A pedestrian trajectory prediction method based on interactive perception diffusion model

Abstract

The embodiment of the application discloses a kind of pedestrian trajectory prediction methods based on interactive perception diffusion model, this method includes: constructing forward diffusion module, obtains the noisy future trajectory of target pedestrian;Social and scene interaction features are regarded as the condition of containing parameter reverse Markov chain, and the noisy future trajectory of target pedestrian is denoised by constructing reverse denoising module;Loss calculation module is constructed, and the pedestrian trajectory prediction model composed of the above module is optimized using loss value, to obtain the optimal pedestrian trajectory prediction model;Test denoising module is constructed, and the initial random noise of test is denoised based on the optimal pedestrian trajectory prediction model, to obtain pedestrian trajectory prediction result.The social encoder and scene encoder are designed to extract social interaction features and scene interaction features respectively in the application, to retain scene spatial information, and the double-guided decoder is designed to strengthen the cooperative guidance of the two interaction features, which further improves the accuracy and rationality of pedestrian trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, specifically relating to a method for predicting pedestrian trajectories based on an interactive perception diffusion model. Background Technology

[0002] In recent years, pedestrian trajectory prediction has been widely applied in fields such as autonomous driving, robotics, and surveillance systems, thus attracting widespread attention from academia and industry. It primarily studies methods for predicting the future trajectory of a target pedestrian based on their historical trajectory. However, the multimodal nature of pedestrian motion trajectories in real-world scenarios, and the complexity of interactions between the target pedestrian and surrounding pedestrians and the environment, pose significant challenges to pedestrian trajectory prediction technology.

[0003] Recently, interaction-aware pedestrian trajectory prediction methods have made significant progress. To improve the accuracy of pedestrian trajectory prediction, some researchers have designed different methods to model social and scene interactions. To model social interactions, Alahi et al. introduced a pooling mechanism to share hidden states among pedestrians; Sadeghian et al. used an attention mechanism to dynamically weight surrounding pedestrians; Mohamed et al. treated pedestrian positions as nodes in a graph and interactions as edges to explicitly represent social interactions. To model scene interactions, Dendorfer et al. first extracted features from RGB scene images and then interacted with the historical trajectory features of pedestrians; Mangalam et al. treated the positions in the historical pedestrian trajectories as 2D Gaussian probability distributions to generate pedestrian motion heatmaps, and then interacted with scene semantic segmentation maps to preserve the spatial information of the scene. Although existing methods have studied social and scene interactions from different perspectives, the above interaction-aware pedestrian trajectory prediction methods do not preserve scene spatial information when modeling both types of interactions simultaneously.

[0004] Recently, due to the reliable performance of diffusion models in generation tasks, Gu et al. first applied diffusion models to pedestrian trajectory prediction methods. This method uses a single encoder to extract social interaction features from the historical trajectories of pedestrians, and then uses these social interaction features as conditions to apply variational inference to generate trajectories. Based on the above architecture, Mao et al. accelerated the sampling process by skipping the initial denoising step during trajectory generation; Bae et al. projected pedestrian motion trajectories into a singular space before applying diffusion models for trajectory prediction. However, these diffusion model-based pedestrian trajectory prediction methods only focus on social interaction and neglect the role of scene interaction, thus weakening the accuracy and reliability of trajectory prediction. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem that pedestrian trajectory prediction methods based on diffusion models ignore scene interaction and cannot retain scene spatial information. To this end, this invention provides a pedestrian trajectory prediction method based on an interaction-aware diffusion model.

[0006] To achieve the aforementioned objective, the present invention proposes a pedestrian trajectory prediction method based on an interactive perception diffusion model, comprising the following steps:

[0007] Step S1: Construct a forward diffusion module. During the training phase, the forward diffusion module is used to add forward noise to the unnoised future trajectory of the target pedestrian. In the process, the noisy future trajectory of the target pedestrian is obtained. ,in, This represents the total number of noise additions during each training session.

