Pedestrian trajectory prediction method and device combined with mechanism model, equipment and medium

Abstract

The invention provides a pedestrian trajectory prediction method, device and equipment combined with a mechanism model and a medium, and relates to the technical field of trajectory prediction, and the method comprises the steps: obtaining the prediction trajectory change of a target pedestrian based on a pre-obtained mechanism model according to the obtained to-be-measured observation information of the target pedestrian; based on the to-be-measured observation information, track change errors are obtained through the error prediction model; and correcting the predicted trajectory change by using the trajectory change error to obtain a target trajectory change, and integrating the target trajectory change with the to-be-measured observation information to obtain a target pedestrian trajectory. According to the invention, the modeling error of the mechanism model is compensated by the trajectory change error predicted by the error prediction model, and the accuracy of pedestrian trajectory prediction is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of trajectory prediction, and in particular to a pedestrian trajectory prediction method, device, equipment and medium combined with a mechanism model. Background Art

[0002] With the development of artificial intelligence, intelligent systems are gradually integrated into human activity scenarios. When predicting human activity behaviors, how to ensure the security of human-intelligent system interaction, support intelligent system service functions, and improve the working efficiency of intelligent systems have become important issues.

[0003] The challenges faced by research related to pedestrian trajectory prediction mainly come from the diversity of human behavior, which is affected by many factors, including individual differences, social factors, environmental changes, etc. At present, pedestrian trajectories are usually predicted by rule-based or mechanism-based methods. This type of method is based on a priori dynamics or social mechanics models and has excellent interpretability, but its final solution still depends on the mechanism model equation specified in advance. Therefore, this method still fails to solve the modeling error problem between real data and performs poorly in terms of robustness and flexibility, resulting in low accuracy in pedestrian trajectory prediction. Summary of the invention

[0004] The problem solved by the present invention is how to improve the accuracy of pedestrian trajectory prediction.

[0005] In order to solve the above problems, the present invention provides a pedestrian trajectory prediction method, device, equipment and medium combined with a mechanism model.

[0006] In a first aspect, the present invention provides a pedestrian trajectory prediction method combined with a mechanism model, comprising:

[0007] According to the acquired observation information of the target pedestrian to be measured, a predicted trajectory change of the target pedestrian is obtained based on a pre-acquired mechanism model;

[0008] Based on the observation information to be measured, the trajectory change error is obtained by using the error prediction model, wherein the error prediction network obtains the trajectory change to be corrected according to the mechanism model and the historical observation trajectory sequence, obtains the corrected error true value according to the trajectory error true value in the historical observation trajectory sequence and the trajectory change to be corrected, and the preset prediction model is trained according to the historical observation trajectory sequence and the corrected error true value;

[0009] The predicted trajectory change is corrected using the trajectory change error to obtain a target trajectory change, and the target trajectory change is integrated with the observation information to be measured to obtain a target pedestrian trajectory.

[0010] Optionally, before obtaining the trajectory change error by using the error prediction model based on the to-be-tested observation information, the method further includes:

[0011] Based on the obtained historical observation trajectory sequence, obtaining the to-be-corrected trajectory change based on the mechanism model;

[0012] Obtaining the historical trajectory change according to the historical observation trajectory sequence, and obtaining the true value of the correction error according to the historical trajectory change and the to-be-corrected trajectory change;

[0013] Constructing a prediction model, and training the prediction model according to the historical observation trajectory sequence and the true value of the correction error to obtain the error prediction model.

[0014] Optionally, the constructing a prediction model, training the prediction model according to the historical observation trajectory sequence and the true value of the correction error to obtain the error prediction model includes:

[0015] Based on a variational autoencoder, constructing the prediction model, and partitioning and partially hiding the historical observation trajectory sequence to obtain a test trajectory sequence;

[0016] According to the historical observation trajectory sequence and the true value of the correction error, performing learning by using the prediction model to obtain a first prediction model, and obtaining the posterior distribution of the latent variable according to the historical observation trajectory sequence;

[0017] Inputting the test trajectory sequence into the first prediction model to obtain a first error, and obtaining the approximate posterior distribution of the latent variable according to the test trajectory sequence;

[0018] According to the test trajectory sequence, obtaining a first trajectory change by using the mechanism model;

[0019] Correcting the first trajectory change according to the first error to obtain a second trajectory change, and obtaining a predicted trajectory according to the second trajectory change and the test trajectory sequence;

[0020] Obtaining the mean square error between the predicted trajectory and the historical observation trajectory sequence to obtain a reconstruction loss, obtaining the KL divergence between the posterior distribution of the latent variable and the approximate posterior distribution of the latent variable, and optimizing the first prediction model based on the reconstruction loss and the KL divergence to obtain the error prediction model.

[0021] Optionally, the prediction model includes an encoder, a decoder, a forward recurrent network, and a backward recurrent network; the obtaining a first prediction model by performing learning by using the prediction model according to the historical observation trajectory sequence and the true value of the correction error includes:

[0022] Mapping the observation trajectory at the start time point in the historical observation trajectory sequence using a multi-layer perceptron to obtain the initial trajectory forward state;

[0023] According to the observation trajectory at the time point to be measured in the historical observation trajectory sequence, using the backward recurrent network to obtain the trajectory backward state at the current time point, where the time point to be measured is two time steps away from the start time point, and the current time point is between the start time point and the time point to be measured;

[0024] According to the initial trajectory forward state and the trajectory backward state at the current time point, using the encoder to encode to obtain the posterior distribution of the latent variable, and using the decoder in the prediction model to decode the posterior distribution of the latent variable to obtain the error to be corrected at the current time point;

[0025] Randomly sampling according to the posterior distribution of the latent variable to obtain an intermediate latent variable, mapping the intermediate latent variable and the error to be corrected, and using the forward recurrent network to obtain the trajectory forward state at the current time point according to the initial trajectory forward state and the mapping result;

[0026] Taking the current time point as the new start time point and the trajectory forward state at the current time point as the new initial trajectory forward state, returning to the step of obtaining the trajectory backward state at the current time point using the backward recurrent network according to the observation trajectory at the time point to be measured in the historical observation trajectory sequence, and cyclically optimizing the encoder and the decoder based on the historical observation trajectory sequence and the true value of the correction error until the observation data at each time point and the error to be corrected at each time point in the historical observation trajectory sequence are all optimized, to obtain the optimized encoder and the optimized decoder;

[0027] Obtaining the first prediction model according to the optimized encoder, the optimized decoder, the backward recurrent network and the forward recurrent network.

