Pedestrian trajectory prediction method and device based on directional perception condition denoising

Abstract

This application proposes a pedestrian trajectory prediction method and device based on direction-aware conditional denoising. By dividing the space around the target pedestrian into several directional sectors, the influence of neighbors in different directions is independently modeled, enabling the model to more accurately reflect the changes in the neighborhood structure in complex scenes. Furthermore, by randomly sampling latent variables, multiple potential future motion patterns are generated, which can simultaneously describe different future motion directions and improve the ability to express the diversity of pedestrian behavior.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and device for predicting pedestrian trajectories based on orientation-aware conditional denoising. Background Technology

[0002] With the development of intelligent transportation systems and the widespread application of autonomous driving technology, pedestrian trajectory prediction, as a key technology in predicting the behavior of traffic participants, plays an important role in path planning, collision warning, and behavioral decision-making. Existing pedestrian trajectory prediction methods typically estimate the future position of a target pedestrian based on its historical trajectory. Traditional methods mainly use statistical models or predict based on the historical motion state of a single target, but they do not fully consider the interaction between the target pedestrian and its neighboring pedestrians, resulting in low prediction accuracy in complex scenarios.

[0003] To improve prediction performance, existing technologies have proposed deep learning-based trajectory prediction methods that model the interaction relationships of multiple pedestrians using Long Short-Term Memory (LSTM) networks and attention mechanisms. However, these methods typically use a holistic neighborhood aggregation approach to process neighbor information, making it difficult to distinguish the differences in the influence of neighbors from different directions on the target pedestrian. In real-world scenarios, neighbors in front of the target pedestrian usually exert a strong constraint on the direction of movement, while neighbors to the side and rear have a relatively weaker influence, resulting in significant differences in interaction relationships across different directions. Furthermore, the future trajectory of a pedestrian exhibits significant uncertainty. Under the same historical observation conditions, the target pedestrian may have multiple plausible future movement paths. While existing methods can generate multiple candidate trajectories, their ability to further correct prediction errors is insufficient, leading to continued bias in the prediction results. Summary of the Invention

[0004] This application proposes a pedestrian trajectory prediction method and device based on direction-aware conditional denoising, which can solve one of the problems existing in the background technology.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] Firstly, a pedestrian trajectory prediction method based on direction-aware conditional denoising is provided, including:

[0007] The historical trajectory of the target pedestrian and the movement status of neighboring pedestrians are obtained. A polar coordinate reference system is established with the target pedestrian as the center. The neighborhood space around the target pedestrian is divided into several directional sectors, and the neighboring pedestrians are distributed in the sectors.

[0008] The historical trajectory is encoded to obtain initial trajectory features;

[0009] The average value of the movement states of all neighboring pedestrians in each sector is encoded to obtain the interaction features;

[0010] By fusing the initial trajectory features and the interaction features, a historical context representation is obtained;

[0011] A random latent variable sampling mechanism is used to randomly sample the historical context representation to obtain the fused historical motion context.

[0012] Furthermore, based on the fused historical motion context, the pedestrian trajectory prediction result is obtained.

[0013] Based on the above technical solution, by dividing the space around the target pedestrian into several directional sectors, the influence of neighbors in different directions is independently modeled, enabling the model to more accurately reflect the changes in neighborhood structure in complex scenes. Furthermore, by randomly sampling latent variables to generate multiple potential future motion patterns, it can simultaneously describe different future motion directions, thereby improving the ability to express the diversity of pedestrian behavior.

[0014] In one possible design approach of the first aspect, the historical trajectory is encoded to obtain initial trajectory features, specifically by: encoding the position sequence of the target pedestrian within the historical observation time t. The input is fed into the embedding layer for feature mapping to obtain the initial trajectory features:

[0015]

[0016] in, This represents the trajectory feature sequence of the target pedestrian. Indicates the length of historical observation time. This represents the feature dimension after encoding the target pedestrian. This is an embedded function.