[0008] Step S2: Construct a reverse denoising module to extract the social interaction features and scene interaction features of the target pedestrian. Calculate the noise prediction value required for denoising the parametric inverse Markov chain using these features. The reverse denoising module treats the social interaction features and scene interaction features as conditions for the parametric inverse Markov chain. Based on the noise prediction value, it adds noise to the target pedestrian's future trajectory. Denoising is performed using a parametric inverse Markov chain to obtain the future trajectory of the target pedestrian after denoising;

[0009] Step S3: Construct a loss calculation module, input the noise prediction value into the loss calculation module, and use the obtained loss value to optimize the pedestrian trajectory prediction model composed of the forward diffusion module, the reverse denoising module and the loss calculation module to obtain the optimal pedestrian trajectory prediction model;

[0010] Step S4: Construct a test denoising module. Based on the optimal pedestrian trajectory prediction model, the test denoising module denoises the initial random noise of the test. After denoising, the pedestrian trajectory prediction results are obtained.

[0011] Optionally, in step S1, the noise addition processing of the forward diffusion module is expressed as follows:

[0012] ;

[0013] in, Represents the noisy future trajectory of a given target pedestrian. Under the conditions, ( The joint conditional probability distribution of ) Indicates the first The noisy future trajectory of the target pedestrian after the second noise addition, symbol The representation is defined as follows: Indicates that in a given Under the conditions, Conditional probability:

[0014] ;

[0015] in, express Obtained from the mean The variance is The normal distribution Represents a predefined variance sequence. Represents the identity matrix.

[0016] Optionally, in step S2, the social interaction features of the target pedestrian are extracted from the historical trajectories of the target pedestrian and surrounding pedestrians using a social encoder; the scene interaction features of the target pedestrian are extracted from the scene semantic segmentation map and the motion heat map of the target pedestrian using a scene encoder.

[0017] Optionally, step S2 includes the following steps:

[0018] Step S21: Construct a social encoder and initialize its parameters. Input the historical trajectories of the target pedestrian and surrounding pedestrians into the social encoder to extract the social interaction features of the target pedestrian.

[0019] Step S22: Input the RGB scene image of the scene where the target pedestrian is located into a pre-trained semantic segmentation neural network with fixed parameters to obtain the scene semantic segmentation map of the scene where the target pedestrian is located.

[0020] Step S23: Treat each position in the historical trajectory of the target pedestrian as the mean of a two-dimensional Gaussian probability distribution, obtain the trajectory scene representation corresponding to that position with a fixed variance, and superimpose the trajectory scene representations corresponding to all positions in the historical trajectory of the target pedestrian to obtain the motion heat map of the target pedestrian.

[0021] Step S24: Determine the pre-trained deep learning model and initialize its parameters. Remove the last global pooling layer and fully connected layer in the pre-trained deep learning model to obtain the scene encoder. Overlay the scene semantic segmentation map and the motion heat map of the target pedestrian and input them into the scene encoder to extract the scene interaction features of the target pedestrian.

[0022] Step S25: Construct a dual-guided decoder to integrate the social interaction features, scene interaction features, and the noisy future trajectory of the target pedestrian. The input is fed into the dual-guided decoder to obtain the noise prediction value required for the parametric inverse Markov chain denoising;

[0023] Step S26: Construct an inverse denoising module. This module treats the social interaction features and scene interaction features as conditions of a parametric inverse Markov chain and uses the noise prediction values ​​to noisyize the future trajectory of the target pedestrian. Perform parametric inverse Markov chain denoising to obtain the denoised future trajectory of the target pedestrian. .

[0024] Optionally, in step S21, the social interaction characteristics of the target pedestrian... Social characteristics and time series characteristics Multilayer perceptron used after splicing The mapping yields:

[0025] ,

[0026] ,

[0027] ,

[0028] Among them, social characteristics Based on feature dimension The multi-head attention mechanism was obtained. and These represent the historical trajectories of the target pedestrian and the surrounding pedestrians, respectively. These represent query, key, and value embedding functions, respectively, and time-series features. Based on temporal convolutional neural network and gated loop unit get.

[0029] Optionally, in step S25, the dual-guided decoder has Each layer of the dual-guided decoder contains a sequentially connected motion guidance time module, a first temporal coding model, a scene guidance space module, and a second temporal coding model.