[0028] Optionally, the obtaining the trajectory change error using the error prediction model based on the to-be-measured observation information includes:

[0029] Performing multi-dimensional feature extraction on the to-be-measured observation information to obtain pedestrian features and static scene features;

[0030] Concatenating the pedestrian features and the static scene features to obtain a comprehensive observation feature;

[0031] Using the error prediction model to obtain the trajectory change error according to the comprehensive observation feature.

[0032] Optionally, the pedestrian features include the features of the pedestrian himself and the features of neighboring pedestrians; the multi-dimensional feature extraction of the to-be-tested observation information to obtain pedestrian features and static scene features includes:

[0033] Extract the pedestrian speed and pedestrian displacement of the target pedestrian in the to-be-tested observation information to obtain the features of the pedestrian himself;

[0034] Obtain neighboring pedestrians within a preset range according to the to-be-tested observation information at each time point, and obtain the relative displacement and relative speed between the neighboring pedestrians and the target pedestrian to obtain the features of neighboring pedestrians;

[0035] Adopt a graph attention mechanism to generate a directed edge pointing from the neighboring pedestrian to the target pedestrian. Based on the directed edge, according to the Euclidean distance, the cosine value of the motion direction angle, and the minimum prediction distance between the features of the pedestrian himself and the features of the neighboring pedestrians, obtain the social influence features of the neighboring pedestrians on the target pedestrian, and obtain the pedestrian self-state features of the target pedestrian before each time point;

[0036] According to the features of the pedestrian himself of the target pedestrian before each time point and the social influence features of each neighboring pedestrian on the target pedestrian, obtain the edge weight of the directed edge;

[0037] Obtain the neighbor influence weight according to the edge weight, and obtain the pedestrian features according to the features of the pedestrian himself, the neighbor influence weight, and the features of the neighboring pedestrians.

[0038] Optionally, the multi-dimensional feature extraction of the to-be-tested observation information to obtain pedestrian features and static scene features includes:

[0039] Perform a dot product operation on the manually annotated scene mask image and the to-be-tested observation information, and input the processing result into a Resnet152 model for feature extraction to obtain static scene features.

[0040] In a second aspect, the present invention provides a pedestrian trajectory prediction device combined with a mechanism model, including:

[0041] A first prediction module, configured to obtain the predicted trajectory change of the target pedestrian based on the mechanism model obtained in advance according to the to-be-tested observation information of the target pedestrian obtained;

[0042] A second prediction module, configured to obtain a trajectory change error by using an error prediction model based on the to-be-measured observation information, where the error prediction network obtains a to-be-corrected trajectory change according to the mechanism model and a historical observation trajectory sequence, obtains a corrected error truth value according to the trajectory error truth value in the historical observation trajectory sequence and the to-be-corrected trajectory change, and trains a preset prediction model according to the historical observation trajectory sequence and the corrected error truth value;

[0043] An integration module, configured to correct the predicted trajectory change by using the trajectory change error to obtain a target trajectory change, and integrate the target trajectory change with the to-be-measured observation information to obtain a target pedestrian trajectory.

[0044] In a third aspect, the present invention provides an electronic device, including a memory and a processor;

[0045] The memory is configured to store a computer program;

[0046] The processor is configured to implement the pedestrian trajectory prediction method combining a mechanism model as described in the first aspect when executing the computer program.

[0047] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the pedestrian trajectory prediction method combining a mechanism model as described in the first aspect is implemented.

[0048] The beneficial effects of the pedestrian trajectory prediction method combining a mechanism model of the present invention are as follows: according to the to-be-measured observation information of the target pedestrian obtained, a predicted trajectory change of the target pedestrian is obtained based on a pre-obtained mechanism model, providing basic prediction data for pedestrian trajectory prediction. Based on the to-be-measured observation information, a trajectory change error is obtained by using an error prediction model as a modeling correction error of the mechanism model. The predicted trajectory change is corrected by using the trajectory change error to obtain a target trajectory change, and the target trajectory change is integrated with the to-be-measured observation information to obtain a target pedestrian trajectory. The trajectory change error predicted by the error prediction model compensates for the modeling error of the mechanism model, effectively improving the accuracy of pedestrian trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic flowchart of the pedestrian trajectory prediction method combining a mechanism model according to an embodiment of the present invention; Figure 1 ;

[0050] Figure 2 is a schematic structural diagram of a trajectory target estimation module according to an embodiment of the present invention;

[0051] Figure 3 is a schematic diagram of a pedestrian's field of view according to an embodiment of the present invention;

[0052] Figure 4 Schematic diagram of the process of the prediction model learning stage according to an embodiment of the present invention;

[0053] Figure 5 Schematic diagram of the process of the prediction model testing stage according to an embodiment of the present invention;

[0054] Figure 6 Schematic diagram of the structure of a pedestrian trajectory prediction device with a combined mechanism model according to an embodiment of the present invention;

[0055] Figure 7 Schematic diagram of the structure of an electronic device according to an embodiment of the present invention;

[0056] Figure 8 Schematic flow of the pedestrian trajectory prediction method with a combined mechanism model according to an embodiment of the present invention Figure 2 . Detailed implementation manners

[0057] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments described herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not used to limit the protection scope of the present invention.

[0058] It should be understood that the various steps described in the method embodiments of the present invention can be executed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this regard.

[0059] The term "including" and its variants used herein are open-ended, that is, "including but not limited to"; the term "based on" is "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts such as "first" and "second" mentioned in the present invention are only used to distinguish different devices, modules, or units, and are not used to limit the order of functions performed by these devices, modules, or units or their interdependent relationships.

