A pedestrian trajectory prediction method based on motion intention extraction

Abstract

The application discloses a pedestrian trajectory prediction method based on motion intention extraction, comprising the following steps: 1) mapping the historical trajectory of the pedestrian and the end position to the hidden space to obtain the corresponding historical hidden space feature representation and the end hidden space feature representation; 2) splicing the historical hidden space feature representation and the end hidden space feature representation together and inputting the spliced feature representation into a conditional variational autoencoder to output the motion intention feature representation of the pedestrian; 3) inputting the motion intention feature representation of the pedestrian into a graph neural network to learn the motion intention interaction relationship between the pedestrians in the group and the pedestrians between the groups respectively and updating the motion intention feature representation of the pedestrian; and 4) splicing the updated motion intention feature representation of the pedestrian and the historical hidden space feature representation together, inputting the spliced feature representation into a fusion network and inputting the fused feature representation into a predictor to obtain the final future trajectory of the pedestrian. The application can effectively extract the motion intention of the pedestrian, fuse the motion interaction information between different pedestrians and improve the precision of the pedestrian motion prediction.

Description

Technical Field

[0001] This invention relates to the technical field of motion intent extraction and pedestrian trajectory prediction, and in particular to a method for pedestrian trajectory prediction based on motion intent extraction. Background Technology

[0002] Predicting the motion of dynamic objects is central to research on the behavioral patterns of autonomous agents (such as humans, social robots, or self-driving cars). In complex and ever-changing environments, safe and smooth path planning truly requires evaluation through prediction. Humans are among the most frequently encountered dynamic objects in everyday traffic environments. Therefore, predicting human motion plays a crucial role in improving the accuracy and versatility of navigation, planning, human-computer interaction, and other autonomous tasks.

[0003] However, due to the highly random and interactive nature of human movement, predicting human movement presents significant requirements and challenges. In particular, when humans navigate around obstacles to achieve their goals, they actively modify their movements and dynamically adjust their paths based on the environment. Existing rule-based prediction methods and time-series prediction methods based on recurrent neural networks lack the ability to adaptively adjust to dynamic environments and do not model and learn about pedestrian movement intentions and their interactions. Therefore, to make accurate predictions, we need to comprehensively consider various factors, such as historical movement trajectories, potential movement intentions, other moving entities in the scene, and group behavior patterns, to help achieve more accurate pedestrian trajectory prediction. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a pedestrian trajectory prediction method based on motion intention extraction. This method can effectively extract the motion intention of pedestrians and integrate motion interaction information between different pedestrians to improve the accuracy of pedestrian motion prediction, enhance the compatibility of predicted trajectories between different pedestrians, and further provide more accurate environmental perception information for downstream autonomous driving tasks.

[0005] To achieve the above objectives, the technical solution provided by this invention is: a pedestrian trajectory prediction method based on motion intent extraction, comprising the following steps:

[0006] 1) Use the historical feature encoder and the endpoint feature encoder to map the pedestrian's historical trajectory and endpoint position to the latent space to obtain the corresponding historical latent space feature representation and endpoint latent space feature representation;

[0007] 2) The historical latent space feature representation and the endpoint latent space feature representation are concatenated together and input into the conditional variational autoencoder to predict the conditional probability distribution of the pedestrian's motion intention, and the output is the feature representation of the pedestrian's motion intention.

[0008] 3) Input the pedestrian's motion intention feature representation into the graph neural network, divide the pedestrians into different groups according to distance and time, learn the motion intention interaction relationship between pedestrians within the group and between groups based on the graph neural network, and update the pedestrian's motion intention feature representation;

[0009] 4) The updated pedestrian motion intention feature representation is concatenated with the historical latent space feature representation and input into the fusion network for feature fusion. Finally, the fused feature representation is input into the predictor to obtain the final pedestrian future trajectory.

[0010] Furthermore, in step 1), the pedestrian's trajectory points refer to the pedestrian's 2D position coordinate sequence, represented as:

[0011]

[0012] In the formula, P i This refers to the sequence of movement trajectory points of the i-th pedestrian, where there are a total of N pedestrians. express It is a 2-dimensional vector. Represents a 2-dimensional vector. and Let be the X-axis and Y-axis components of pedestrian i at time t in the pixel coordinate system, with a total of T time steps. These T time steps are divided into two parts: historical time steps and future time steps. The historical trajectory of pedestrian i is the set of trajectory points from the historical time steps, and the future trajectory is the set of trajectory points from the future time steps. Therefore, the historical trajectory and future trajectory of pedestrian i are represented as follows:

[0013]

[0014]

[0015] In the formula, t obs This indicates the number of steps in the historical timeline. This represents the set of trajectory points of pedestrian i over historical time steps. express It is a 2-dimensional vector. and These are the X-axis and Y-axis components of pedestrian i in the pixel coordinate system at time t1, totaling t. obs A historical time step; from t obs +1 to t pred Y represents a future time step. i pred This represents the set of trajectory points for pedestrian i at future time steps. express It is a 2-dimensional vector. and These are the X-axis and Y-axis components of pedestrian i in the pixel coordinate system at time t2, totaling t. pred -t obs There are several future time steps; predicting the trajectory points of these future time steps is the target; and the endpoint of pedestrian i's movement is taken as the trajectory point of the last step in the future time steps, represented as:

[0016]

[0017] In the formula, d i t represents the final position of pedestrian i's movement. pred This indicates the last time step in the future. and Is the pedestrian i in t pred The X and Y axis components in the pixel coordinate system at time step; using two independent multilayer perceptrons, the historical trajectory and endpoint position of pedestrian i are mapped to the latent space, resulting in the corresponding historical latent space feature representation and endpoint latent space feature representation, expressed as:

[0018]

[0019]

[0020] In the formula, and These are the historical latent space feature representations and the destination latent space feature representations for pedestrian i, respectively. his and f des These represent the historical feature encoder and the endpoint feature encoder, respectively, both of which are composed of multiple fully connected layers.