[0030] Optionally, in step S25, the first The input to the motion guidance time module in the layer dual-guided decoder is the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder Social interaction characteristics and the total number of noise additions The output is :

[0031] ,

[0032] ,

[0033] in, , , These are implemented based on fully connected layers and have preset parameters. , , The parameterized embedded function Indicates feature splicing, Indicates the time embedding function:

[0034] ,

[0035] in, Represents a predefined variance sequence, matrix Sampling is performed using a standard Gaussian distribution;

[0036] After the first time-series coding model and... Perform residual connection to obtain , Scene interaction features Total number of noise additions Together, they serve as input to the scene guidance space module, and the output of the scene guidance space module is... ;

[0037] After the second time-series coding model and After performing residual join, we obtain the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder The final layer dual-guided decoder outputs the noisy future trajectory features of the target pedestrian. The noise prediction value required for denoising the parametric inverse Markov chain.

[0038] Optionally, in step S26, the parametric inverse Markov chain denoising is represented as:

[0039] ,

[0040] in, Indicates a given social interaction feature and scene interaction features Under the conditions, ( The joint conditional probability distribution of ) Indicates the first The denoised future trajectory of the target pedestrian after the first denoising step. Indicates the first After secondary noise addition, the target pedestrian's future trajectory is further noised. The probability distribution, Indicates that in a given , , Under the conditions, The conditional probability distribution.

[0041] Optionally, the loss function in step S3 is expressed as:

[0042] ,

[0043] in, Indicates forward noise and noise prediction values Between Norm expectation This represents the noisy future trajectory of the target pedestrian. This indicates the total number of times noise was added. This represents the social interaction characteristics of the target pedestrian. This represents the scene interaction features of the target pedestrian. They all follow a certain probability distribution.

[0044] Optionally, in step S4, the denoising operation of the test denoising module is represented as follows:

[0045] ,

[0046] in, and They represent the first Second and third Predicted future trajectory of the target pedestrian after denoising. Represents a predefined variance sequence. , , This represents the predicted noise value. This represents the social interaction characteristics of the target pedestrian. This represents the scene interaction features of the target pedestrian. It is a random noise variable.

[0047] The beneficial effects of this invention are as follows: This invention treats social and scene interaction features as conditions of a parametric inverse Markov chain, performs denoising operations on the noisy future trajectory of the target pedestrian, designs a social encoder and a scene encoder to model social and scene interactions respectively to preserve scene spatial information, and uses a dual-guided decoder to enhance the collaborative guidance of the two interactions, generating a more reasonable trajectory, avoiding the trajectory from crossing obstacles or road boundaries, and improving the accuracy and reliability of pedestrian trajectory prediction. Attached Figure Description

[0048] Figure 1 This is a flowchart of a pedestrian trajectory prediction method based on an interactive perception diffusion model according to an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0050] Figure 1 This is a flowchart of a pedestrian trajectory prediction method based on an interactive sensing diffusion model according to an embodiment of the present invention. The following is an example... Figure 1 To illustrate some specific implementation processes of the present invention, we will use examples, such as... Figure 1 As shown, the pedestrian trajectory prediction method based on the interactive perception diffusion model includes the following steps:

[0051] Step S1: Construct a forward diffusion module. During the training phase, the forward diffusion module is used to add forward noise to the unnoised future trajectory of the target pedestrian. In the process, the noisy future trajectory of the target pedestrian is obtained. ,in, The total number of noise additions during each training session. From the predetermined range In one embodiment of the present invention, the selection is randomized. ;

[0052] During the training phase, the unnoised future trajectory of the target pedestrian is... It is known.

[0053] The noise addition process of the forward diffusion module can be expressed as follows:

[0054] ,

[0055] in, Represents the noisy future trajectory of a given target pedestrian. Under the condition that, ( The joint conditional probability distribution of ) Indicates the first The noisy future trajectory of the target pedestrian after the second noise addition, symbol The representation is defined as follows: Indicates that in a given Under the conditions, The conditional probability of can be expressed as:

[0056] ,

[0057] in, express Obeying the mean The variance is The normal distribution This represents a predefined series of variances, whose values ​​increase sequentially. Represents the identity matrix.