[0060] It should be noted that the modifications of "one" and "plurality" mentioned in the present invention are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0061] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes, and are not used to limit the scope of these messages or information.

[0062] In view of the problems existing in the above-mentioned related technologies, this embodiment provides a pedestrian trajectory prediction method, device, equipment and medium combined with a mechanism model.

[0063] like Figures 1 to 5 ,as well as Figure 8 As shown, an embodiment of the present invention provides a pedestrian trajectory prediction method combined with a mechanism model, including:

[0064] Step S1: According to the acquired observation information of the target pedestrian to be measured, the predicted trajectory change of the target pedestrian is obtained based on the pre-acquired mechanism model.

[0065] Specifically, the movement trajectory of pedestrians is affected by a variety of forces from the environment, among which the most important influencing factors are the attraction of moving targets and the repulsion of static obstacles, which are obtained by the social mechanics model. The social mechanics model includes a trajectory target estimation module and a static obstacle segmentation module, which are used to obtain the attraction of moving targets and the repulsion of static obstacles based on the observed information of the target pedestrian to be measured.

[0066] (1) Trajectory Target Estimation Module

[0067] In the trajectory prediction task, since the pedestrian moving target is invisible in the pedestrian trajectory prediction process, it is necessary to build a model to estimate and sample the trajectory target, and then use the generated target to calculate the target attraction. Figure 2 The trajectory target estimation module shown makes predictions of trajectory targets.

[0068] The observation information to be measured includes pedestrian activity scene information and pedestrian observation trajectory information. The pedestrian activity scene information is image information. The RGB scene image I of the pedestrian activity scene information is intercepted, and the target pedestrian observation trajectory information at time point t is recorded as The scene image I is processed using a semantic segmentation network to obtain a semantic map S, where the semantic map S includes walkable areas and non-walkable areas. The trajectory heat map H(t, i, j) is obtained by using formula (1), which is a Gaussian probability distribution map. Its height and width are consistent with the scene image I, and each time point corresponds to a processing channel.

[0069]

[0070] Among them, H(t, i, j) represents the trajectory heat map, t represents the observation segment time point, and its value ranges from 1 to t obs , i and j respectively represent the two-dimensional pixel subscripts of the scene picture I, and p t represents the position coordinates of the target pedestrian at time point t, and (x, y) represents the coordinates of any pixel point on the scene picture I.

[0071] The trajectory heat map H(t, i, j) and the semantic map S are concatenated in the channel dimension to form the scene-trajectory heat map Hs. The U-net encoder is used to encode the scene-trajectory heat map Hs to obtain the intermediate result H m , which is input into the U-net target heat map decoder for decoding to generate the spatial probability distribution map of the final position. After super-sampling and clustering, the final trajectory target p is output T .

[0072] After obtaining the trajectory target p T , calculate the target attractiveness of the target pedestrian. According to the current position p of the pedestrian t and the trajectory target p T , the expected moving direction e of the target pedestrian at time point t is obtained according to formula (2) t , that is, the unit direction vector of the position coordinate p of the target pedestrian at time point t pointing to the position coordinate of the trajectory target p t . The expected speed is calculated according to formula (3) T represents the average speed required to reach the trajectory target within the remaining time. The expected velocity is The target attractiveness F represents the tendency of the target pedestrian to change the current speed goal , that is, the acceleration required for the current speed of the target pedestrian to become the expected speed , which is calculated by formula (4).

[0073]

[0074] Among them, T represents the total number of time points from the starting point to the trajectory target p of the target pedestrian T , and represents the time step.

[0075]

[0076] (2) Static obstacle segmentation module

[0077] ​Based on the assumption that only static obstacles within the field of view of the target pedestrian will affect the trajectory of the target pedestrian, the field of view of the target pedestrian is defined as a square area with the diagonal along the current velocity direction and side length r env . As shown in Figure 3 . Figure (a) shows a schematic diagram of the field of view of the target pedestrian marked based on the scene picture I, and Figure (b) shows a schematic diagram of the field of view of the pedestrian under the corresponding manually annotated mask picture. Among them, area a is the static obstacle within the field of view of the target pedestrian, which will generate an environmental repulsive force on the pedestrian.

[0078] With the help of the homography matrix, the pedestrian trajectory coordinates are converted from the world coordinate system to the pixel coordinate system based on the image. According to the current velocity direction of the target pedestrian, the field of view range is delimited, and the side length r env of the field of view is set to the empirical value of 50. Based on the manually annotated mask picture, the center coordinates p obs of the obstacles within the field of view of the target pedestrian are calculated, and the static obstacle repulsive force F env is calculated according to formula (5).

[0079]

[0080] Among them, k env represents the environmental force coefficient.

[0081] The static obstacle repulsive force F env and the target attraction force F goal are spliced to obtain the resultant force F, and based on the mechanism model, that is, formulas (6) and (7), the predicted trajectory changes of the target pedestrian are calculated, including the displacement changes △x and △v.

[0082] △x = 1 / 2(F / m)t 运 2 + vt 运 (6)

[0083] △v = (F / m)t 运 (7)

[0084] Among them, m represents the mass of the target pedestrian itself, and t 运 represents the movement time of the target pedestrian from the current time point to the next time point.

[0085] Step S2: Used to obtain the trajectory change error based on the to-be-measured observation information by using an error prediction model. Among them, the error prediction network obtains the trajectory change to be corrected according to the mechanism model and the historical observation trajectory sequence, obtains the corrected error truth value according to the trajectory error truth value in the historical observation trajectory sequence and the trajectory change to be corrected, and trains a preset prediction model according to the historical observation trajectory sequence and the corrected error truth value.

[0086] It should be noted that before the error prediction model predicts the trajectory change error, the prediction model should be constructed and trained first.

[0087] That is, before obtaining the trajectory change error using the error prediction model based on the to-be-measured observation information, it further includes:

[0088] Based on the obtained historical observation trajectory sequence, the to-be-corrected trajectory change is obtained based on the mechanism model.

