Method and device for dynamically analyzing human flow track

Abstract

The invention discloses a people flow track dynamic analysis method and device, and relates to the field of track processing, and the method comprises the steps: tracking people flow through video acquisition by employing angular point track statistics, carrying out the counting, and generating a people flow track diagram; based on the real path and the model path, comparing the people flow travel distribution rule of the to-be-transformed area; determining a decision-making habit of the people flow in the region by combining a people flow behavior decision-making model of the region of interest; and according to the people flow trajectory diagram and a decision analysis result, completing analysis of people flow trajectories and providing reconstruction suggestions for the to-be-reconstructed region. The invention further discloses a people stream trajectory dynamic analysis device which comprises a people stream data acquisition module, a central processing module, a people stream trajectory core element acquisition module and an identification module. Based on monitoring of the people stream motion video of the to-be-transformed area, the pedestrian trajectory tendency in the to-be-transformed area is known and analyzed from multiple angles; and the area is more closely transformed by considering the actual street state and the pedestrian tendency, so that the transformation efficiency of the old urban area is improved.

Description

Technical Field

[0001] The present invention relates to the field of trajectory processing, and in particular to a method and device for dynamic analysis of pedestrian trajectory. Background Art

[0002] Although most cities are rapidly advancing urbanization, the existing layout of the original roads and the travel habits and modes of residents in the area cannot be effectively considered in the process of rapid urbanization. Even if the old urban area is renovated after the urban advancement, the renovation results cannot be quickly accepted by the residents of the old urban area, and cannot meet the actual living needs of the residents, resulting in unsatisfactory renovation results. In the end, the renovation cost of the old urban area is high and the renovation accuracy is average. Therefore, a new method of analyzing pedestrian travel modes is urgently needed to assist the renovation of the old urban area, so as to solve the existing dilemma.

[0003] Chinese patent document CN202010933190.8 discloses "a method and device for determining regional crowd trajectory". It includes: obtaining pedestrian trajectory data of the area to be analyzed, and the first attribute information of the direction zone corresponding to the camera screen in the area to be analyzed; wherein the direction zone is the area between two business areas in the area to be analyzed, and the business area represents the area where a business is handled in the area to be analyzed, and the first attribute information includes information for indicating the direction of pedestrian flow between the two business areas; determining the pedestrian flow between the two business areas corresponding to the direction zone according to the pedestrian trajectory data and the first attribute information; determining the regional crowd trajectory corresponding to the area to be analyzed according to the pedestrian flow and the pedestrian flow direction represented by the direction zone. In the above scheme, by dividing the business area and the direction zone between the two business areas in the camera screen, the pedestrian flow between the two business areas can be determined according to the pedestrian trajectory data in the camera screen and the attribute information of the direction zone, and the corresponding regional crowd trajectory can be determined according to the pedestrian flow and the pedestrian flow direction. Therefore, there is no need to perform cross-camera recognition, and only the regional crowd trajectory needs to be analyzed for each camera’s image, thereby avoiding the problem of low accuracy in determining the regional crowd trajectory caused by the low accuracy of cross-camera tracking and recognition, that is, improving the accuracy of determining the regional crowd trajectory. However, the execution method of this patent cannot effectively provide effective transformation suggestions for urban planning and construction based on the current regional crowd flow distribution status. Summary of the invention

[0004] The present invention mainly solves the technical problem that it is impossible to efficiently and accurately perform urban area reconstruction through crowd trajectory analysis, and provides a crowd trajectory dynamic analysis method and device. The present invention effectively extracts crowd distribution information in a video through corner point trajectory statistics and generates a crowd trajectory map, and quickly obtains the travel pattern of crowds in the area by comparing the real path with the model path. The crowd behavior decision model is used to accurately capture the decision habits of pedestrians in the area when faced with choices. Based on the crowd trajectory distribution map and the crowd decision habits, the reconstruction area is reconstructed according to the habits of residents in the area, thereby improving the reconstruction efficiency.