[0017] In one possible design of the first aspect, the motion state includes: the relative Euclidean distance between the neighboring pedestrian and the target pedestrian, the speed of the neighboring pedestrian, and the relative azimuth angle between the neighboring pedestrian and the target pedestrian;

[0018] No. Each sector Average speed of pedestrians in the middle neighborhood It can be represented as follows:

[0019] in, and Let these represent the positions of the j-th neighbor pedestrian at the current time and the historical observation time, respectively;

[0020] At that moment, the Each sector The average Euclidean distance between all neighboring pedestrians in the area can be expressed as follows:

[0021] in, Indicates the current position of the target pedestrian;

[0022] No. Each sector The average azimuth of the pedestrians in the middle neighborhood can be expressed as follows:

[0023] in, Indicates the first The polar angle of a neighboring pedestrian relative to the target pedestrian.

[0024] In one possible design approach of the first aspect, the average value of the movement states of all neighboring pedestrians in each sector is encoded, and the embedding function used is: The embedding function contains two MLP layers: the first layer uses the ReLU activation function, and the second layer uses the tanh activation function.

[0025] In one possible design approach of the first aspect, the integration of the initial trajectory features and the interaction features specifically includes:

[0026] Align the initial trajectory features with the interaction features according to time steps;

[0027] Furthermore, the aligned initial trajectory features are concatenated with the interaction features and then input into the mapping layer to form the historical context representation.

[0028] In one possible design approach of the first aspect, a random latent variable sampling mechanism is employed to randomly sample the historical context representation, specifically including:

[0029] Sample the latent variable z from the standard Gaussian distribution:

[0030] Through learnable nonlinear mapping functions Embed the latent variables into the same feature space as the historical context representation:

[0031] Embedding latent variables The fused historical motion context is obtained by fusing it with the historical motion context representation c. :

[0032] in, ,in This indicates the dimension of the latent variable.

[0033] In one possible design approach of the first aspect, the pedestrian trajectory prediction result is obtained based on the fused historical motion context, specifically including:

[0034] The fused historical motion context is encoded sequentially using the Transformer self-attention mechanism and the multi-head attention mechanism to obtain the final encoded feature h:

[0035] in, Indicates the first One point of attention, To output the projection matrix, Concat indicates concatenation.

[0036] Decode the final encoded features to obtain the pedestrian trajectory prediction result.

[0037] In one possible design approach of the first aspect, decoding the final encoded features to obtain the pedestrian trajectory prediction result specifically includes:

[0038] Introduced during the decoding phase Each sport style branch corresponds to a potential sport style, based on the fused representation. Each mode branch generates its corresponding prediction result through an independent decoder:

[0039] in, Indicates different prediction models, As a nonlinear transformation mapping, these decoders share a backbone network structure to extract common motion features, while generating prediction results of different modes through independent output layers, thus enabling the model to predict multiple potential future trajectories simultaneously.

[0040] Assuming the future observation duration is This module only predicts The intermediate target points will yield intermediate features. Perform a linear transformation to obtain the prediction results. It can be represented as:

[0041] in, To predict the number of modes, This represents the number of intermediate target points. Represents two-dimensional coordinates;

[0042] Let the last observation point be Then the complete intermediate target point sequence for:

[0043] After obtaining the sequence of intermediate target points, a piecewise linear interpolation method is used to generate a complete trajectory between adjacent intermediate target points. Let the two adjacent intermediate target points be: The corresponding time is: For any timestamp t within the interval, the position of the trajectory for:

[0044] The complete trajectory can be obtained by interpolating across all intervals. for:

[0045] in, Where D represents the duration of future observations, and D represents the dimension of the complete trajectory. It is a linear interpolation calculation of the trajectory position at time t, used to calculate the future... Each time step The complete trajectory can then be obtained. .