[0058] Based on the above formula and the properties of Gaussian transition, for any The conditional probability distribution of the noisy future trajectory of the target pedestrian can be expressed as:

[0059] ,

[0060] in, This represents the unnoised future trajectory of the target pedestrian. , .

[0061] Then the first After the noise is added, the conditional probability distribution of the noisy future trajectory of the target pedestrian can be expressed as:

[0062] ,

[0063] Furthermore, From The sampling yielded:

[0064] ,

[0065] in, In one embodiment of the present invention, the forward noise is described as follows: When the total number of noise additions When large enough, .

[0066] In one embodiment of the present invention, , and linearly increase to .

[0067] Step S2: Construct a reverse denoising module to extract the social interaction features and scene interaction features of the target pedestrian. Calculate the noise prediction value required for denoising the parametric inverse Markov chain using these features. The reverse denoising module treats the social interaction features and scene interaction features as conditions for the parametric inverse Markov chain. Based on the noise prediction value, it adds noise to the target pedestrian's future trajectory. Denoising is performed using a parametric inverse Markov chain to obtain the future trajectory of the target pedestrian after denoising;

[0068] In one embodiment of the present invention, the social interaction features of the target pedestrian are extracted from the historical trajectories of the target pedestrian and surrounding pedestrians using a social encoder, wherein the surrounding pedestrians refer to other pedestrians in the same scene as the target pedestrian; the scene interaction features of the target pedestrian are extracted from the scene semantic segmentation map and the motion heat map of the target pedestrian using a scene encoder to preserve scene spatial information.

[0069] Further, step S2 includes the following steps:

[0070] Step S21: Construct a social encoder and initialize its parameters. Input the historical trajectories of the target pedestrian and surrounding pedestrians into the social encoder to extract the social interaction features of the target pedestrian.

[0071] In one embodiment of the present invention, the social interaction feature Social characteristics and time series characteristics Multilayer perceptron used after splicing The mapping yields the following equation:

[0072] ,

[0073] ,

[0074] ,

[0075] Among them, social characteristics Based on feature dimension The multi-head attention mechanism was obtained. and These represent the historical trajectories of the target pedestrian and the surrounding pedestrians, respectively. These represent query, key, and value embedding functions, respectively, and time-series features. Based on temporal convolutional neural network and gated loop unit get.

[0076] In one embodiment of the present invention, the feature dimension The embedding function All are implemented using fully connected layers with an input dimension of 2 and an output dimension of 256; the temporal convolutional neural network The input dimension is 2, the output dimension is 256, and the kernel size is 3; the gated recurrent unit The input dimension is 256, and the hidden layer dimension is 256; the multilayer perceptron The input dimension is 512 and the output dimension is 256.

[0077] Step S22: Input the RGB scene image of the scene where the target pedestrian is located into a pre-trained semantic segmentation neural network with fixed parameters to obtain the scene semantic segmentation map of the scene where the target pedestrian is located.

[0078] In one embodiment of the present invention, the pre-trained semantic segmentation neural network with fixed parameters can be U-Net, and the input RGB scene image size is [missing information]. ,in, 3 represents the image dimensions, 3 represents the image channels, and the output scene semantic segmentation map size is [size missing]. , where 6 represents the number of categories in the scene semantic segmentation graph.

[0079] Step S23: Treat each position in the historical trajectory of the target pedestrian as the mean of a two-dimensional Gaussian probability distribution, obtain the trajectory scene representation corresponding to that position with a fixed variance, and superimpose the trajectory scene representations corresponding to all positions in the historical trajectory of the target pedestrian to obtain the motion heat map of the target pedestrian.

[0080] The trajectory scene is represented as a two-dimensional Gaussian probability distribution.

[0081] In one embodiment of the present invention, the variance is 1.25, and the size of the motion heatmap of the target pedestrian is [missing value]. 8 represents the number of trajectory points in the historical trajectory of the target pedestrian.