[0005] The above technical problem of the present invention is mainly solved by the following technical solution: The present invention comprises S1 obtaining through video and using corner point trajectory statistics to track the flow of people to count and generate a flow of people trajectory map; S2 compares the distribution of passenger flow in the area to be transformed based on the real path and the model path; S3 combines the crowd behavior decision model of the area of ​​interest to confirm the decision-making habits of the crowd in the area; S4 completes the analysis of the pedestrian flow trajectory based on the pedestrian flow trajectory map and decision analysis results and puts forward renovation suggestions for the area to be renovated.

[0006] Preferably, the corner point trajectory statistics include corner point extraction and corner point tracking. During the corner point extraction, a single Gaussian distribution background model is used to separate each frame of the image, and a foreground image obtained by separation is used to track the moving object of the image using a particle filter to obtain information points with significant features in each frame of the video, and then the KLT method is used to calculate the tracking information points to complete the corner point tracking.

[0007] Preferably, when the corner point extraction step is completed and the corner point tracking is performed, a corner point of any frame is selected and its weight is set to n. The closer the points in adjacent frames are to the corner point, the closer the weights are, that is, the larger the weights are. The corner points with corresponding weights in adjacent frames are connected to generate the trajectory of the corner points, thereby obtaining a pedestrian flow trajectory map.

[0008] Preferably, the cross-section counting method, the static snapshot method, the motion trajectory method and the pedestrian tracking method are combined to compare the real path and the model path of the area to be transformed. The cross-section counting method cuts the streets of the selected area into sections and counts the actual number of pedestrians. The static snapshot method records and counts the behavioral preferences of pedestrians in different streets.

[0009] Preferably, the motion trajectory method records the direction choices of pedestrians in the statistical area at the intersection of the street; the pedestrian tracking method measures the preference of pedestrians in path selection when making a decision on the travel path, and provides multi-angle statistical feedback on the actual decision habits and non-occurring decision tendencies of pedestrians in the area through real path.

[0010] Preferably, a space syntax viewshed model is established during the model path comparison, and the spatial relationship of streets in the area is analyzed from a spatial level through the space syntax viewshed model. The equal views and visibility in the viewshed model are analyzed, and then combined with the statistical results of the real path, unreasonable pedestrian decision-making habit statistical results are filtered out and the statistical results are corrected.

[0011] Preferably, when establishing the crowd behavior decision model of the interest area, the information points with significant features in each frame of the video are first defined as interest points, and the interest points are further divided into high interest points, low interest points and no interest points according to the degree of information concentration. Different interest points are further refined and distinguished by a threshold method. When the difference between the high information point and the threshold in any area changes with the change of frame number, the interest point classification of the point is modified.

[0012] Preferably, the calculation formula of the interest area is as follows: Among them, R1 represents the radius of the high interest area; R2 represents the radius of the low interest area; 2θ represents the angle of the high interest area; θ1 and θ2 represent the angles between the center line connecting the two interest points and the pedestrian p and the moving direction of the pedestrian p, respectively; L1 and L2 represent the lateral and longitudinal distances from the pedestrian p to the interest point i; L3 and L4 represent the lateral and longitudinal distances from the pedestrian p to the interest point j.

[0013] As a preference, based on the area of ​​interest, a relationship model between crowd density and area radius of the point of interest and a crowd behavior decision model within the area of ​​interest are established. With the crowd density and area radius of the point of interest as judgment conditions, the crowd behavior decision model is used to execute a rule-based pedestrian behavior decision algorithm for the appearance of people moving in opposite and same directions at different densities in the area of ​​interest, eliminate pedestrian overlap and reconfirm the decision-making habits of pedestrians in the area on the basis of improving judgment accuracy.

[0014] A pedestrian flow trajectory dynamic analysis device, comprising: The crowd data collection module includes a number of crowd video data collection and transmission devices, which are connected to the central processing module; The central processing module processes the crowd flow trajectory data in a centralized manner to generate crowd flow analysis results; The core element acquisition module of the pedestrian flow trajectory performs video analogy processing and obtains the core element results in the pedestrian flow trajectory data; the recognition module determines the travel mode of users in different areas and time periods through the analysis results of mobile phone signaling data.