[0089] Specifically, first, a preset number of historical observation trajectory sequences are obtained, usually the pedestrian observation trajectory features at 20 time points, such as the time features at 20 time points, pedestrian own features, static obstacle features, neighbor pedestrian features, etc., and a mechanism model is introduced to perform prediction according to the historical observation trajectory sequence, and the to-be-corrected trajectory change of the pedestrian at each time point is obtained as the dataset for subsequent prediction model training.

[0090] The historical trajectory change is obtained according to the historical observation trajectory sequence, and the true value of the correction error is obtained according to the historical trajectory change and the to-be-corrected trajectory change.

[0091] Specifically, the historical trajectory change of the pedestrian at each time point is extracted based on the historical observation trajectory sequence, and the difference between the historical trajectory change of the pedestrian at each time point and the to-be-corrected trajectory change of the pedestrian at each time point is calculated to obtain the true value of the correction error at each time point as the dataset for subsequent prediction model training.

[0092] A prediction model is constructed, and the prediction model is trained according to the historical observation trajectory sequence and the true value of the correction error to obtain the error prediction model.

[0093] Specifically, the prediction model can be constructed according to a neural network model, such as a Convolutional Neural Networks (CNN). The prediction model is trained according to the historical observation trajectory sequence and the true value of the correction error to obtain the error prediction model, so as to improve the accuracy and prediction rate of trajectory change error prediction, and further increase the accuracy of pedestrian trajectory prediction.

[0094] In a preferred embodiment of the present invention, the constructing the prediction model, training the prediction model according to the historical observation trajectory sequence and the true value of the correction error to obtain the error prediction model:

[0095] Based on a variational autoencoder, the prediction model is constructed, and the historical observation trajectory sequence is divided and partially hidden to obtain a test trajectory sequence.

[0096] Specifically, the Variational Autoencoder (VAE) introduces a probability model on the basis of the autoencoder and is trained by maximizing the log-likelihood of the data. The goal of the variational autoencoder is not only to learn the compressed representation of the data, but also to ensure that these representations have good probability distribution characteristics in the latent space, and the latent space has a certain interpretability, which helps to better understand and analyze the training data, thereby improving the robustness of the prediction model to generate higher-quality prediction data. Using the variational autoencoder to construct a prediction model can effectively improve the accuracy of the trajectory change error.

[0097] Divide the historical observation trajectory sequence and partially hide the data therein to obtain a test sequence. For example, the historical observation trajectory sequence includes historical observation trajectories at 20 time points. Divide the historical observation trajectory sequence, use the historical observation trajectories at the 1st - 8th time points as the observation trajectory sequence, use the historical observation trajectories at the 9th - 20th time points as the prediction data sequence, and hide the prediction data sequence to obtain the observation trajectory sequence as the test trajectory sequence, that is, the test set for the prediction model training. The observation trajectory sequence is used to support the prediction of the trajectory coordinates of its prediction data sequence.

[0098] The initial target probability distribution of the variational autoencoder is shown in Equation (8):

[0099]

[0100] where i represents the target pedestrian, t H represents the number of time points in the prediction data sequence, t F represents the number of time points in the observation trajectory sequence, 1:t H represents the time period of the observation trajectory sequence in the test trajectory sequence, that is, the observation time period, t H+1 :t H +t F represents the time period of the prediction data sequence in the test trajectory sequence, that is, the prediction time period, represents the comprehensive observation features extracted from the observation trajectory sequence of the target pedestrian i, d represents the coordinate residuals at each time point in the prediction time period, represents the initial target probability distribution of the variational autoencoder, represents the trajectory change error in the prediction time period of the target pedestrian i, d t i represents the trajectory change error of the target pedestrian i at time point t, represents the trajectory change error of the target pedestrian i obtained before the current time point t.

[0101] Introduce the latent variable z, and the target probability distribution function will evolve into Equation (9):

[0102]

[0103] where z t represents the latent variable corresponding to the time point t, represents the posterior probability distribution of the trajectory change error with respect to the sampling result of the latent variable and the comprehensive observation features of the target pedestrian i, represents the posterior probability distribution of the latent variable with respect to the comprehensive observation features of the target pedestrian i.

[0104] The training of the trajectory prediction model mainly includes a learning stage and a testing stage. Learning stage:

[0105] Based on the historical observation trajectory sequence and the true value of the correction error, the prediction model is used for learning to obtain a first prediction model, and the posterior distribution of the latent variable is obtained according to the historical observation trajectory sequence.

[0106] Specifically, the prediction model learns based on the historical observation trajectory sequence and the true value of the correction error to obtain a well-trained first prediction model. At this time, the first prediction model has high accuracy. Therefore, the posterior distribution of the latent variable is obtained according to the historical observation trajectory sequence, that is, the true posterior distribution of the latent variable, as the basic data for subsequent model tuning.

[0107] Specifically, as Figure 4 shown, the prediction model includes an encoder, a decoder, a forward recurrent network, and a backward recurrent network. In this embodiment, on the basis of the prediction model constructed by the original VAE model, a backward recurrent network GRU b and a forward recurrent network GRU h are added. During the learning process of the prediction model, the historical observation trajectory sequences are all observation segments for the learning process of the prediction model. The step of using the prediction model to learn based on the historical observation trajectory sequence and the true value of the correction error to obtain a first prediction model includes:

[0108] Using a multi-layer perceptron to map the observation trajectory at the start time point in the historical observation trajectory sequence to obtain an initial trajectory forward state. For example Figure 4 in, if the start time point is t - 1, then the initial trajectory forward state is h t-1 .

[0109] According to the observation trajectory at the time point to be measured in the historical observation trajectory sequence, the backward recurrent network is used to obtain the trajectory backward state at the current time point, where the time point to be measured is two time steps away from the start time point, and the current time point is between the start time point and the time point to be measured. As Figure 4Among them, the time point to be measured is t+1, and the observation trajectory of the time point to be measured t+1 in the historical observation trajectory sequence includes the comprehensive observation feature O of the current time point t t and / or the backward state b of the time point to be measured t+1 (If the time point to be measured is the first time point for calculating the backward state, the observation trajectory of the time point to be measured can be mapped by using a multi-layer perceptron to obtain the trajectory backward state of the current time point), and input it into the backward recurrent neural network GRU b to obtain the trajectory backward state b of the current time point t t .

