City-level large-scale crowd activity track generation method based on mobile phone signaling

Abstract

The invention provides a city-level large-scale crowd activity track generation method based on mobile phone signaling, and belongs to the field of intelligent transportation, and the method comprises the steps: S1, obtaining signaling data of a specific crowd in a selected city in one day, carrying out the preprocessing operation, and carrying out the grid division of a map region; s2, processing each preprocessed track sequence by using a mask auto-encoder model to obtain the probability that each track point belongs to different grids; s3, constructing a loss function of trajectory accuracy and distribution diversity, and finally obtaining parameters of a pre-training stage model by reconstructing training parameters of mask trajectory points; s4, generating a complete final track by using the pre-training stage model; and S5, covering the corresponding latitude and longitude coordinates of the specific personnel at a certain moment in the generated trajectory points, and finally forming a complete final trajectory of the specific personnel. According to the method disclosed by the invention, the large-scale crowd activity trajectory data consistent with the original signaling trajectory data in distribution can be generated on the premise of not exposing personal privacy.

Description

Technical Field

[0001] The present invention belongs to the field of smart transportation, and in particular relates to a method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling. Background Art

[0002] In recent years, with the popularity of mobile devices, especially mobile phones, user trajectory data collected by network operators has become an important source of information for studying individual and group mobility patterns. Based on the mobile big data extracted from mobile device communications, it contains rich user geographic information and time information, which greatly promotes the study of the mobility trajectories of millions of individual users within a few months and provides an important basis for the semantic analysis of individual behavior. Analysis based on user trajectory data is of great significance to the research and application of big data such as smart cities.

[0003] However, since mobile phone signaling data contains a large amount of personal information and sensitive data, the processing and use of this data must strictly comply with privacy protection laws and regulations. Therefore, this type of data is strictly limited to a few internal departments of communication operators in most cases. Due to the lack of extensive access to signaling data, researchers face many difficulties when conducting research in related fields. This limitation in data acquisition not only affects the depth and breadth of research, but also delays the accumulation of knowledge and technological progress in this field. Therefore, how to fully release the huge capacity of signaling data while protecting personal privacy, and how to carry out geospatial analysis and population mobility statistics without involving privacy has become an urgent problem to be solved. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling, comprising the following steps:

[0005] Step S1: obtaining the signaling data of a specific group of people in a selected city for one day, performing a preprocessing operation on the signaling data, and dividing the map area into grids to obtain a preprocessed trajectory sequence of equal length;

[0006] Step S2: Construct a masked autoencoder model, including: a mask module, an encoder, a decoder and a loss function module; first, use the mask module to sample different mask rates greater than 0.5 for each of the preprocessed trajectory sequences, mask each of the preprocessed trajectory sequences according to the mask rate, and obtain a masked trajectory sequence; pass the masked trajectory sequence through the encoder and the decoder to obtain the probability that each trajectory point belongs to a different grid ;

[0007] Step S3: construct the loss function of trajectory accuracy and distribution diversity, and finally obtain the parameters of the pre-training stage model by reconstructing the training parameters of the mask trajectory points;

[0008] Step S4: in the trajectory generation stage, the pre-training stage model is used to generate trajectory points in a step-by-step manner, with a certain number of trajectory points generated in each step until the obtained trajectory points form a complete final trajectory;

[0009] Step S5: Given the corresponding longitude and latitude coordinates of a specific person at a certain moment, input them into the initial trajectory sequence position corresponding to that moment. On the basis of step S4, after each step of generating trajectory points, overwrite the above-known corresponding longitude and latitude coordinates of the specific person at a certain moment in the generated trajectory points. Repeat multiple steps of generating trajectory points according to step S4 until a complete final trajectory of the specific person is finally formed. Repeat step S5 until a predetermined number of final trajectories of specific persons are generated. The collection of trajectories is the city-level large-scale crowd activity trajectory.

[0010] Beneficial effects:

[0011] 1. The present invention designs a loss function for trajectory accuracy and distribution diversity, which considers the authenticity and diversity of the network-generated trajectory respectively. Existing methods only consider the error of predicting user trajectories when generating trajectories, and tend to generate relatively consistent trajectories when finally generating trajectories. They can only represent the most likely trajectory of individual users, but cannot describe the diverse trajectories of user groups. The loss function proposed by the present invention can effectively solve this problem.