[0046] In one possible design of the first aspect, the pedestrian trajectory prediction method based on direction-aware conditional denoising further includes:

[0047] An initial predicted trajectory is obtained by sampling from the complete trajectory;

[0048] Using the historical trajectories of the target pedestrian and the neighboring pedestrians as conditional inputs, the initial predicted trajectory of Sohu is input into the conditional denoising network;

[0049] Furthermore, the prediction error is gradually reduced through multi-step residual correction to output the final predicted trajectory.

[0050] In a second aspect, an electronic device is provided, comprising: a processor and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the pedestrian trajectory prediction method based on orientation-aware conditional denoising as described in any possible implementation of the first aspect. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a framework diagram of the pedestrian trajectory prediction method based on the direction-aware conditional denoising framework provided in the embodiments of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0056] This embodiment provides a pedestrian trajectory prediction method that can simultaneously achieve directional interaction modeling, multimodal trajectory prediction, and trajectory error refinement.

[0057] This embodiment provides a pedestrian trajectory prediction method based on a direction-aware conditional denoising framework, addressing the problems of insufficient neighborhood direction interaction representation, limited multimodal prediction capabilities, and insufficient trajectory refinement capabilities in existing technologies. This embodiment includes a motion coding module and a multimodal and uncertainty modeling module. The motion coding module includes a target pedestrian motion coding module and a neighboring pedestrian feature coding module; the multimodal and uncertainty modeling module includes a multimodal trajectory prediction module and a trajectory refinement module.

[0058] (a) Target pedestrian motion coding module

[0059] The target pedestrian motion encoding module is used to encode the historical trajectory of the target pedestrian. Specifically, it includes encoding the position sequence of the target pedestrian within the historical observation time. The input is fed into the embedding layer for feature mapping to obtain the initial trajectory features of the target pedestrian:

[0060] in:

[0061] This represents the trajectory feature sequence of the target pedestrian. Indicates the length of historical observation time. This represents the feature dimension after encoding the target pedestrian. Embedding function. It consists of two fully connected layers, with ReLU and tanh activation functions used in the middle to enhance feature representation capabilities.

[0062] (ii) Neighbor pedestrian feature coding module

[0063] The neighbor pedestrian feature encoding module is used to extract directional interaction information of neighboring pedestrians around the target pedestrian. Specifically, this includes: establishing a polar coordinate reference system centered on the target pedestrian; and dividing the neighborhood space around the target pedestrian into 12 directional sectors, namely:

[0064] In this embodiment, polar coordinates are established with the target pedestrian as the clock center (i.e., the polar coordinate center), and the surrounding neighbor space is divided into 12 equal sectors, which are denoted as follows: and define the angle Indicates neighboring pedestrians With the target pedestrian The relative position can be expressed by the formula:

[0065] in pedestrian timestamp The location.

[0066] Then, the movement status of neighboring pedestrians within each sector in each direction is statistically analyzed, and let the i-th sector be... Each sector The number of pedestrians in the inner neighboring area is The pedestrian movement characteristics corresponding to this sector are represented as follows: The motion states include: relative distance; average velocity; and relative azimuth angle.

[0067] Speed ​​characteristics:

[0068] During pedestrian movement, individuals typically judge whether to slow down, yield, or follow based on the speed of surrounding pedestrians. Therefore, the speed information of neighboring pedestrians plays a crucial role in predicting the future trajectory of a target pedestrian. To this end, this embodiment calculates the average speed of all neighboring pedestrians within each sector to reflect the overall movement trend in that direction. Each sector The average speed of pedestrians in the middle neighborhood can be expressed as follows:

[0069] in, and Let represent the positions of the j-th neighboring pedestrian at the current time and the starting time of the historical observation, respectively.

[0070] Distance features:

[0071] The spatial distance between neighboring pedestrians and the target pedestrian also affects the interaction intensity. Generally, closer neighboring pedestrians have a more significant impact on the target pedestrian, while farther pedestrians have a relatively weaker impact. Therefore, this embodiment calculates the average Euclidean distance between neighboring pedestrians and the target pedestrian within each sector to characterize the spatial interaction relationship in that direction. Each sector The average Euclidean distance of all neighboring pedestrians in the middle ( (Time) can be represented as follows:

[0072] in, This indicates the current position of the target pedestrian.