[0110] According to the initial trajectory forward state and the trajectory backward state of the current time point, use the encoder to encode to obtain the posterior distribution of the latent variable, and use the decoder in the prediction model to decode the posterior distribution of the latent variable to obtain the error to be corrected at the current time point

[0111] Specifically, as Figure 4 shown, according to the initial trajectory forward state h t-1 and the trajectory backward state b of the current time point t to obtain the posterior distribution of the latent variable, and use the decoder in the prediction model to decode the posterior distribution of the latent variable to obtain the error to be corrected d at the current time point t .

[0112] Randomly sample according to the posterior distribution of the latent variable to obtain an intermediate latent variable, use, for example, a multi-layer perceptron φ zd to map the intermediate latent variable and the error to be corrected, and use the forward recurrent neural network to obtain the trajectory forward state h of the current time point according to the initial trajectory forward state h t-1 and the encoding result t ;

[0113] Take the current time point as the new start time point and the trajectory forward state of the current time point as the new initial trajectory forward state, return to the step of obtaining the trajectory backward state of the current time point by using the backward recurrent neural network according to the observation trajectory of the time point to be measured in the historical observation trajectory sequence, and perform cyclic optimization on the encoder and the decoder based on the historical observation trajectory sequence and the true value of the correction error until the observation data of each time point and the error to be corrected of each time point in the historical observation trajectory sequence are all optimized, obtain the optimized encoder and the optimized decoder, and obtain the first prediction model according to the optimized encoder, the optimized decoder, the backward recurrent neural network and the forward recurrent neural network

[0114] Finally, the learned first prediction model is represented by formulas (11) and (12):

[0115]

[0116] Among them, q in formula (11) φ (z t |b t ,h t-1 ) represents the posterior distribution of the latent variable z, that is, the latent variable z is related to the observed information and historical prediction information. In this formula, is the prediction result of the trajectory change error before time t, is the comprehensive observation feature of the observation segment of pedestrian i. In formula (12) represents the trajectory change error d of pedestrian i at time t i t of the posterior distribution.

[0117] In this embodiment, a forward recurrent network and a backward recurrent network are used to learn the prediction model, which can effectively learn the context information of the historical observation trajectory sequence, thereby improving the learning accuracy of the prediction model.

[0118] Testing stage:

[0119] Input the test trajectory sequence into the first prediction model to obtain a first error, and obtain an approximate posterior distribution of the latent variable according to the test trajectory sequence.

[0120] Specifically, using the test trajectory sequence as a test set, the accuracy of the first prediction model is tested. Input the test trajectory sequence into the first prediction model to predict the trajectory change error of the prediction segment, obtain a first error, and obtain an approximate posterior distribution of the latent variable according to the test trajectory sequence.

[0121] According to the test trajectory sequence, use the mechanism model to obtain a first trajectory change, and the first trajectory change is the pedestrian trajectory change predicted by the mechanism model according to the prediction segment in the test trajectory sequence.

[0122] As Figure 5 shown, after learning the prediction model, a first prediction model is obtained. To verify the accuracy of the first prediction model, it is tested according to the prediction trajectory sequence, and the first prediction model is optimized according to the test results to obtain a final trajectory prediction model. The inputting the test trajectory sequence into the first prediction model to obtain a first error and obtaining an approximate posterior distribution of the latent variable according to the test trajectory sequence includes:

[0123] The observed trajectory at the first time point in the test trajectory sequence is mapped by the multi-layer perceptron φ h to the initial forward state As Figure 5 shown by the dashed part in; then the forward recurrent network GRU hUpdate the forward state h at each step according to the input of the intermediate result latent variable z and the trajectory change error d at the corresponding time point, as Figure 5 shown, after obtaining the forward state h t-1 , according to the forward state h tt-1 obtain the approximate posterior distribution p of the latent variable θ (z t |h t-1 ), represented by formula (13), and decode the approximate posterior distribution p of the latent variable θ (z t |h t-1 ) to obtain the trajectory change error d at the current time point t t .

[0124]

[0125] Optimize the first prediction model according to the results of the learning stage and the results of the test stage:

[0126] As Figure 5 shown in the lower part, correct the first trajectory change △x t according to the first error d t’ , that is, add the two to get the second trajectory change △x t , and then obtain the predicted trajectory according to the second trajectory change and the test trajectory sequence. For example, given that the pedestrian trajectory at the 8th time point is x and the second trajectory change at the 9th time point is △x t , then the pedestrian predicted trajectory at the 9th time point is x + △x t .

[0127] Obtain the mean square error between the predicted trajectory and the historical observation trajectory sequence to obtain the reconstruction loss, obtain the KL divergence between the posterior distribution of the latent variable and the approximate posterior distribution of the latent variable, and optimize the first prediction model based on the reconstruction loss and the KL divergence to obtain the error prediction model.

[0128] Specifically, use the commonly used loss function of the VAE model, the Evidence Lower Bound (ELBO), to perform backpropagation on the prediction model to optimize the prediction model and obtain the final error prediction model. The formula for the Evidence Lower Bound of the loss function is shown in formula (10):

[0129]

[0130] where D KL [q φ (z t |b t ,h t-1 )||p θ(z t |h t-1 )] represents the KL divergence between the latent variable posterior distribution and the latent variable approximate posterior distribution, h t-1 Represents the forward state at time point t-1, which can be based on GRU h Network acquisition, p θ (z t |h t-1 ) represents the approximate posterior distribution of the latent variable z at time point t, q φ (z t |b t ,h t-1 ) represents the posterior distribution of the latent variable z at time point t, z t ~q φ (·|b t ,h t-1 ) represents the time point t from the distribution q φ (z t |b t ,h t-1 ) is sampled to obtain z t , p ζ (d i t |z t ,h t-1 ) represents the trajectory change error d of pedestrian i at time t i t In practical applications, in order to avoid cumulative errors and simplify calculations, the mean square error between the corrected predicted trajectory and the true value of the target trajectory is used to approximate logp ζ (d i t |z t ,h t-1 )item.