[0012] 2. The present invention designs a multi-step autoregressive trajectory generation architecture that generates trajectory points incrementally, which can balance the quality of the generated trajectory and the time required. When generating the final trajectory, the existing method only generates a fixed number of trajectory points per step, and a large number of trajectory points generated per step will result in a low quality of the generated trajectory. Since the number of trajectory points generated in the end does not change, a small number of trajectory points generated per step will result in a large number of steps and a long generation time. The multi-step generation method for incrementally generating trajectory points proposed in the present invention can improve the quality of the generated trajectory in the early stage, and also reduce the trajectory generation time in the later stage.

[0013] 3. The present invention designs a method for generating trajectories based on specific spatiotemporal information. By inputting the user's current spatiotemporal position into the model, a specific trajectory can be obtained. Existing methods cannot insert the user's spatiotemporal information into the model, and cannot generate user trajectories with specific spatiotemporal characteristics. The trajectory generation method proposed in the present invention can solve this problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1It is a flow chart of a method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling according to the present invention;

[0015] Figure 2 This is a framework diagram of the method for generating large-scale city-level crowd activity trajectories based on mobile phone signaling. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] Embodiment 1

[0018] like Figure 1 As shown, an embodiment of the present invention provides a method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling, comprising the following steps:

[0019] Step S1: obtaining the signaling data of a specific group of people in the selected city for one day, performing preprocessing operations on the signaling data, and dividing the map area into grids to obtain a preprocessed trajectory sequence of equal length;

[0020] Step S2: Construct a masked autoencoder model, including: a mask module, an encoder, a decoder, and a loss function module; first, use the mask module to sample different mask rates greater than 0.5 for each preprocessed trajectory sequence, mask each preprocessed trajectory sequence according to the mask rate, and obtain a masked trajectory sequence; pass the masked trajectory sequence through the encoder and decoder to obtain the probability that each trajectory point belongs to a different grid ;

[0021] Step S3: construct the loss function of trajectory accuracy and distribution diversity, and finally obtain the parameters of the pre-training stage model by reconstructing the training parameters of the mask trajectory points;

[0022] Step S4: in the trajectory generation stage, the pre-training stage model is used to generate trajectory points in a step-by-step manner, with a certain number of trajectory points generated in each step until the obtained trajectory points form a complete final trajectory;

[0023] Step S5: Given the corresponding longitude and latitude coordinates of a specific person at a certain moment, input them into the initial trajectory sequence position corresponding to that moment. On the basis of step S4, after each step of generating trajectory points, overwrite the above-known corresponding longitude and latitude coordinates of the specific person at a certain moment in the generated trajectory points. Repeat multiple steps of generating trajectory points according to step S4 until a complete final trajectory of the specific person is finally formed. Repeat step S5 until a predetermined number of final trajectories of specific persons are generated. The collection of trajectories is the city-level large-scale crowd activity trajectory.

[0024] In one embodiment, the above step S1: obtaining the signaling data of a specific group of people in a selected city for one day, performing preprocessing operations on the signaling data, and dividing the map area into grids to obtain a preprocessed trajectory sequence of equal length, specifically includes:

[0025] Step S11: Acquire original signaling data, including: user ID, base station ID, base station longitude, base station latitude, arrival time and departure time;

[0026] Step S12: Group the user IDs, sort all the original signaling data in ascending order according to the arrival time, and convert the original signaling data of each user into sequence data of equal length in units of 15 minutes. The length of the sequence data is set to , the sequence data of one day corresponds to The value is 24;

[0027] Step S13: According to simple rules, the signaling data of users with ping-pong or drifting phenomena or those who have been in the same area for a given period of time and have not traveled are eliminated. Specifically, if the difference between the distance between the i+1th trajectory point and the i-th trajectory point is greater than 3 km, and the distance between the jth trajectory point and the i-th trajectory point is less than 1.5 km in the subsequent trajectory changes within 20 minutes or within 5 times, it indicates that the user data is abnormal, and the i+1 to j-1th trajectory points can be eliminated;

[0028] Step S14: Divide the map area into 100m×100m grids, regard the base stations in the same grid as the same base station, set the ID of each grid, and obtain a pre-processed trajectory sequence. Each pre-processed trajectory sequence includes: user ID, grid ID, and current time.