[0073] Angular features:

[0074] Although the space surrounding the target pedestrian is essentially continuous, dividing it into 12 discrete sectors may result in some loss of angular information. To more precisely describe the positional relationships of neighboring pedestrians relative to the target pedestrian, this embodiment further calculates the average angle of neighboring pedestrians relative to the target pedestrian within each sector to supplement spatial orientation information. Each sector The average azimuth of the pedestrians in the middle neighborhood can be expressed as follows:

[0075] in, Indicates the first The polar angle of a neighboring pedestrian relative to the target pedestrian.

[0076] Combining the above three characteristics, the first Each sector The characteristics of an expert's movement are:

[0077] in, The embedding function consists of two MLP layers. The first layer uses the ReLU activation function, and the second layer uses the tanh activation function to map the statistically obtained interaction features to a high-dimensional feature space.

[0078] Based on this, the model fuses the target pedestrian motion encoding results with the directional interaction representation. Specifically, this includes: aligning the target pedestrian's historical motion features and directional interaction features according to time steps; concatenating the two and inputting them into a mapping layer to form a unified historical context representation; the historical context representation is used for subsequent multimodal trajectory prediction.

[0079] As can be seen from the above, the historical movement characteristics of the target pedestrian can be represented as follows: ,in Indicates the duration of historical observations. The historical motion characteristics of the target pedestrian are represented by the following dimension: the interaction characteristics of neighboring pedestrians within each sector are represented by... The overall neighbor movement characteristics of the 12 sectors are represented as follows: , This represents the interaction feature dimension of neighboring pedestrians within each sector.

[0080] To align neighbor interaction features with the time-series features of the target pedestrian, this embodiment first uses a linear transformation layer to... Perform a dimension mapping to convert it into a representation consistent with the time dimension. Then, the target pedestrian's motion is encoded. Interaction coding with neighboring pedestrians Merging them along the time dimension yields the historical motion context:

[0081] This context feature It also includes the target pedestrian's own motion state information as well as the interaction information of surrounding pedestrians, which can more comprehensively depict the dynamic environment in which the pedestrian is located and serve as an important input for the subsequent trajectory prediction module.

[0082] (III) Multimodal trajectory prediction module

[0083] The multimodal trajectory prediction module is used to generate multiple potential future trajectories. Specifically, it includes: randomly sampling a unified historical context representation to introduce latent variables; mapping the latent variables and historical context together to a unified feature space; inputting a self-attention structure to calculate the correlation between different time steps; and outputting key points of the future trajectories under multiple potential motion patterns. Each latent pattern corresponds to an independent decoding channel; each decoding channel outputs key target points in the future trajectory; and finally, the complete future trajectory sequence is recovered through interpolation.

[0084] To address motion uncertainties, this embodiment introduces a random latent variable sampling mechanism during the forward computation process. For the same historical trajectory input, latent variables are sampled from a standard Gaussian distribution:

[0085] in, ,in This represents the dimension of the latent variable. This latent variable is used to characterize the random perturbations of the pedestrian during future movement. Subsequently, a learnable nonlinear mapping function is used... Embed the noise into the same feature space as the historical context:

[0086] in, Indicates the number of historical observation steps. This represents the feature dimension after mapping.

[0087] Based on this, the latent variable embeddings are fused with historical motion context features to obtain the input representation for subsequent spatiotemporal modeling:

[0088] To further model the dependencies of historical trajectories in the time dimension and the spatial interaction between the target pedestrian and its neighbors, this embodiment uses the Transformer self-attention mechanism to encode the fused historical motion context.