[0131] Step S3: using the trajectory change error to correct the predicted trajectory change to obtain a target trajectory change, and integrating the target trajectory change with the observation information to be measured to obtain a target pedestrian trajectory.

[0132] Specifically, the estimated change error is summed with the predicted trajectory change to obtain the target trajectory change. At this time, the target trajectory change is a more realistic trajectory change of the target pedestrian at the next time point. The observed information to be measured is then integrated according to the target trajectory change, that is, the target pedestrian trajectory at the next time point is obtained based on the observation information at the current time point and the target trajectory change.

[0133] In a preferred embodiment of the present invention, obtaining the trajectory change error by using an error prediction model based on the observation information to be measured includes:

[0134] Perform multi-dimensional feature extraction on the to-be-tested observation information to obtain pedestrian features and static scene features;

[0135] Concatenate the pedestrian features and the static scene features to obtain comprehensive observation features;

[0136] Use the error prediction model to obtain the trajectory change error according to the comprehensive observation features.

[0137] Specifically, the pedestrian trajectory is a complex result jointly affected by various internal and external factors. The factors used for predictive model analysis in the field of pedestrian trajectory prediction mainly include three categories: (1) Pedestrian features, including the pedestrian's own features and the features of neighboring pedestrians. The pedestrian's own features are obtained from the data in the historical trajectory observation sequence and usually include the pedestrian's own position coordinates, speed, acceleration, and potential moving targets, etc.; The features of neighboring pedestrians contain position conflicts or social impacts brought by surrounding moving objects (such as neighboring pedestrians), usually the trajectory features of neighboring pedestrians; (2) Static scene features, including static obstacles, geometric topology information, and semantic information in the scene, etc. Extract multi-dimensional features from the dataset, and aggregate various influencing factors to obtain comprehensive observation features, which are used as the basis for predicting the target pedestrian's trajectory using the error prediction model (the pedestrian's own features, the features of neighboring pedestrians, and static scene features), and can provide rich information in multiple dimensions, thereby realizing accurate prediction of the pedestrian trajectory.

[0138] The pedestrian features can be directly obtained by extracting information from the to-be-tested observation information (usually image information), and the static scene features can be extracted using a feature extraction model. Concatenate the pedestrian features and the static scene features to obtain comprehensive observation features It is represented by formula (17):

[0139]

[0140] Among them, represents the pedestrian features, and S i represents the static scene features.

[0141] In a preferred embodiment of the present invention, the performing multi-dimensional feature extraction on the to-be-tested observation information to obtain pedestrian features and static scene features includes:

[0142] Extract the pedestrian speed and pedestrian displacement of the target pedestrian in the to-be-tested observation information to obtain the pedestrian's own features.

[0143] Specifically, for target pedestrian i, obtain its two-dimensional spatial coordinates at time point t from the to-be-tested observation information (usually image information) where R is the set of real numbers, and here R 2Indicates that the x - coordinate point is a two - dimensional real - valued vector, and its observed segment trajectory is The predicted segment trajectory is denoted as Since pedestrian trajectory datasets usually come from video sampling and have a constant sampling frequency (taking the ETH / UCY dataset as an example, the time interval between each point is 0.4 s), the pedestrian speed at each time point can be calculated accordingly. The pedestrian's own characteristics of the target pedestrian are denoted as

[0144] According to the to - be - measured observation information at each time point, neighboring pedestrians within a preset range are obtained, and the relative displacement and relative speed between the neighboring pedestrians and the target pedestrian are obtained to obtain the neighboring pedestrian characteristics.

[0145] Specifically, for target pedestrian i, all neighboring pedestrians within a preset range r i are obtained based on the to - be - measured observation information at each time point to construct a neighboring set J, that is According to historical experience, r i is set to the empirical value 2. For neighbor j ∈ J, the relative displacement and relative speed between the neighboring pedestrian and the target pedestrian are obtained to obtain the neighboring pedestrian characteristics, and the neighboring pedestrian characteristics are denoted as

[0146] The graph attention mechanism is used to generate a directed edge from the neighboring pedestrian to the target pedestrian. Based on the directed edge, according to the Euclidean distance, the cosine value of the motion direction angle, and the minimum prediction distance between the pedestrian's own characteristics and the neighboring pedestrian characteristics, the social influence characteristics of the neighboring pedestrian on the target pedestrian are obtained, and the pedestrian's own state characteristics of the target pedestrian before each time point are obtained.

[0147] Specifically, neighbor pedestrian j and target pedestrian i are regarded as nodes in the graph, and the attention mechanism is used to generate a directed edge from neighbor pedestrian j to target pedestrian i. According to the Euclidean distance, the cosine value of the motion direction angle, and the minimum prediction distance between the pedestrian's own characteristics and the neighboring pedestrian characteristics, the social influence characteristics of the neighboring pedestrian on the target pedestrian are obtained. The pedestrian's own state characteristics of the target pedestrian before each time point can be obtained according to the image information corresponding to the time point.

[0148] According to the pedestrian's own characteristics of the target pedestrian before each time point and the social influence characteristics of each neighboring pedestrian on the target pedestrian, the edge weight of the directed edge is obtained. It is represented by formula (14):

[0149]

[0150] where LeakyReLU() is the activation function, fq is a feature extraction network with learnable parameters, is the pedestrian's own state feature of the target row i before time t - 1, f k is a feature extraction network with learnable parameters, represents the social influence feature generated by neighbor pedestrian j on target pedestrian i at time point t.

[0151] Obtain the neighbor influence weight according to the edge weight It is represented by formula (15), and the pedestrian feature H is obtained according to the pedestrian's own feature, the neighbor influence weight and the neighbor pedestrian feature i t , which is represented by formula (16):

[0152]

[0153] where exp() represents the natural exponential function, represents the edge weight of neighbor pedestrian k pointing to target pedestrian i, f s is a feature extraction network with learnable parameters, f n is a feature extraction network with learnable parameters.