[0029] In one embodiment, the above step S2: constructs a masked autoencoder model, including: a mask module, an encoder, a decoder and a loss function module; first, the mask module is used to sample different mask rates greater than 0.5 for each preprocessed trajectory sequence, and each preprocessed trajectory sequence is masked according to the mask rate to obtain a masked trajectory sequence; the masked trajectory sequence is passed through the encoder and the decoder to obtain the probability that each trajectory point belongs to a different grid , specifically including:

[0030] Step S21: Sampling a probability according to Gaussian distribution , from the preprocessed trajectory sequence Medium Sampling Masks of different indicators are used to perform mask operations, where is the length of the preprocessed trajectory sequence, as shown in formulas (1) to (2):

[0031] (1)

[0032] (2)

[0033] in, represents the trajectory sequence after mask processing, express The masked trajectory sequence in , that is, the corresponding grid ID is designated as ; express The unmasked trajectory sequence in ; is the number of grids;

[0034] Step S22: define the code of each grid as , add position coding , which is finally mapped to dimensional vector, as shown in formula (3), which uses the encoding of each grid plus the position encoding in the sequence to represent the initial sequence:

[0035] (3)

[0036] in, The superscript 0 indicates that this is the encoding of level 0; Indicates positional encoding; represents one-hot encoding;

[0037] Step S23: The proportion of randomly discarded part of the masked data, and its index set is expressed as , Indicates data that has not been discarded;

[0038] (4)

[0039] exist Connect an extra class vector at the beginning , which constitutes the input of the encoder:

[0040] (5)

[0041] Step S24: Input encoder, the encoder consists of four sub-layers in order: layer normalization, multi-head self-attention layer, layer normalization and feedforward neural network; first, Layer Input Perform layer normalization and pass it into the multi-head self-attention layer as input to obtain ; Then use residual connection to input Added to the output of the multi-head self-attention layer, we get ; After layer normalization and feed-forward neural network layer, we get ; Finally, stack the encoders After the layer, the layer is normalized again to achieve trajectory embedding coding, as shown in formulas (6) to (9):

[0042] (6)

[0043] (7)

[0044] (8)

[0045] (9)

[0046] in, is layer normalization, is a multi-head self-attention layer, is a feed-forward neural network, is the output of the encoder;

[0047] Step S25: Enter the decoder module and use the decoder to fill in the previously blocked positions. Map the encoder space to the decoder space, is the bias term of the mapping, represents the trajectory points that have not been discarded, Represents the code of the masked position in the prediction step S21, and then introduces a new position code , and pass the current representation to a decoder with the same structure as the encoder but different parameters to obtain the final decoder output , as shown in formulas (10) to (14):

[0048] (10)

[0049] (11)

[0050] (12)

[0051] (13)

[0052] (14)

[0053] Step S26: Through an additional linear layer , a GELU activation function and a layer normalization to convert it into the encoding ,Will and Multiply the map back to the grid to get a weight coefficient for each grid class and add this weight coefficient to an additional bias And process it through the Softmax layer to get each trajectory point belonging to different Probability , as shown in formulas (15) to (17):

[0054] (15)

[0055] (16)

[0056] (17).

[0057] In one embodiment, the above step S3: constructing a loss function of trajectory accuracy and distribution diversity, and finally obtaining the parameters of the pre-training stage model by reconstructing the training parameters of the mask trajectory points, specifically includes:

[0058] Step S31: represents the likelihood probability that the i-th trajectory point belongs to different grids, Represents a grid with masked operation, using trajectory accuracy loss function Predict the true position of the masked trajectory points, As shown in formula (18):

[0059] (18)

[0060] in, Represents the likelihood probability that the mask trajectory points belong to different grids;

[0061] Trajectory Accuracy Loss Function The goal is to maximize the log-likelihood of the actual position of the mask trajectory point, ensuring that the probability of the correct position is maximized;

[0062] Step S32: To obtain more diverse trajectories, design a distribution diversity loss function As shown in formula (19):

[0063] (19)