[0015] The beneficial effects of the present invention are as follows: based on monitoring the movement video of pedestrians in the area to be renovated, the present invention uses the corner point trajectory statistics method to obtain the pedestrian flow in the video frame by frame, perform statistics, and generate a pedestrian trajectory map. The corner point statistics method can not only protect the privacy of pedestrians, but also more accurately feedback the dynamic trajectory of pedestrians in the video. The distribution law of pedestrian travel in the renovated area is obtained by comparing the real path of the monitoring area with the model path. The information statistics in the actual path are combined with the model path in space to more comprehensively feedback the diversion trend of the overall pedestrian travel in the area. On the basis of understanding the trend of pedestrian flow, the interest area in the video is planned, and the interest area is used to supplement the decision-making habits of pedestrians in different areas from the level of social factors. Finally, the pedestrian trajectory tendencies in the area to be renovated are analyzed from multiple angles, and the actual street status and pedestrian tendencies are taken into account to carry out a closer transformation of the area, thereby improving the transformation efficiency of the old city. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of comparative analysis between the real path and the model path of the present invention.

[0017] Figure 2 It is a flow chart of pedestrian decision making in an area of ​​interest of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0019] Embodiment: A kind of embodiment of this embodiment is as follows Figure 1 and Figure 2 As shown, it includes obtaining the corner point trajectory statistics through the video to track the flow of people for counting and generating a flow of people trajectory map. The single Gaussian distribution background model can be used to separate the pedestrian information in the video to obtain the foreground image. For the separated foreground image, the particle filter is used to track the flow of people moving in the image, so as to obtain the information points with significant characteristics about the flow of people in the image, that is, the corner points. The commonly used corner point extraction method is the KLT method, which is widely used due to its fast calculation speed and good tracking and detection effect. This embodiment uses this method to obtain the corner points. Motion corner point detection is to find the corresponding pixel points in the continuous motion frame image. Considering two adjacent image frames with constant brightness and continuous time (small motion displacement), if all the pixel points on the corresponding window move in one direction, a new pixel coordinate is obtained, that is, the (x, y) point at time t is (x+dx, y+dy) at time t+T. Usually the corner point area data meets this requirement. The possible corner points in a frame of video image are obtained by KLT tracking, and the main corner points with large displacement changes are extracted for subsequent processing.

[0020] After performing corner point trajectory statistical tracking and extracting the main corner points of the motion area, it is necessary to establish the motion trajectory of the moving object based on this, that is, to connect the corner points that remain unchanged relative to the moving object in the previous and next frames. This embodiment solves this problem through a corner point trajectory statistical tracking algorithm. Assume that the corner points of the moving object in the previous frame can also be tracked in the next frame, and the displacement of the corner points of the previous frame in the next frame is within a certain range. Taking a main corner point as an example, the corner point trajectory statistical tracking steps are as follows: the weight of the corner point is marked as 10, and the 8 points closest to the corner point have the highest correlation with the corner point, so the weight is marked as 9. The weights of the 16 points outside these 8 points are marked as 8, and so on, until the fifth layer of points closest to the corner point, its weight is marked as 5. Although the points outside also have a certain correlation with the corner point, they have little effect on the results in subsequent calculations and processing, so the weights of the remaining unmarked points are all marked as 0. The center point is the selected corner point, and the weight is 10. If the corner point with a fixed position relative to the moving object in the first frame can also be tracked in the second frame, then assume that the corner point A in the first frame becomes a corner point with a constant position relative to the object in the second frame, and the weights of the points around A change. In the first frame, the first layer outside A is marked as 9, that is, the weight is 9. In the second frame, the relative position of corner point A is A', and B is marked as 8 in the second layer outside A', and the weight is the average of the two frame marks: (9+8) / 2=8.5. Since the displacement of the corner point with a fixed position relative to the moving object in two consecutive frames changes little, some points can have two or even more weights as the number of image frames increases. The larger the weight, the greater the correlation with this corner point. The weighted average can be used to calculate the average weight of each point in the image. Connect the corner point with the point with the largest average weight among the nearest 8 points, and so on. The corresponding corner points can be connected to obtain the trajectory movement map of the corner point. The innovation of this algorithm lies in marking the weight of the midpoint of the image according to the correlation around the corner point, and obtaining the weight change of the corner point trajectory over time, thereby obtaining the trajectory displacement map of the moving object.