[0082] Step S24: Determine the pre-trained deep learning model and initialize its parameters. Remove the last global pooling layer and fully connected layer in the pre-trained deep learning model to obtain the scene encoder. Overlay the scene semantic segmentation map and the motion heat map of the target pedestrian and input them into the scene encoder to extract the scene interaction features of the target pedestrian, thereby preserving the scene spatial information.

[0083] In one embodiment of the present invention, the pre-trained deep learning model is ResNet-18, and only the model components before the last global pooling layer are selected, that is, the last global pooling layer and the fully connected layer in the pre-trained deep learning model are removed. The output dimension of the scene interaction features is 256.

[0084] Step S25: Construct a dual-guided decoder to integrate the social interaction features, scene interaction features, and the noisy future trajectory of the target pedestrian. The input is fed into the dual-guided decoder to obtain the noise prediction value required for the parametric inverse Markov chain denoising;

[0085] In one embodiment of the present invention, the dual-guided decoder has Each layer of the dual-guided decoder contains a sequentially connected motion guidance time module, a first temporal coding model, a scene guidance space module, and a second temporal coding model.

[0086] Among them, the layer( The input to the motion guidance time module in the dual-guided decoder is the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder Social interaction characteristics and the total number of noise additions The output is :

[0087] ,

[0088] ,

[0089] in, , , These are implemented based on fully connected layers and have preset parameters. , , The parameterized embedded function Indicates feature splicing, Indicates the time embedding function:

[0090] ,

[0091] Among them, matrix Sampling is performed on a standard Gaussian distribution.

[0092] Subsequently After the first time-series coding model and... Perform residual connection to obtain , Scene interaction features Total number of noise additions Together, they serve as input to the scene guidance space module. The structure of the scene guidance space module is similar to that of the motion guidance time module, and the specific processing procedure will not be described in detail. The output of the scene guidance space module is... .

[0093] at last After the second time-series coding model and After performing residual join, we obtain the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder The last layer, that is, the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder That is, the noise prediction value required for denoising the parametric inverse Markov chain.

[0094] In one embodiment of the present invention, ; It is implemented using a fully connected layer with an input dimension of 256 and an output dimension of 256. , It is implemented by a fully connected layer with an input dimension of 512 and an output dimension of 256; the first and second temporal coding models can be transformer encoders with 1 layer and an input and output dimension of 256.

[0095] Step S26: Construct an inverse denoising module. This module treats the social interaction features and scene interaction features as conditions of a parametric inverse Markov chain and uses the noise prediction values ​​to noisyize the future trajectory of the target pedestrian. Perform parametric inverse Markov chain denoising to obtain the denoised future trajectory of the target pedestrian. .

[0096] The parametric inverse Markov chain denoising can be expressed as:

[0097] ,

[0098] in, Indicates a given social interaction feature and scene interaction features Under the conditions, ( The joint conditional probability distribution of ) Indicates the first The denoised future trajectory of the target pedestrian after the first denoising step. Indicates the first After secondary noise addition, the target pedestrian's future trajectory is further noised. The probability distribution, Indicates that in a given , , Under the conditions, The conditional probability distribution.

[0099] through After noise reduction From The sampling yielded:

[0100] ,

[0101] in, It is a random noise variable. In one embodiment of the present invention, , and They represent The mean and variance of.

[0102] In one embodiment of the present invention, the variance Set as The mean It can be reparameterized and represented as:

[0103] ,

[0104] in, This represents the predicted noise value.

[0105] Step S3: Construct a loss calculation module, input the noise prediction value into the loss calculation module, and use the obtained loss value to optimize the pedestrian trajectory prediction model composed of the forward diffusion module, the reverse denoising module and the loss calculation module to obtain the optimal pedestrian trajectory prediction model;

[0106] In step S3, the loss function used by the loss calculation module is obtained by maximizing the log-likelihood of the future trajectory of the target pedestrian using a variational lower bound, and can be expressed as:

[0107] ,

[0108] in, Indicates the desired operation. Indicates After the first denoising step, the future denoising trajectory of the target pedestrian. The probability distribution, Indicates the desired operation, where It follows a certain probability distribution. This indicates that in the inverse denoising module, given... , , Under the conditions, The conditional probability distribution, This indicates that in the forward diffusion module, given... Under the conditions, The conditional probability distribution.