[0064] Since only the trajectory accuracy loss function is used This may lead to insufficient diversity in the model output. In order to obtain trajectories with more diverse distributions, the present invention designs a distribution diversity loss function: , if the probability distribution is concentrated on a certain value, the logarithmic mean is larger; if the probability distribution is more uniform, the logarithmic mean is smaller. This loss function regularizes the probability distribution, thereby enhancing the distribution diversity of the model;

[0065] Step S33: Combining the above two different loss functions, a loss function that comprehensively considers the accuracy and distribution diversity of the reconstructed trajectory is obtained. for:

[0066] (20)

[0067] in, is a parameter; is set to 0.001;

[0068] Step S34: Use loss function Perform parameter training and back propagation of the model to obtain the parameters of the updated model. The pre-training process ends and the pre-trained model parameters are retained. .

[0069] In one embodiment, the above step S4: trajectory generation stage, using the pre-training stage model, adopts a method of gradually increasing the generation of trajectory points, and each step generates a certain number of trajectory points incrementally until the obtained trajectory points form a complete final trajectory, specifically including:

[0070] Step S41: The pre-trained model parameters The conditional probability model defined is denoted as , the generation process starts from At the beginning, all the data initially input into the pre-trained model are masked and used as the input of the first step through the pre-trained model to obtain the first generated trajectory sequence. A certain proportion of trajectory points at the beginning of the sampling sequence are used as the trajectory points generated in the first step. , and then As a condition, The subsequent trajectory point mask is used as the input of the second step through the pre-trained model to obtain the second generated trajectory sequence. The trajectory point data with a larger proportion at the beginning of the sampling sequence is used as the trajectory point generated in the second step. sampling , and so on, the sampling ratio of each step is calculated in step S42;

[0071] Step S42: Sampling ratio per round As shown in formula (21):

[0072] (twenty one)

[0073] in, Indicates the current sampling round number, represents the total number of sampling rounds, in the embodiment of the present invention Take 12; relatively The variable is monotonically increasing, The value range of is between 0 and 1, and is used to indicate the sampling ratio of each step in step S41;

[0074] Since a relatively small number of trajectory points are selected in the initial sampling, and the number of trajectory points is increased in the subsequent sampling, in the early stage, due to the small number of trajectory points, only a relatively small number of trajectory points are generated, so as to avoid introducing large errors due to insufficient trajectory information and reducing the quality of trajectory generation. In the later stage, in order to reduce the number of steps of trajectory generation, it is necessary to increase the number of trajectory points generated in each step and reduce the running time. The sampling ratio designed by the embodiment of the present invention can meet the balance requirements of the quality of generated trajectories and the time of generated trajectories.

[0075] Step S43: According to the sampling ratio , according to the sampling method of step S41, from Continuous sampling , until the sum of the sampling ratios of the trajectory is 1, the number of generated trajectory points is equal to the number of original trajectory points, forming a complete trajectory sequence, and all the trajectories obtained in the last step are taken out as a complete final trajectory.

[0076] In one embodiment, the above step S5: given the corresponding latitude and longitude coordinates of a specific person at a certain moment, and inputting them into the initial trajectory sequence position corresponding to the moment, on the basis of step S4, after each step of generating trajectory points, the above-known corresponding latitude and longitude coordinates of the specific person at a certain moment are covered in the generated trajectory points, and multiple steps of generating trajectory points are repeated according to step S4 until a complete final trajectory of the specific person is finally formed, and step S5 is repeated until a predetermined number of final trajectories of the specific person are generated, and the set of the final trajectories is the city-level large-scale crowd activity trajectory, which specifically includes:

[0077] Step S51: Set the residence and work place of specific personnel according to the corresponding tasks, generate large-scale crowd activity trajectories, set the spatial position of part of the evening time period in the input sequence of step S4 to the residence in advance, set the spatial position of part of the daytime time period to the residence, and set other positions as masks for subsequent trajectory generation;

[0078] Step S52: Based on step S4, After the wheel trajectory is generated, to The track of the time period is retained, and the tracks of the evening and daytime preset in step S51 are overlaid on the retained track. The overlaid track is used to continue to generate the i+1th round using step S4 until the sum of the sampling ratios of the multi-step tracks is 1. After the final complete track is obtained, the step of generating the track is terminated;

[0079] Step S53: Repeat steps S51 to S52 until the final trajectories of a predetermined number of specific persons are obtained.