[0021] After the corner point trajectory tracking is completed, the trajectory consistency crowd counting is performed. After obtaining the motion trajectory diagram of the crowd, it is necessary to judge the trajectory to determine which are the trajectories of the same moving object, and then perform crowd counting statistics. This embodiment uses a trajectory consistency clustering method to achieve the above goals. The crowd movement is mainly divided into n directions. The crowd advances with a certain purpose and process. Therefore, the movement trajectory is slow and changes slightly. There are rarely sudden jumps in the direction of the previous and next frames. Through this algorithm, it can be inferred that the trajectory moves along the above n directions is the same moving object. For two or more people moving in the same direction at the same time, the cross-sectional area of ​​the same movement direction is detected. If it exceeds a certain range, it can be determined as two or more people. In this way, the number of people within a certain range can be calculated.

[0022] After the pedestrian flow trajectory map is drawn, the distribution of pedestrian travel in the area to be transformed is compared based on the real path and the model path, and the allocation of the actual measurement point survey and data processing are determined. When the cross-sectional counting method is used for field research, there are 7 members, each of whom is assigned 5 observation points. A two-minute video is recorded for each observation point during the time period, and a total of 12 sets of data are obtained for each observation point. In order to reduce the influence of other interference factors, the survey data at different times are recorded and summarized separately, and the average of six times is taken, and then multiplied by 30 to obtain the hourly pedestrian flow as the actual hourly pedestrian flow data of the street. At the static snapshot method level, the behavioral activities in the urban street space are classified through static snapshots, the quality of the street environment and the impact of spatial layout on people's behavior are explored, and the street sections in the study area are numbered. During execution, the selection of research tools and task allocation are determined. The number of people who choose different directions of each intersection is counted in the form of recorded videos. At the intersection of the road, the streets are marked as four movement directions A, B, C, and D in a clockwise direction, which represent different movement directions of pedestrians. At the pedestrian tracking level, the starting point and destination of the pedestrian path are determined. The starting point of the pedestrian is determined as the unit door where the respondent lives, and the destination is set as three types of facilities that are frequently used in daily life, such as schools, restaurants, and bus stops. Two methods are used to obtain pedestrian path data: one is to extract pedestrians' reported paths through structured questionnaires; the other is to select pedestrians on site for tracking and draw pedestrian path maps. Two methods are adopted and the data are supplemented with each other.

[0023] The above four methods are used to collect data on real paths, and the distribution patterns of pedestrians in street space are presented from the perspective of data. The cross-sectional counting method is used to find differences in pedestrian flow distribution at different times, and to determine the attractiveness of streets of different road grades to pedestrians. The static snapshot data is used to determine the total amount of pedestrian behavior on urban secondary roads and the degree of influence of time. The motion trajectory method is used to count the direction choices of different roads at intersections, and it is found that streets with a larger proportion of pedestrian choices are associated with high-frequency commercial facilities. From the perspective of pedestrian tracking data, the choice of paths with public transportation points as destinations clearly defines the number of relatively obvious routes. The four field survey methods fully explain the pedestrian activities in the street space in the study area.

[0024] In addition to field research to understand the distribution of pedestrians in real paths, model path simulation analysis calibrates the correctness of pedestrian distribution from the existing street level. When model path simulation analysis is performed, when establishing a space syntax line segment model, it is necessary to first determine the area to be renovated, then determine the buffer area of ​​the area to be renovated, then draw the axis map of the area, and finally convert the axis model into a line segment model. In the space syntax line segment model, the maximum and average values ​​of the standardized angle penetration and standardized angle integration of the line segment model under different radii are analyzed respectively. It is expressed by the high and low correlation coefficients of the two types of parameters, penetration and integration. That is, the line segment model selects quantitative parameters, among which penetration measures the penetration and selectivity potential of urban streets; integration reflects the accessibility of streets in the urban system. In the study of urban street space, penetration and accessibility are important indicators of daily travel. They can directly compare cities of different sizes and shapes, which is of great significance to the quantitative study of urban space. Based on the results of the real path survey, the degree of penetration and integration during the model path analysis are used as consideration indicators to complement each other to judge the crowd decision-making tendencies and travel distribution patterns in the video, thereby improving the accuracy of grasping the crowd distribution patterns.

