram of static centroid and dynamic centroid of panic pedestrian general walking posture, pocket inserting walking posture and running posture in the embodiment of the present application;

[0047] Figure 4 Dynamic centroid and static centroid coordinate diagram of panic pedestrian running posture in the embodiment of the present application;

[0048] Figure 5 Centroid trajectory diagram of accelerated running behavior in the embodiment of the present application. DETAILED DESCRIPTION

[0049] The present application will be described in detail below in combination with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and gives detailed implementation mode and specific operation process, but the protection scope of the present application is not limited to the following embodiments.

[0050] The embodiment discloses a panic pedestrian accelerated running behavior recognition method based on a dynamic centroid model, as shown in the following formula (1): Figure 1 The main steps include:

[0051] S1. Obtain the panic pedestrian image frame by frame from the video;

[0052] S2. Establish the coordinate system according to the panic pedestrian image, and construct the dynamic centroid model of the panic pedestrian;

[0053] S3. Based on the dynamic centroid model, use the geometric discrimination formula and the dynamics discrimination formula to determine whether the panic pedestrian is accelerated running.

[0054] The steps of the recognition method in the embodiment are described in detail below.

[0055] In the embodiment, the panic pedestrian accelerated running video is shot and observed in an open space as a verification scene video, and the video acquisition frequency f collection = 29.97 fps (Frames Per Second), for the convenience of calculation, f collection = 30 fps is used in the following, and the video single frame resolution is 1920*1080 pixels, the video frame image is intercepted by using MATLAB, and the interception frequency is FrameRate = 30.

[0056] In the embodiment, the upper left corner position of the video frame image is defined as the coordinate origin, the horizontal direction is the x axis, the horizontal coordinate value represents 1920 pixels, and the vertical direction is the y axis, and the vertical coordinate value represents 1080 pixels.

[0057] Further, the 2D skeleton key point coordinates of the panic pedestrian are extracted from the video image by using the pose estimation algorithm Openpose, Openpose first estimates the feature map of each joint node of the human body through multiple convolution layers, pooling layers and full connection layers, then the limbs of the same person are connected according to the limb connection relationship by using the bipartite graph matching to determine the connection relationship between the key points. Finally, as shown in the following formula (2), the number of extracted 2D skeleton key points is 21, including: head, left ear, right ear, left eye, right eye, nose, left mouth corner, right mouth corner, neck, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle. Figure 2

[0058] In the embodiment, the construction process of the dynamic centroid model of the panic pedestrian includes:

[0059] The 2D skeleton key points of the panic pedestrian are extracted from the panic pedestrian image by using the pose estimation algorithm;

[0060] The 2D skeleton key points of the panic pedestrian are connected according to the limb connection relationship by using the bipartite graph matching.​

[0061] The coordinates of the panicking pedestrian's 2D skeleton key points are used to calculate the dynamic center of mass coordinates of the panicking pedestrian.

[0062] Furthermore, the coordinates of the dynamic center of mass of the panicking pedestrian are calculated as follows:

[0063] The eight 2D skeleton key points (serial numbers 14-21) of the face of the same panicking pedestrian in the panicking pedestrian image are calculated and processed into one head node, which is marked as 0.

[0064] The calculation formula of the head node is:

[0065]

[0066] Among them, i is the key point number of the facial 2D skeleton, (x i ,y i ) are the 2D bone key point coordinates.

[0067] Calculate the dynamic center of mass coordinates of different segments respectively, the formula is as follows:

[0068]

[0069] Among them, k is the number of segments, p is the number of key points at the proximal end of the segment, d is the number of key points at the distal end of the segment, (x k ,y k ) is the segment dynamic centroid coordinate, (x p ,y p ) is the coordinate of the key point at the proximal end of the segment, (x d ,y d ) is the coordinate of the key point at the distal end of the segment, l p is the dynamic center of mass coefficient of the initial segment, l d is the dynamic center of mass coefficient of the terminal segment.

[0070] The proximal key point of the segment is the 2D bone key point closer to the human torso among the two 2D bone key points connected by a segment, and the distal key point of the segment is the 2D bone key point farther away from the human torso among the two 2D bone key points connected by a segment. The dynamic center of mass coefficient of the starting segment and the dynamic center of mass coefficient of the terminal segment are calculated based on anatomical knowledge and statistical data, and are used to reflect the contribution of different body parts to the dynamic center of mass of panicked pedestrians.