Claims

1. A method for generating large-scale urban crowd activity trajectories based on mobile phone signaling, characterized in that: include: Step S1: obtaining the signaling data of a specific group of people in a selected city for one day, performing a preprocessing operation on the signaling data, and dividing the map area into grids to obtain a preprocessed trajectory sequence of equal length; Step S2: Construct a masked autoencoder model, including: a mask module, an encoder, a decoder and a loss function module; first, use the mask module to sample different mask rates greater than 0.5 for each of the preprocessed trajectory sequences, mask each of the preprocessed trajectory sequences according to the mask rate, and obtain a masked trajectory sequence; pass the masked trajectory sequence through the encoder and the decoder to obtain the probability that each trajectory point belongs to a different grid ; Step S3: construct the loss function of trajectory accuracy and distribution diversity, and finally obtain the parameters of the pre-training stage model by reconstructing the training parameters of the mask trajectory points; Step S4: in the trajectory generation stage, the pre-training stage model is used to generate trajectory points in a step-by-step manner, with a certain number of trajectory points generated in each step until the obtained trajectory points form a complete final trajectory; Step S5: Given the corresponding longitude and latitude coordinates of a specific person at a certain moment, and input them into the initial trajectory sequence position corresponding to that moment, on the basis of step S4, after each step of generating trajectory points, the corresponding longitude and latitude coordinates of the specific person at a certain moment known above are covered in the generated trajectory points, and multiple steps of generating trajectory points are repeated according to step S4 until a complete final trajectory of the specific person is finally formed, and step S5 is repeated until a predetermined number of final trajectories of the specific person are generated, and the collection of the final trajectories is the city-level large-scale crowd activity trajectory.

2. The method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling according to claim 1 is characterized in that: The step S2: constructs a masked autoencoder model, including: a mask module, an encoder, a decoder and a loss function module; first, the mask module is used to sample different mask rates greater than 0.5 for each of the preprocessed trajectory sequences, and each of the preprocessed trajectory sequences is masked according to the mask rate to obtain a masked trajectory sequence; the masked trajectory sequence is passed through the encoder and the decoder to obtain the probability that each trajectory point belongs to a different grid , including: Step S21: Sampling a probability according to Gaussian distribution , from the preprocessed trajectory sequence Medium Sampling Masks of different indicators are used to perform mask operations, where is the length of the preprocessed trajectory sequence, as shown in formulas (1) to (2): (1) (2) in, represents the trajectory sequence after mask processing, express The masked trajectory sequence in express The unmasked trajectory sequence in ; is the number of grids; Step S22: define the code of each grid as , add position coding , which is finally mapped to dimensional vector, as shown in formula (3), which uses the encoding of each grid plus the position encoding in the sequence to represent the initial sequence: (3) in, The superscript 0 indicates that this is the encoding of level 0; Indicates positional encoding; represents one-hot encoding; Step S23: The proportion of randomly discarded part of the masked data, and its index set is expressed as , Indicates data that has not been discarded; (4) exist Connect an extra class vector at the beginning , which constitutes the input of the encoder: (5) Step S24: Input encoder, the encoder consists of four sub-layers in order: layer normalization, multi-head self-attention layer, layer normalization and feedforward neural network; first, Layer Input Perform layer normalization and pass it into the multi-head self-attention layer as input to obtain ; Then use residual connection to input Added to the output of the multi-head self-attention layer, we get ; After layer normalization and feed-forward neural network layer, we get ; Finally, stack the encoders After the layer, the layer is normalized again to achieve trajectory embedding coding, as shown in formulas (6) to (9): (6) (7) (8) (9) in, is layer normalization, is a multi-head self-attention layer, is a feed-forward neural network, is the output of the encoder; Step S25: Enter the decoder module, use the decoder to fill in the previously masked positions, mapping the encoder space to the decoder space, is the bias term of the mapping, represents the trajectory points that have not been discarded, Represents the code of the masked position in the prediction step S21, and then introduces a new position code , and pass the current representation to a decoder with the same structure as the encoder but different parameters to obtain the final decoder output , as shown in formulas (10) to (14): (10) (11) (12) (13) (14) Step S26: Through an additional linear layer , a GELU activation function and a layer normalization to convert it into the encoding ,Will and Multiply the map back to the grid to get a weight coefficient for each grid class and add this weight coefficient to an additional bias And process it through the Softmax layer to get each trajectory point belonging to different Probability , as shown in formulas (15) to (17): (15) (16) (17)。 3. The method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling according to claim 2 is characterized in that: The step S3: constructs a loss function of trajectory accuracy and distribution diversity, and finally obtains the parameters of the pre-training stage model by reconstructing the training parameters of the mask trajectory points, specifically including: Step S31: represents the likelihood probability that the i-th trajectory point belongs to different grids, Represents a grid with masked operation, using trajectory accuracy loss function Predict the true position of the masked trajectory points, As shown in formula (18): (18) in, Represents the likelihood probability that the mask trajectory points belong to different grids; Step S32: To obtain more diverse trajectories, design a distribution diversity loss function As shown in formula (19): (19) Step S33: Combining the above two different loss functions, a loss function that comprehensively considers the accuracy and distribution diversity of the reconstructed trajectory is obtained. for: (20) in, is the parameter; Step S34: Use loss function Perform parameter training and back propagation of the model to obtain the parameters of the updated model. The pre-training process ends and the pre-trained model parameters are retained. .