[0025] This embodiment also includes combining the pedestrian behavior decision model of the interest area to confirm the decision-making habits of pedestrians in the area. In actual road reconstruction, in addition to the actual street structure pattern, social factors in the actual environment also affect pedestrians' decisions. The pedestrian interaction range usually refers to the range of pedestrians' visual contact and communication with other pedestrians during walking. This range can be understood as the range of pedestrians' perception and influence on other pedestrians, and can also be called the pedestrian's social range or perception range. The pedestrian's perception ability will affect the pedestrian's decision-making and actions, thereby affecting the behavior and movement of the entire crowd. Specifically, the pedestrian's perception ability will affect their cognition and understanding ability, such as the ability to perceive obstacles, paths, other pedestrians, etc. in the environment. These perception information will be used by pedestrians to make decisions, such as choosing a path, adjusting speed, avoiding obstacles, etc. At the same time, the pedestrian's perception ability will also affect their understanding and judgment of other pedestrians, such as predicting the actions of other pedestrians, avoiding collisions with other pedestrians, etc. These decisions and actions will produce a chain reaction in the entire crowd, thereby affecting the behavior and movement of the entire crowd. In actual situations, pedestrians do not care about the global information in the motion scene, but only pay attention to the local area information of their own interest. Pedestrians interact with static information such as buildings, obstacles, pedestrian facilities, and other dynamic information such as other pedestrians in a local area to determine the pedestrian's movement path. Different interaction ranges will cause pedestrians to choose different paths. Therefore, there is no need to consider the global information of the motion scene, but only extract the local area information that the pedestrian is interested in. This embodiment fills the pedestrian's points of interest into interest areas through the flow tendency of the crowd in the video, and distinguishes the size of the attention of the points of interest, to achieve different divisions of the interest areas, which can more accurately describe the interaction range and behavior of pedestrians. This helps to further grasp the law of pedestrian flow.

[0026] When determining the region of interest, firstly, the pedestrian flow video is recorded synchronously in the horizontal and overhead view to collect the pedestrian movement scene information; secondly, the pedestrian flow video recorded horizontally is observed by an eye tracker to obtain the pedestrian's gaze frequency and gaze duration and other eye movement behavior data related to the gaze intensity; then, the time threshold method is used to extract the points of interest and remove the outliers; finally, the pedestrian attention formula is introduced to measure the degree of stimulation of static and dynamic factors on pedestrians, and the different divisions of the region of interest are achieved by clustering internal filling. Among them, the points of interest are points with rich and high information content. These points are the most representative points in the pedestrian movement scene, and are points that select useful information from a lot of information. They are the guarantee of the effectiveness of pedestrian walking perception. The points of interest include high interest points, low interest points, and no interest points. The division of points of interest can be achieved by various methods. Low interest points are distinguished by the threshold method, and high interest points are distinguished by introducing the attention formula. The point with the highest attention is the high interest point. When the pedestrian's gaze time on a certain point exceeds the threshold Ti, the gaze point is considered to be an interest point. The attention formula is introduced to measure the degree of stimulation of static and dynamic factors on pedestrians, which represents the attention of pedestrians to the interest point i at the current position. The calculation method is as follows: Among them, N(i,p) represents the number of people at the point of interest i, N(p) represents the total number of subjects, con(c,i) represents a further comparison of the attention of the point of interest; N(p,i) represents the frequency of the subject's gaze at the point of interest i at the current pedestrian p position; N(I) represents the total frequency of the subject's interest points at the current pedestrian p position. The interest area is composed of three parameters: interest level, radius, and angle.