[0071] The dynamic centroid coordinates of the panicking pedestrian are calculated using the segment dynamic centroid coordinates. The formula is as follows:

[0072]

[0073] Among them, (x k ,y k) is the segment dynamic center of mass coordinate, m k is the mass of the panicking pedestrian in the kth segment, and M is the total mass of the panicking pedestrian.

[0074] Furthermore, the accelerated running behavior is identified. The static center of mass of the panic pedestrian is the center of mass of the panic pedestrian in a static state. The static center of mass and dynamic center of mass of the running posture are compared with the normal walking posture and the walking posture with the hands in pockets. Figure 3 As shown in the figure, P D.C is the dynamic center of mass, P S.C It is the static center of mass. It can be found that the positions of the static center of mass and the dynamic center of mass of the general walking posture and the walking posture with hands in pockets basically coincide, while the position of the dynamic center of mass in the running posture is biased towards the direction of running.

[0075] Therefore, combining the geometric characteristics of running posture and walking posture, we obtain the geometric and dynamic discriminant formulas for running posture, general walking posture, and walking posture with hands in pockets. The obtained 2D skeleton key points, segment angles, and dynamic center of mass data are substituted into the following geometric and dynamic discriminant formulas:

[0076]

[0077] Among them, such as Figure 4 As shown, β1 is the angle between the segment connecting the left hip and left knee and the segment connecting the left knee and left ankle (segments 8-9 and 9-10), β2 is the angle between the segment connecting the right hip and right knee and the segment connecting the right knee and right ankle (segments 11-12 and 12-13), x(DC) represents the horizontal coordinate value of the dynamic center of mass of the panicking pedestrian, x(SC) represents the horizontal coordinate value of the static center of mass of the panicking pedestrian, dir(DC) represents the direction of the dynamic center of mass of the panicking pedestrian, Indicates the acceleration direction of panicked pedestrians, Indicates the speed and direction of panicking pedestrians.

[0078] If the 2D skeleton key points, segment angles and dynamic center of mass data satisfy the geometric discriminant formula and the dynamic discriminant formula, it means that the panicked pedestrian is running at an accelerated speed.

[0079] The center of mass trajectory of a panicked pedestrian when he accelerates is as follows Figure 5 As shown in Figure 1, the upper curve represents the dynamic center of mass of a panicked pedestrian running at high speed, and the lower curve represents the static center of mass of a panicked pedestrian running at high speed. The acceleration process lasts for two seconds, and a total of 62 frames of images are extracted.

[0080] According to the comparison between the dynamic center of mass and the static center of mass of the panic runner in accelerated running, firstly, it can be seen that the dynamic center of mass of the panic runner has a flat platform at the peak, while the static center of mass of the panic runner has no platform at the peak and is relatively sharp. Secondly, in the time meaning of the dynamic center of mass and the static center of mass of the panic runner, the dynamic center of mass of the panic runner is ahead of the static center of mass. Taking the twentieth frame as an example, the dynamic center of mass is one frame time difference ahead of the static center of mass, that is, 0.033 seconds (1 / 30). Then, the peak time of the dynamic center of mass of the panic runner is earlier than that of the static center of mass. Taking the third peak platform in the figure as an example (from left to right), the dynamic center of mass of the panic runner reaches the quasi-peak at the twentieth frame, and the static center of mass of the panic runner reaches the peak at the twenty-second frame. The dynamic center of mass is two sampling periods ahead of the static center of mass, that is, 0.067 seconds (2 / 30). Therefore, the dynamic center of mass curve can exhibit the characteristics of accelerated running ahead of the static center of mass curve in identifying accelerated running behavior.