[0154] In a preferred embodiment of the present invention, the pedestrian feature includes a static scene feature; the multi - dimensional feature extraction of the to - be - measured observation information to obtain the pedestrian feature and the static scene feature includes:

[0155] Perform a dot - product operation on the manually labeled scene mask image and the to - be - measured observation information, and input the processing result into the Resnet152 model for feature extraction to obtain the static scene feature S i .

[0156] As Figure 6 shown, a pedestrian trajectory prediction device 600 provided by an embodiment of the present invention includes:

[0157] A first prediction module 610, configured to obtain the predicted trajectory change of the target pedestrian based on the to - be - measured observation information of the target pedestrian and a pre - obtained mechanism model;

[0158] A second prediction module 620, where the error prediction network obtains the trajectory change to be corrected according to the mechanism model and the historical observation trajectory sequence, obtains the corrected error truth value according to the trajectory error truth value in the historical observation trajectory sequence and the trajectory change to be corrected, and trains a preset prediction model according to the historical observation trajectory sequence and the corrected error truth value;

[0159] An integration module 630 is configured to correct the predicted trajectory change by using the trajectory change error to obtain a target trajectory change, and integrate the target trajectory change with the to-be-detected observation information to obtain a target pedestrian trajectory.

[0160] As Figure 7 shown, an electronic device 700 provided by an embodiment of the present invention includes a memory 710 and a processor 720; the memory 710 is configured to store a computer program; the processor 720 is configured to, when executing the computer program, implement the pedestrian trajectory prediction method in combination with the mechanism model as described above.

[0161] Or rather, an electronic device 700 includes a memory 710 and a processor 720 coupled to the memory 710; the memory 710 is configured to store a computer program; the processor 720 is configured to, when executing the computer program, perform the following operations:

[0162] Based on the to-be-detected observation information of the target pedestrian obtained, obtain the predicted trajectory change of the target pedestrian based on a pre-obtained mechanism model;

[0163] Based on the to-be-detected observation information, use an error prediction model to obtain a trajectory change error, wherein, based on the mechanism model and the historical observation trajectory sequence, obtain a to-be-corrected trajectory change, obtain a corrected error truth value according to the trajectory error truth value in the historical observation trajectory sequence and the to-be-corrected trajectory change, and train a preset prediction model according to the historical observation trajectory sequence and the corrected error truth value to obtain the error prediction network;

[0164] Correct the predicted trajectory change by using the trajectory change error to obtain a target trajectory change, and integrate the target trajectory change with the to-be-detected observation information to obtain a target pedestrian trajectory.

[0165] A computer-readable storage medium provided by an embodiment of the present invention has a computer program stored thereon, and when the computer program is executed by a processor, the pedestrian trajectory prediction method in combination with the mechanism model as described above is implemented.

[0166] Or rather, a non-volatile computer-readable storage medium has a computer program stored thereon, and when the computer program is executed by a processor, the processor is caused to perform the following operations:

[0167] Based on the to-be-detected observation information of the target pedestrian obtained, obtain the predicted trajectory change of the target pedestrian based on a pre-obtained mechanism model;

[0168] Based on the to-be-measured observation information, a trajectory change error is obtained by using an error prediction model. Among them, a to-be-corrected trajectory change is obtained according to the mechanism model and the historical observation trajectory sequence, a corrected error truth value is obtained according to the trajectory error truth value in the historical observation trajectory sequence and the to-be-corrected trajectory change, and a preset prediction model is trained according to the historical observation trajectory sequence and the corrected error truth value to obtain the error prediction network;

[0169] The predicted trajectory change is corrected by using the trajectory change error to obtain a target trajectory change, and the target trajectory change and the to-be-measured observation information are integrated to obtain a target pedestrian trajectory.

[0170] Now, an electronic device 700 that can be used as a server or a client of the present invention will be described. It is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device 700 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device 700 can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0171] The electronic device 700 includes a computing unit that can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The computing unit, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0172] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiments of the present invention. In addition, the functional units in each embodiment of the present invention can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0173] Although the present invention is disclosed as above, the scope of protection of the present invention is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will all fall within the scope of protection of the present invention.

Claims

1. A pedestrian trajectory prediction method combined with a mechanism model, characterized in that Including: Based on the to-be-tested observation information of the obtained target pedestrian, obtain the predicted trajectory change of the target pedestrian based on the pre-obtained mechanism model; Based on the to-be-tested observation information, use the error prediction model to obtain the trajectory change error, wherein the error prediction network obtains the to-be-corrected trajectory change according to the mechanism model and the historical observation trajectory sequence, obtains the corrected error truth value according to the trajectory error truth value in the historical observation trajectory sequence and the to-be-corrected trajectory change, and trains the preset prediction model according to the historical observation trajectory sequence and the corrected error truth value; Use the trajectory change error to correct the predicted trajectory change to obtain the target trajectory change, and integrate the target trajectory change with the to-be-tested observation information to obtain the target pedestrian trajectory.

2. The pedestrian trajectory prediction method of the binding mechanism model according to claim 1, characterized in that Before the step of using the error prediction model to obtain the trajectory change error based on the to-be-tested observation information, it further includes: Based on the obtained historical observation trajectory sequence, obtain the to-be-corrected trajectory change based on the mechanism model; Obtain the historical trajectory change according to the historical observation trajectory sequence, and obtain the corrected error truth value according to the historical trajectory change and the to-be-corrected trajectory change; Construct a prediction model, and train the prediction model according to the historical observation trajectory sequence and the corrected error truth value to obtain the error prediction model.