[0027] The interest area can be divided into high interest area, low interest area and no interest area. Considering that the no interest area has no effect on the pedestrian movement pattern, this embodiment mainly considers the high interest area and the low interest area. The interest point is considered to be a high interest point if the subject's attention to a certain interest point in the continuous number of frames is calculated by the attention formula. The distance from the position of the high interest point in the first frame to the pedestrian at the current position is the radius of the high interest area R1, the maximum distance from all the interest points of the pedestrian to the pedestrian at the current position is the radius of the low interest area R2, and the maximum angle formed by two points of all the interest points of the subject and the pedestrian at the current position is the high interest area angle 2θ. In the process of interest point detection, some pedestrians or obstacles are not regarded as interest points, but these points are actually part of the interest area and can be regarded as interest points through the internal filling method. The points of interest of the pedestrian at the current pedestrian p position are a, b, c, d, e, and f. Pedestrian b is the subject's high interest point. The distance from the pedestrian p position to the interest point b is recorded as the high interest area radius R1, and the distance from the pedestrian p position to the farthest interest point f is recorded as the low interest area radius R2. The angle of the high interest area is the angle between the interest points a and c and the current pedestrian p. However, the components of the interest area are clustered and filled as interest points. The angle formed by the current pedestrian p and the non-interest point l and the interest point a is less than 2θ, but the distance from the non-interest point l to the current pedestrian p is greater than the low interest area radius R2, so it cannot be clustered and filled as an interest point. The distance from the non-interest point m to the current pedestrian p is less than the low interest area radius R2, but the angle formed by the current pedestrian p and the non-interest point m and the interest point a is greater than 2θ, so it cannot be clustered and filled as an interest point. Assume that the high interest point of pedestrian p is i, the maximum distance interest point is j, and pedestrian p forms the maximum angle with interest points i and j (the assumption does not represent the actual situation). The calculation formula of the region of interest parameters is as follows: Among them, R1 represents the radius of the high interest area; R2 represents the radius of the low interest area; 2θ represents the angle of the high interest area; θ1 and θ2 represent the angles between the center line connecting the two interest points and the pedestrian p and the moving direction of the pedestrian p, respectively; L1 and L2 represent the lateral and longitudinal distances from the pedestrian p to the interest point i; L3 and L4 represent the lateral and longitudinal distances from the pedestrian p to the interest point j.

[0028] In general, pedestrians want to reach their destination as quickly as possible, so they usually choose to go directly to their destination. However, pedestrians may encounter obstacles or slower pedestrians during their travel, which may lead to conflicts. In order to avoid conflicts, pedestrians need to seek a balance between conflict avoidance and detour distance to choose the optimal speed and direction. In addition, pedestrians also need to maintain a sufficient safety distance and adjust their moving speed to avoid obstacles. However, in the actual movement process, pedestrians have their own thoughts, and their behavioral decisions are affected by various factors. The same pedestrian may make different behavioral decisions at different times in the same environment. According to the detection results of people in the area of ​​interest, a rule-based algorithm is used in the high interest area and the low interest area to ensure that the area boundary transitions naturally and establish an algorithm for pedestrian behavior decision-making. In this embodiment, according to the distance statistics of the area of ​​interest, the pedestrian behavior decision in the area of ​​interest is modeled, and the modeling method of the agent is used to model the area of ​​interest, the behavior decision modeling and the feedback modeling are respectively performed. The mathematical model of the radius of each area of ​​interest and the local pedestrian flow density is corrected to determine the relationship between the radius of the high interest area and the radius of the low interest area and the local pedestrian flow density. Taking into account the influence of vacancies, the angle of the high-interest area and the local pedestrian flow density were corrected, so as to re-establish the relationship model between the local pedestrian flow density and the radius of the high-interest and low-interest areas. Secondly, the pedestrian behavior decision-making model was established in the area of ​​interest. According to the personnel detection judgment conditions in the area of ​​interest, a rule-based pedestrian behavior decision algorithm was established for the appearance of people in opposite and same directions at different densities in the area of ​​interest, and the method of eliminating the overlap of pedestrian positions was supplemented to improve the accuracy of the judgment results. Finally, the analysis of the pedestrian flow trajectory was completed according to the pedestrian flow trajectory map and the decision analysis results, and modification suggestions for the area to be transformed were put forward.

Claims

1. A method for dynamic analysis of pedestrian flow trajectories, characterized in that: The following steps are involved: S1 uses corner point trajectory statistics to track the flow of people through video acquisition and count them to generate a flow trajectory map; S2 compares the distribution pattern of pedestrian flow in the area to be transformed based on the real path and the model path; S3 combines the crowd behavior decision model of the area of ​​interest to confirm the decision-making habits of the crowd in the area; S4 completes the analysis of pedestrian flow trajectories based on the pedestrian flow trajectory map and decision analysis results and proposes renovation suggestions for the areas to be renovated.