[0109] Based on step S26 The mean of the reparameterization involved The loss function can be simplified to minimizing the noise prediction error, expressed as:

[0110] ,

[0111] in, Indicates forward noise and noise prediction values Between Norm expectation, where, All of them follow a certain probability distribution. In one embodiment of the present invention, , , ,when hour, .

[0112] Step S4: Construct a test denoising module. Based on the optimal pedestrian trajectory prediction model, the test denoising module denoises the initial random noise of the test. After denoising, the pedestrian trajectory prediction results are obtained.

[0113] Let the initial random noise of the test be denoted as . During the testing phase, the test denoising module can be used to denoise the initial random noise based on the social interaction features and scene interaction features. China Economic After denoising, the pedestrian trajectory prediction results are obtained. The predicted future trajectory of the target pedestrian obtained during the denoising process can be represented as: ,in, Indicates the first Predicted future trajectory of the target pedestrian after denoising.

[0114] The specific denoising operation of the test denoising module can be expressed as follows:

[0115] ,

[0116] in, and They represent the first Second and third Predicted future trajectory of the target pedestrian after denoising. Represents a predefined variance sequence. , , This represents the predicted noise value. It is a random noise variable. In one embodiment of the present invention, .

[0117] Using publicly available online pedestrian trajectory prediction databases as test subjects, such as the ETH / UCY database, the average displacement error (ADE) and endpoint displacement error (FDE) of the pedestrian trajectory prediction of this invention reached 0.16 and 0.26, respectively.

[0118] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A pedestrian trajectory prediction method based on an interactive perception diffusion model, characterized in that, The method includes the following steps: Step S1: Construct a forward diffusion module. During the training phase, the forward diffusion module is used to add forward noise to the unnoised future trajectory of the target pedestrian. In the process, the noisy future trajectory of the target pedestrian is obtained. ,in, This represents the total number of noise additions during each training session. Step S2: Construct a reverse denoising module to extract the social interaction features and scene interaction features of the target pedestrian. Calculate the noise prediction value required for denoising the parametric inverse Markov chain using these features. The reverse denoising module treats the social interaction features and scene interaction features as conditions for the parametric inverse Markov chain. Based on the noise prediction value, it adds noise to the target pedestrian's future trajectory. Denoising is performed using a parametric inverse Markov chain to obtain the future trajectory of the target pedestrian after denoising; Step S3: Construct a loss calculation module, input the noise prediction value into the loss calculation module, and use the obtained loss value to optimize the pedestrian trajectory prediction model composed of the forward diffusion module, the reverse denoising module and the loss calculation module to obtain the optimal pedestrian trajectory prediction model; Step S4: Construct a test denoising module. Based on the optimal pedestrian trajectory prediction model, the test denoising module denoises the initial random noise of the test. After denoising, the pedestrian trajectory prediction results are obtained; Step S2 includes the following steps: Step S21: Construct a social encoder and initialize its parameters. Input the historical trajectories of the target pedestrian and surrounding pedestrians into the social encoder to extract the social interaction features of the target pedestrian. Step S22: Input the RGB scene image of the scene where the target pedestrian is located into a pre-trained semantic segmentation neural network with fixed parameters to obtain the scene semantic segmentation map of the scene where the target pedestrian is located. Step S23: Treat each position in the historical trajectory of the target pedestrian as the mean of a two-dimensional Gaussian probability distribution, obtain the trajectory scene representation corresponding to that position with a fixed variance, and superimpose the trajectory scene representations corresponding to all positions in the historical trajectory of the target pedestrian to obtain the motion heat map of the target pedestrian. Step S24: Determine the pre-trained deep learning model and initialize its parameters. Remove the last global pooling layer and fully connected layer in the pre-trained deep learning model to obtain the scene encoder. Overlay the scene semantic segmentation map and the motion heat map of the target pedestrian and input them into the scene encoder to extract the scene interaction features of the target pedestrian. Step S25: Construct a dual-guided decoder to integrate the social interaction features, scene interaction features, and the noisy future trajectory of the target pedestrian. The input is fed into the dual-guided decoder to obtain the noise prediction value required for the parametric inverse Markov chain denoising, wherein the dual-guided decoder has Each layer of dual-guided decoders contains sequentially connected motion guidance time modules, a first temporal coding model, a scene guidance space module, and a second temporal coding model. Step S26: Construct an inverse denoising module. This module treats the social interaction features and scene interaction features as conditions of a parametric inverse Markov chain and uses the noise prediction values ​​to noisyize the future trajectory of the target pedestrian. Perform parametric inverse Markov chain denoising to obtain the denoised future trajectory of the target pedestrian. .