[0089] Specifically, the merged historical movement context features The input is fed into the Transformer module, and a query, key, and value are generated through linear mapping: , , .in , and This is the learnable parameter matrix.

[0090] Subsequently, a multi-head self-attention mechanism is employed to capture feature relationships in different subspaces:

[0091] The outputs of multiple attention heads are concatenated and subjected to a linear transformation to obtain the final encoded features:

[0092] in, Indicates the first One point of attention, This is for outputting the projection matrix.

[0093] By employing multi-head attention computation, the model can simultaneously learn the dynamic dependencies of historical trajectories over time, as well as the spatial interaction patterns between the target pedestrian and its neighbors. This results in a richer spatiotemporal motion representation. This provides a foundation for subsequent multimodal trajectory prediction.

[0094] To address intent uncertainty, this embodiment introduces [a specific feature] during the decoding phase. Each branch represents a potential motion style, such as: straight ahead, veering left, veering right, accelerating, decelerating, etc. This is based on the fused representation. Each mode branch generates its corresponding prediction result through an independent decoder:

[0095] in, Indicates different prediction models, This is a nonlinear transformation mapping. These decoders share a backbone network structure to extract common motion features, while generating predictions of different modes through independent output layers, enabling the model to predict multiple potential future trajectories simultaneously.

[0096] Assuming the future observation duration is This module only predicts The intermediate target points, and the intermediate features obtained above After performing a linear transformation, the prediction result can be expressed as:

[0097] in To predict the number of modes, This represents the number of intermediate target points. Represents two-dimensional coordinates.

[0098] Since intermediate target points only cover a portion of the timestamp locations in the future trajectory, it is necessary to further convert the intermediate target point sequence into a complete trajectory sequence. To ensure trajectory continuity, the model adds the last observation point as an anchor point to the sequence before interpolation. Let the last observation point be... The complete sequence of intermediate target points is then:

[0099] This design can avoid discontinuous jumps between the predicted trajectory and the historical trajectory.

[0100] After obtaining the sequence of intermediate target points, this module uses piecewise linear interpolation to generate a complete trajectory between adjacent intermediate target points. Let two adjacent intermediate target points be: The corresponding time is: For any timestamp t within the interval, the position of the trajectory is:

[0101] By interpolating across all intervals, the complete trajectory can be obtained as follows:

[0102] in, It is the number of future observation steps. It is a linear interpolation calculation of the trajectory position at time t, used to calculate the future... Each time step The complete trajectory can then be obtained. .

[0103] (v) Trajectory Refinement Module

[0104] The trajectory refinement module is used to correct errors in the initial predicted trajectory. Specifically, it includes: sampling the initial predicted trajectory from the trajectory distribution output by the multimodal trajectory prediction module; using the historical trajectories of the target pedestrian and neighboring pedestrians as conditional inputs; inputting the initial predicted trajectory into a conditional denoising network; gradually reducing the prediction error through multi-step residual correction; and outputting the final predicted trajectory. The conditional denoising network updates the prediction residual at each step to improve trajectory continuity and smoothness.

[0105] To further improve prediction accuracy, this embodiment introduces a conditional diffusion model (DDPM) to refine the initial trajectory.

[0106] Since the above operations have generated a trajectory sequence that approximates the future trajectory, in order to save the time cost of training and inference, this module does not need to generate a complete trajectory through Gaussian distribution like the traditional diffusion model. Instead, it refines the trajectory by modeling the prediction error, thereby reducing the uncertainty of trajectory prediction.

[0107] The initial trajectory obtained by linear interpolation is Let the true trajectory be The residual is then defined as:

[0108] The goal of the diffusion model is to learn the probability distribution of the residual.

[0109] In the conditional diffusion framework, the model's conditional information is composed of both historical observation trajectories and initial prediction trajectories:

[0110] in, For the historical trajectory of the target pedestrian, Historical observation tracks of neighboring pedestrians.