2. A method for dynamic analysis of pedestrian flow trajectories according to claim 1, characterized in that: The corner point trajectory statistics include corner point extraction and corner point tracking. During corner point extraction, a single Gaussian distribution background model is used to separate each frame of the image. The separated foreground image is used to track the moving object in the image using a particle filter to obtain information points with significant features in each frame of the video. The KLT method is then applied to calculate and track the information points to complete corner point tracking.

3. A method for dynamic analysis of pedestrian flow trajectories according to claim 2, characterized in that: When the corner point extraction step is completed and corner point tracking is performed, a corner point of any frame is selected and its weight is set to n. The closer the point in the adjacent frame is to the corner point, the closer the weight is, that is, the larger the weight is. The corner points with corresponding weights in the adjacent frames are connected to generate the trajectory of the corner point, and then a pedestrian trajectory map is obtained.

4. A method for dynamic analysis of pedestrian flow trajectories according to claim 1, characterized in that: The cross-section counting method, the static snapshot method, the motion trajectory method and the pedestrian tracking method are combined to compare the actual path and the model path of the area to be transformed. The cross-section counting method is used to cut the streets of the selected area into sections and count the actual number of pedestrians. The static snapshot method is used to record and count the behavioral preferences of pedestrians in different streets.

5. A method for dynamic analysis of pedestrian flow trajectories according to claim 4, characterized in that: The motion trajectory method records the direction choices of pedestrians in the statistical area when they are at street intersections; the pedestrian tracking method measures the path selection preferences of pedestrians when making travel path decisions, and uses multi-angle statistics of real paths to feedback the actual decision habits and non-occurring decision tendencies of pedestrians in the area.

6. A method for dynamic analysis of pedestrian flow trajectories according to claim 5, characterized in that: When comparing the model paths, a space syntax viewshed model is established. The spatial relationship of streets in the area is analyzed from a spatial level through the space syntax viewshed model. The iso-viewshed and visibility in the viewshed model are analyzed. Then, combined with the statistical results of the real paths, unreasonable pedestrian decision-making habit statistical results are filtered out and the statistical results are corrected.

7. A method for dynamic analysis of pedestrian flow trajectories according to claim 2, characterized in that: When establishing the crowd behavior decision model for the interest area, the information points with significant features in each frame of the video are first defined as interest points, and then the interest points are divided into high-interest points, low-interest points and no-interest points according to the degree of information aggregation. Different interest points are further refined and distinguished by the threshold method. When the difference between the high-information point and the threshold in any area changes with the change of frame number, the interest point classification of this point is modified.

8. A method for dynamic analysis of pedestrian flow trajectories according to claim 7, characterized in that: The calculation formula of the interest area is as follows: Among them, R1 represents the radius of the high interest area; R2 represents the radius of the low interest area; 2θ represents the angle of the high interest area; θ1 and θ2 represent the angles between the center line connecting the two interest points and pedestrian p and the movement direction of pedestrian p, respectively; L1 and L2 represent the horizontal and vertical distances from pedestrian p to interest point i; L3 and L4 represent the horizontal and vertical distances from pedestrian p to interest point j.

9. A method for dynamic analysis of pedestrian flow trajectories according to any one of claims 7 or 8, characterized in that: Based on the area of ​​interest, a relationship model between crowd density and area radius of the point of interest and a crowd behavior decision model within the area of ​​interest are established. Using crowd density and area radius of the point of interest as judgment conditions, the crowd behavior decision model is used to execute a rule-based pedestrian behavior decision algorithm for the occurrence of people moving in opposite and same directions at different densities in the area of ​​interest, eliminating pedestrian overlap and reconfirming the decision-making habits of pedestrians in the area on the basis of improving judgment accuracy.

10. A pedestrian trajectory dynamic analysis device, adapted for the method for identifying a mode of transportation according to any one of claims 1 to 9, characterized in that: include: The crowd data collection module includes several crowd video data collection and transmission devices, which are connected to the central processing module; The central processing module processes the crowd flow trajectory data in a centralized manner to generate crowd flow analysis results; The core element acquisition module of the pedestrian flow trajectory performs video analogy processing and obtains the core element results in the pedestrian flow trajectory data; The identification module determines the travel mode of users in different areas and time periods through the analysis of mobile phone signaling data.

Images