2. The method according to claim 1, characterized in that, In step S1, the noise addition process of the forward diffusion module is represented as follows: ， in, Represents the noisy future trajectory of a given target pedestrian. Under the conditions, ( The joint conditional probability distribution of ) Indicates the first The noisy future trajectory of the target pedestrian after the second noise addition, symbol The representation is defined as follows: Indicates that in a given Under the conditions, Conditional probability: ， in, express Obtained from the mean The variance is The normal distribution Represents a predefined variance sequence. Represents the identity matrix.

3. The method according to claim 1, characterized in that, In step S2, the social interaction features of the target pedestrian are extracted from the historical trajectories of the target pedestrian and surrounding pedestrians using a social encoder; the scene interaction features of the target pedestrian are extracted from the scene semantic segmentation map and the motion heat map of the target pedestrian using a scene encoder.

4. The method according to claim 1, characterized in that, In step S21, the social interaction characteristics of the target pedestrian Social characteristics and time series characteristics Multilayer perceptron used after splicing The mapping yields: ， ， ， Among them, social characteristics Based on feature dimension The multi-head attention mechanism was obtained. and These represent the historical trajectories of the target pedestrian and the surrounding pedestrians, respectively. These represent query, key, and value embedding functions, respectively, and time-series features. Based on temporal convolutional neural network and gated loop unit get.

5. The method according to claim 1, characterized in that, In step S25, the first The input to the motion guidance time module in the layer dual-guided decoder is the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder Social interaction characteristics and the total number of noise additions The output is : ， ， in, , , These are implemented based on fully connected layers and have preset parameters. , , The parameterized embedded function Indicates feature splicing, Indicates the time embedding function: ， in, Represents a predefined variance sequence, matrix Sampling is performed using a standard Gaussian distribution; After the first time-series coding model and... Perform residual connection to obtain , Scene interaction features Total number of noise additions Together, they serve as input to the scene guidance space module, and the output of the scene guidance space module is... ; After the second time-series coding model and After performing residual join, we obtain the first... The noisy future trajectory features of the target pedestrian output by the layer dual-guided decoder The final layer dual-guided decoder outputs the noisy future trajectory features of the target pedestrian. The noise prediction value required for denoising the parametric inverse Markov chain.

6. The method according to claim 1, characterized in that, In step S26, the parametric inverse Markov chain denoising is represented as follows: ， in, Indicates a given social interaction feature and scene interaction features Under the conditions, ( The joint conditional probability distribution of ) Indicates the first The denoised future trajectory of the target pedestrian after the first denoising step. Indicates the first After secondary noise addition, the target pedestrian's future trajectory is further noised. The probability distribution, Indicates that in a given , , Under the conditions, The conditional probability distribution.

7. The method according to claim 1, characterized in that, The loss function in step S3 is expressed as follows: ， in, Indicates forward noise and noise prediction values Between Norm expectation This represents the noisy future trajectory of the target pedestrian. This indicates the total number of times noise was added. This represents the social interaction characteristics of the target pedestrian. This represents the scene interaction features of the target pedestrian. They all follow a certain probability distribution.

8. The method according to claim 1, characterized in that, In step S4, the denoising operation of the test denoising module is represented as follows: ， in, and They represent the first Second and third Predicted future trajectory of the target pedestrian after denoising. Represents a predefined variance sequence. , , This represents the predicted noise value. This represents the social interaction characteristics of the target pedestrian. This represents the scene interaction features of the target pedestrian. It is a random noise variable.

Images