3. The pedestrian trajectory prediction method of the binding mechanism model according to claim 2, wherein The step of constructing a prediction model and training the prediction model according to the historical observation trajectory sequence and the corrected error truth value to obtain the error prediction model includes: Based on the variational autoencoder, construct the prediction model, and divide and partially hide the historical observation trajectory sequence to obtain a test trajectory sequence; According to the historical observation trajectory sequence and the corrected error truth value, use the prediction model for learning to obtain a first prediction model, and obtain the posterior distribution of the latent variable according to the historical observation trajectory sequence; Input the test trajectory sequence into the first prediction model to obtain a first error, and obtain the approximate posterior distribution of the latent variable according to the test trajectory sequence; According to the test trajectory sequence, obtain the first trajectory change by using the mechanism model; Correct the first trajectory change according to the first error to obtain a second trajectory change, and obtain the predicted trajectory according to the second trajectory change and the test trajectory sequence; Obtain the mean square error between the predicted trajectory and the historical observation trajectory sequence to obtain the reconstruction loss, obtain the KL divergence between the posterior distribution of the latent variable and the approximate posterior distribution of the latent variable, and optimize the first prediction model based on the reconstruction loss and the KL divergence to obtain the error prediction model.

4. The pedestrian trajectory prediction method of the binding mechanism model according to claim 3, characterized in that The prediction model includes an encoder, a decoder, a forward recurrent network, and a backward recurrent network; the step of using the prediction model to learn according to the historical observation trajectory sequence and the corrected error truth value to obtain a first prediction model includes: Use a multi-layer perceptron to map the observation trajectory at the start time point in the historical observation trajectory sequence to obtain the initial trajectory forward state; According to the observation trajectory at the time point to be measured in the historical observation trajectory sequence, use the backward recurrent network to obtain the backward state of the trajectory at the current time point, where the time point to be measured is two time steps away from the start time point, and the current time point is between the start time point and the time point to be measured; According to the initial forward state of the trajectory and the backward state of the trajectory at the current time point, use the encoder to encode to obtain the posterior distribution of the latent variable, and use the decoder in the prediction model to decode the posterior distribution of the latent variable to obtain the error to be corrected at the current time point; Randomly sample according to the posterior distribution of the latent variable to obtain an intermediate latent variable, map the intermediate latent variable and the error to be corrected, and use the forward recurrent network to obtain the forward state of the trajectory at the current time point according to the initial forward state of the trajectory and the mapping result; Take the current time point as the new start time point and the forward state of the trajectory at the current time point as the new initial forward state of the trajectory, return to the step of using the backward recurrent network to obtain the backward state of the trajectory at the current time point according to the observation trajectory at the time point to be measured in the historical observation trajectory sequence, and perform cyclic optimization on the encoder and the decoder based on the historical observation trajectory sequence and the true value of the correction error until the observation data at each time point in the historical observation trajectory sequence and the error to be corrected at each time point are all optimized, and obtain the optimized encoder and the optimized decoder; Obtain the first prediction model according to the optimized encoder, the optimized decoder, the backward recurrent network, and the forward recurrent network; 5. The pedestrian trajectory prediction method according to the binding mechanism model described in claim 1, characterized in that The obtaining of the trajectory change error by using the error prediction model based on the to-be-measured observation information includes: Perform multi-dimensional feature extraction on the to-be-measured observation information to obtain pedestrian features and static scene features; Concatenate the pedestrian features and the static scene features to obtain comprehensive observation features; Use the error prediction model to obtain the trajectory change error according to the comprehensive observation features; 6. The pedestrian trajectory prediction method according to the binding mechanism model described in claim 5, characterized in that, The pedestrian features include the features of the pedestrian itself and the features of neighboring pedestrians; the performing of multi-dimensional feature extraction on the to-be-measured observation information to obtain pedestrian features and static scene features includes: Extract the pedestrian speed and pedestrian displacement of the target pedestrian in the to-be-measured observation information to obtain the features of the pedestrian itself; Obtain neighboring pedestrians within a preset range according to the to-be-measured observation information at each time point, and obtain the relative displacement and relative speed between the neighboring pedestrians and the target pedestrian to obtain the features of neighboring pedestrians; Adopt a graph attention mechanism to generate a directed edge pointing from the neighboring pedestrian to the target pedestrian, and based on the directed edge, obtain the social influence features of the neighboring pedestrian on the target pedestrian according to the Euclidean distance, the cosine value of the movement direction angle, and the minimum prediction distance between the features of the pedestrian itself and the features of the neighboring pedestrian, and obtain the features of the state of the pedestrian itself of the target pedestrian before each of the time points; Based on the pedestrian's own characteristics of the target pedestrian before each time point and the social influence characteristics of each of the neighbor pedestrians on the target pedestrian, obtain the edge weight of the directed edge; Based on the edge weight, obtain the neighbor influence weight, and based on the pedestrian's own characteristics, the neighbor influence weight, and the neighbor pedestrian characteristics, obtain the pedestrian characteristics.

7. The pedestrian trajectory prediction method of the binding mechanism model according to claim 5, characterized in that, The multi-dimensional feature extraction of the to-be-tested observation information obtains pedestrian characteristics and static scene characteristics, including: Perform a dot product process on the manually annotated scene mask image and the to-be-tested observation information, and input the processing result into the Resnet152 model for feature extraction to obtain static scene characteristics.

8. A pedestrian trajectory prediction device combined with a mechanism model, characterized in that, Including: A first prediction module, configured to, based on the to-be-tested observation information of the target pedestrian obtained, obtain the predicted trajectory change of the target pedestrian based on a pre-obtained mechanism model; A second prediction module, configured to, based on the to-be-tested observation information, use an error prediction model to obtain a trajectory change error, where the error prediction network obtains a to-be-corrected trajectory change according to the mechanism model and a historical observation trajectory sequence, obtains a corrected error truth value according to the trajectory error truth value in the historical observation trajectory sequence and the to-be-corrected trajectory change, and trains a preset prediction model according to the historical observation trajectory sequence and the corrected error truth value; An integration module, configured to use the trajectory change error to correct the predicted trajectory change to obtain a target trajectory change, and integrate the target trajectory change with the to-be-tested observation information to obtain the target pedestrian trajectory.

9. An electronic device, characterized in that, Including a memory and a processor; The memory is used to store a computer program; The processor is configured to, when executing the computer program, implement the pedestrian trajectory prediction method combining a mechanism model according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by the processor, the pedestrian trajectory prediction method combining a mechanism model according to any one of claims 1 to 7 is implemented.

Images