4. The method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling according to claim 3 is characterized in that: The step S4: trajectory generation stage, using the pre-training stage model, adopts a method of gradually increasing the generation of trajectory points, and each step generates a certain number of trajectory points incrementally until the obtained trajectory points form a complete final trajectory, specifically including: Step S41: The pre-trained model parameters The conditional probability model defined is denoted as , the generation process starts from At the beginning, all the data initially input into the pre-trained model are masked and used as the input of the first step through the pre-trained model to obtain the first generated trajectory sequence. A certain proportion of trajectory points at the beginning of the sampling sequence are used as the trajectory points generated in the first step. , and then As a condition, The subsequent trajectory point mask is used as the input of the second step through the pre-trained model to obtain the second generated trajectory sequence. The trajectory point data with a larger proportion at the beginning of the sampling sequence is used as the trajectory point generated in the second step. sampling , and so on, the sampling ratio of each step is calculated in step S42; Step S42: Sampling ratio per round As shown in formula (21): (21) in, Indicates the current sampling round number, represents the total number of sampling rounds; relatively The variable is monotonically increasing, The value range of is between 0 and 1, and is used to indicate the sampling ratio of each step in step S41; Step S43: According to the sampling ratio , according to the sampling method of step S41, from Continuous sampling , until the sum of the sampling ratios of the trajectory is 1, the number of generated trajectory points is equal to the number of original trajectory points, forming a complete trajectory sequence, and all the trajectories obtained in the last step are taken out as a complete final trajectory.

5. The method for generating city-level large-scale crowd activity trajectories based on mobile phone signaling according to claim 4 is characterized in that: The step S5 is as follows: given the corresponding longitude and latitude coordinates of a specific person at a certain moment, and inputting them into the initial trajectory sequence position corresponding to the moment, on the basis of step S4, after each step of generating trajectory points, the corresponding longitude and latitude coordinates of the specific person at a certain moment are covered in the generated trajectory points, and the trajectory points are generated repeatedly in multiple steps according to step S4, until a complete final trajectory of the specific person is finally formed, until the final trajectories of a predetermined number of specific persons are generated, and the set of the final trajectories is the city-level large-scale crowd activity trajectory, which specifically includes: Step S51: Set the residence and work place of specific personnel according to the corresponding tasks, generate large-scale crowd activity trajectories, set the spatial position of part of the evening time period in the input sequence of step S4 to the residence in advance, set the spatial position of part of the daytime time period to the residence, and set other positions as masks for subsequent trajectory generation; Step S52: Based on step S4, After the wheel trajectory is generated, to The track of the time period is retained, and the tracks of the evening and daytime preset in step S51 are overlaid on the retained track. The overlaid track is used to continue to generate the i+1th round using step S4 until the sum of the sampling ratios of the multi-step tracks is 1. After the final complete track is obtained, the step of generating the track is terminated; Step S53: Repeat steps S51 to S52 until the final trajectories of a predetermined number of specific persons are obtained.

Images