[0111] During the training phase, the diffusion model generates noise residuals by progressively adding Gaussian noise:

[0112] in, Indicates standard Gaussian noise. These are the noise dispatch coefficients during the diffusion process. The training objective of this module is to predict the noise term. That is, learning a noise prediction function:

[0113] Used to estimate the noise added in the current time step. During training, network parameters are optimized by minimizing the mean squared error loss.

[0114] In this way, the network can gradually learn how to recover the original residual structure from noisy residuals. During the inference phase, the diffusion model starts with random Gaussian noise and gradually generates residual estimates through a multi-step iterative denoising process. Initialization begins with:

[0115] Then, the residuals are updated step by step according to the reverse diffusion process:

[0116] in For random noise, This indicates the first step in the reverse diffusion process. The noise standard deviation corresponding to each time step is used to control the injection intensity of random perturbations in the current iteration.

[0117] The final predicted residuals can be obtained: And correct the initial trajectory.

[0118] In this implementation, the model was implemented on an NVIDIA 3090 GPU server using the PyTorch framework. The model parameters are set as follows: Polar coordinates are established centered on the target pedestrian, dividing the surrounding neighborhood space into 12 equal sectors, each with an angle of 30 degrees. The hidden features of the MLP layer are set to 32, and the activation function is ReLU. The trajectory data is set to a historical observation length of 8 time steps and a future observation length of 12 time steps, with each time step being 0.4 seconds. A "leave-one-out" cross-validation strategy is adopted, meaning the model is trained on four scenarios and tested on the remaining scenario.

[0119] The core innovation of this embodiment lies in generating a historical context representation with directional structure through direction-aware interactive modeling, generating a multimodal future trajectory distribution based on this, and then refining the predicted trajectory step by step through a conditional denoising mechanism, thereby realizing an integrated prediction framework for direction interactive modeling, multimodal prediction and trajectory error correction.

[0120] Compared with the prior art, this embodiment has the following beneficial effects:

[0121] 1. Improve directional interaction modeling capabilities: By dividing the space around the target pedestrian into 12 directional sectors, the influence of neighbors in different directions is modeled independently, enabling the model to more accurately reflect the changes in neighborhood structure in complex scenes.

[0122] 2. Enhance the ability to represent multimodal trajectories: By randomly sampling latent variables to generate multiple potential future motion patterns, it can simultaneously describe different future motion directions, thereby improving the ability to represent the diversity of pedestrian behavior.

[0123] 3. Improve trajectory prediction accuracy: The initial trajectory is progressively corrected by a conditional denoising network, which reduces prediction error and improves trajectory continuity and stability.

[0124] 4. Improve adaptability to complex scenarios: This embodiment is applicable to pedestrian trajectory prediction tasks in complex dynamic environments such as autonomous driving, intelligent monitoring, and human-machine collaborative navigation.

[0125] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0126] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0127] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0128] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0129] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0130] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0131] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0132] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A pedestrian trajectory prediction method based on directional perception condition denoising, characterized in that, include: The historical trajectory of the target pedestrian and the movement status of neighboring pedestrians are obtained. A polar coordinate reference system is established with the target pedestrian as the center. The neighborhood space around the target pedestrian is divided into several directional sectors, and the neighboring pedestrians are distributed in the sectors. The historical trajectory is encoded to obtain initial trajectory features; The average value of the movement states of all neighboring pedestrians in each sector is encoded to obtain the interaction features; By fusing the initial trajectory features and the interaction features, a historical context representation is obtained; A random latent variable sampling mechanism is used to randomly sample the historical context representation to obtain the fused historical motion context; and, based on the fused historical motion context, the pedestrian trajectory prediction result is obtained.

2. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 1, characterized in that, Encode the historical trajectory to obtain an initial trajectory feature, specifically: a position sequence of the target pedestrian within a historical observation time t Input into an embedding layer for feature mapping to obtain the initial trajectory feature: in, This represents the trajectory feature sequence of the target pedestrian. Indicates the length of historical observation time. This represents the feature dimension after encoding the target pedestrian. This is an embedded function.

3. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 1, characterized in that, The motion state includes: the relative Euclidean distance between the neighboring pedestrian and the target pedestrian, the speed of the neighboring pedestrian, and the relative azimuth angle between the neighboring pedestrian and the target pedestrian. No. Each sector Average speed of pedestrians in the middle neighborhood It is expressed as follows: in, and These represent the positions of the j-th neighbor pedestrian at the current time and the starting time of historical observation, respectively. At that moment, the Each sector The average Euclidean distance between all neighboring pedestrians in the area can be expressed as follows: in, This indicates the current position of the target pedestrian. No. Each sector The average azimuth angle of the pedestrians in the middle neighborhood is expressed as follows: in, Indicates the first The polar angle of a neighboring pedestrian relative to the target pedestrian.

4. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 1, characterized in that, The average value of the movement states of all neighboring pedestrians in each sector is encoded using the following embedding function: The embedding function contains two MLP layers: the first layer uses the ReLU activation function, and the second layer uses the tanh activation function.

5. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 1, characterized in that, The fusion of the initial trajectory features and the interaction features specifically includes: Align the initial trajectory features with the interaction features according to time steps; and The aligned initial trajectory features are concatenated with the interaction features and then input into the mapping layer to form the historical context representation.

6. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 1, characterized in that, A random latent variable sampling mechanism is employed to randomly sample the historical context representation, specifically including: Sample the latent variable z from the standard Gaussian distribution: Through learnable nonlinear mapping functions Embed the latent variables into the same feature space as the historical context representation: Embedding latent variables The fused historical motion context is obtained by fusing it with the historical motion context representation c. : in, ,in This indicates the dimension of the latent variable.

7. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 1, characterized in that, Based on the fused historical motion context, the pedestrian trajectory prediction result is obtained, specifically including: The fused historical motion context is encoded sequentially using the Transformer self-attention mechanism and the multi-head attention mechanism to obtain the final encoded feature h: in, Indicates the first One point of attention, To output the projection matrix, Concat indicates concatenation. Decode the final encoded features to obtain the pedestrian trajectory prediction result.

8. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 7, characterized in that, Decoding the final encoded features to obtain the pedestrian trajectory prediction result specifically includes: Introduced during the decoding phase Each sport style branch corresponds to a potential sport style, based on the fused representation. Each mode branch generates its corresponding prediction result through an independent decoder: in, Indicates different prediction models, As a nonlinear transformation mapping, these decoders share a backbone network structure to extract common motion features, while generating predictions of different modes through independent output layers. This enables the model to predict multiple potential future trajectories simultaneously. Assuming the future observation duration is This module only predicts The intermediate target points will yield intermediate features. Perform a linear transformation to obtain the prediction results. It can be represented as: in, To predict the number of modes, The number of intermediate target points. Represents two-dimensional coordinates. Let the last observation point be Then the complete intermediate target point sequence for: After obtaining the sequence of intermediate target points, a piecewise linear interpolation method is used to generate a complete trajectory between adjacent intermediate target points. Let the two adjacent intermediate target points be: The corresponding time is: For any timestamp t within the interval, the position of the trajectory for: The complete trajectory can be obtained by interpolating across all intervals. for: where T pred is the future observation duration.

9. The pedestrian trajectory prediction method based on direction-aware conditional denoising as described in claim 8, characterized in that, The pedestrian trajectory prediction method based on direction-aware conditional denoising also includes: An initial predicted trajectory is obtained by sampling from the complete trajectory; Using the historical trajectories of the target pedestrian and the neighboring pedestrians as conditional inputs, the initially predicted trajectory from Sohu is input into a conditional denoising network; and, The prediction error is gradually reduced through multi-step residual correction to output the final predicted trajectory.

10. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the pedestrian trajectory prediction method based on orientation-aware conditional denoising as described in any one of claims 1-9.

Images