[0081] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for recognizing panic running behavior of pedestrians based on a dynamic center-of-mass model, characterized in that, The method comprises: frame-by-frame acquisition of panic pedestrian images from a video; establishment of a coordinate system according to the panic pedestrian images, and construction of a dynamic centroid model of the panic pedestrian; determination of whether the panic pedestrian is accelerating and running based on the dynamic centroid model, a geometric discrimination formula and a dynamic discrimination formula; wherein the geometric discrimination formula is, β1 is an included angle between a segment connecting a left hip and a left knee and a segment connecting the left knee and a left ankle, and β2 is an included angle between a segment connecting a right hip and a right knee and a segment connecting the right knee and a right ankle; the dynamic discrimination formula is, x(D.C) represents the dynamic center of mass horizontal coordinate value of the panic pedestrian, x(S.C) represents the static center of gravity horizontal coordinate value of the panic pedestrian, dir(D.C) represents the dynamic center of mass direction of the panic pedestrian, represents the panic pedestrian acceleration direction, represents the panic pedestrian speed direction.

2. The method of claim 1, wherein the method is based on a dynamic center of mass model. the construction process of the dynamic centroid model of the panic pedestrian comprises: extraction of 2D skeletal key points of the panic pedestrian from the panic pedestrian images by using a pose estimation algorithm; segment connection of the 2D skeletal key points of the panic pedestrian according to a limb connection relationship through bipartite graph matching; calculation of dynamic centroid coordinates of the panic pedestrian by using coordinates of the 2D skeletal key points of the panic pedestrian.

3. The method of claim 2, wherein the method is based on a dynamic center of mass model. the extraction process of the 2D skeletal key points of the panic pedestrian by the pose estimation algorithm comprises: input of the panic pedestrian images and the coordinate system, and identification of specific orientations of each panic pedestrian in the images; calculation of key points of human joints of the panic pedestrian through convolution layers, pooling layers and fully connected layers; numbering of the key points as the 2D skeletal key points of the panic pedestrian, and output of the 2D skeletal key points.

4. The method of claim 2, wherein the method is characterized by, the 2D skeletal key points comprise: a head top, a left ear, a right ear, a left eye, a right eye, a nose, a left mouth corner, a right mouth corner, a neck, a left shoulder, a right shoulder, a left elbow, a right elbow, a left wrist, a right wrist, a left hip, a right hip, a left knee, a right knee, a left ankle and a right ankle.

5. The method of claim 2, wherein the method is characterized by, the calculation process of the dynamic centroid coordinates of the panic pedestrian comprises: all 2D skeletal key points of a same panic pedestrian face in the panic pedestrian images are calculated and processed as one head node marked as 0; dynamic centroid coordinates of different segments are calculated respectively; the dynamic centroid coordinates of the panic pedestrian are calculated by using the segment dynamic centroid coordinates.

6. The method of claim 5, wherein the method is based on a dynamic center of mass model. the calculation processing formula of the head node is: Wherein, i is the face 2D skeleton key point serial number, (x i , y i ) is the 2D skeleton key point coordinate.

7. The method of claim 5, wherein the method is characterized by, the calculation formula of the dynamic centroid of the different segments is: where k is the number of segments, p is the number of proximal key points of a segment, d is the number of distal key points of a segment, (x k , y k ) is the dynamic centroid coordinate of a segment, (x p , y p ) is the coordinate of a proximal key point of a segment, (x d , y d ) is the coordinate of a distal key point of a segment, l p is the initial segment dynamic centroid coefficient, and l d is the terminal segment dynamic centroid coefficient.

8. The method of claim 7, wherein the method is characterized by, the segment proximal key point is a 2D skeletal key point closer to a human trunk among two 2D skeletal key points connected by a segment, the segment distal key point is a 2D skeletal key point farther away from the human trunk among the two 2D skeletal key points connected by the segment, and the initial segment dynamic centroid coefficient and the terminal segment dynamic centroid coefficient are calculated based on anatomical knowledge and statistical data, and are used to reflect contribution degrees of different body parts to the dynamic centroid of the panic pedestrian.

9. The method of claim 5, wherein the method is characterized by, the calculation formula of the dynamic centroid of the panic pedestrian is: where (x k , y k ) is the dynamic centroid coordinate of the segment, m k is the panic pedestrian body mass of the kth segment, and M is the total panic pedestrian body mass.

10. An electronic identification device, characterized in that comprise: one or more processors; a memory; one or more programs stored in the memory, the one or more programs comprising instructions for performing the method for identifying an accelerating and running behavior of a panic pedestrian based on a dynamic centroid model according to any one of claims 1-9.

Images