[0021] Furthermore, step 2) includes the following steps:

[0022] 2.1) Concatenate the historical latent space feature representation of pedestrian i with the endpoint latent space feature representation to obtain the motion information feature representation, as follows:

[0023]

[0024] In the formula, Let [·||·] represent the motion information feature representation of pedestrian i, where [·||·] represents the concatenation operation, that is, concatenating two vectors together along the feature dimension;

[0025] 2.2) The motion information feature representation of the pedestrian is input into a conditional variational autoencoder, which consists of an encoder and a decoder, both of which are composed of multiple fully connected layers; The input is fed into the encoder to obtain the posterior probability distribution of the latent variable z. Here, the latent variable z is a random variable representing the pedestrian's endpoint position; then, the latent variable z is sampled and input into the decoder along with the historical latent space feature representation to obtain the conditional probability distribution of the pedestrian's movement intention. Finally, by solving the marginal distribution of the conditional probability distribution of pedestrian movement intention, we can obtain the conditional probability distribution of pedestrian i's movement intention based on historical movement information, which is expressed as:

[0026]

[0027] In the formula, This represents the predicted movement intention feature of pedestrian i. The conditional probability distribution representing the movement intention of pedestrian i based on historical movement information is used. The decoder will ultimately output a feature vector, which represents the movement intention feature of pedestrian i.

[0028] Furthermore, step 3) includes the following steps:

[0029] 3.1) Input the pedestrian's motion intention feature into a graph neural network, and divide pedestrians into different groups according to distance and time proximity. Pedestrian i and pedestrian j are represented as nodes v. i and v j With undirected edge e i,j By making connections, the pedestrians within the group construct a fully connected graph, represented as follows:

[0030] G b ={V b E b}

[0031]

[0032]

[0033] In the formula, G b Let N represent the b-th group, which is the b-th fully connected graph. There are a total of B fully connected graphs, each of which is a set of nodes and edges within a graph. b V represents the set of indices of the nodes contained in the b-th fully connected graph. b E is the set of nodes in the b-th fully connected graph. b It is the set of edges of the b-th fully connected graph; here, a mask M is used. i,j To indicate whether pedestrian i and pedestrian j are in the same group, it is represented as:

[0034]

[0035] In the formula, M i,j This indicates whether pedestrian i and pedestrian j are in the same group. This represents pedestrian i at time step t. Let j be a pedestrian at time step s. If the minimum of the 2-norms of the two is... Greater than the set threshold d thresh Then M i,j Setting it to 0 indicates that pedestrian i and pedestrian j are not in the same group; otherwise, M i,j A value of 1 indicates that the two are in the same group;

[0036] 3.2) Based on the self-attention network and graph message passing mechanism, the movement intentions of pedestrians in the same group are updated to learn the mutual influence of movement intentions among different pedestrians in the same group. The updated movement intention of pedestrian i is:

[0037]

[0038] In the formula, σ is the updated representation of pedestrian i's motion intention features, where σ is a non-linear activation function used to enhance the network's non-linear expressive power. This indicates a weighted summation of the motion intention features of other pedestrians j within the same group. Let represent the set of other pedestrians in the same group as pedestrian i, and W be the weight matrix of the graph neural network used to learn the interaction relationships of movement intentions between different pedestrians in the group; α ij The influence weight of pedestrian j relative to pedestrian i is obtained through adaptive learning via a self-attention network.

[0039] Furthermore, step 4) includes the following steps:

[0040] 4.1) The updated pedestrian motion intention feature representation and the historical latent space feature representation are input into a fusion network for feature fusion, making the predicted trajectory more consistent with the actual motion situation and more compatible with the motion patterns of different pedestrians; for pedestrian i, the historical latent space feature representation and the updated motion intention feature representation are first concatenated together as the feature representation before fusion:

[0041]

[0042] In the formula, This represents the pre-fusion feature representation of pedestrian i before input to the fusion network, where [·||·] represents the vector concatenation operation along the feature dimension direction;

[0043] 4.2) The obtained pre-fusion feature representations are input into the fusion network for feature fusion. A total of K rounds of feature fusion are performed. The feature representation of the (k+1)th round is obtained by updating the feature representation of the kth round. The iterative process is as follows:

[0044]

[0045] In the formula, Let i be the feature representation of pedestrian i in the (k+1)th round. Let i be the feature representation of pedestrian i in the k-th round. Let g(·) be the feature representation of pedestrian j in the k-th round, and g(·) be a transform encoder used to process the feature representation of the k-th round. Perform remapping, f s (·,·) is a similarity network used to calculate the feature similarity between pedestrian i and pedestrian j. The result is a scalar, and the calculation process is as follows:

[0046]

[0047] In the formula, ReLu(·) represents the ReLU activation function, which maps the result to a value between [0,1]. θ(·) and φ(·) are both transform encoders consisting of three fully connected layers, which map the feature representations of pedestrian i and pedestrian j in the k-th round to the learnable latent space, respectively. [·||·] represents the vector concatenation operation along the feature dimension. This represents a single fully connected layer used to map the concatenated feature representations to a scalar.

[0048] 4.3) After K rounds of iteration, the final fused feature representation is obtained. It integrates both its own historical motion information and the motion intention information of other pedestrians in the surrounding environment, and finally represents the fused features. The input is fed into a predictor consisting of three fully connected layers. Finally, it passes through a fully connected layer to map the feature dimension to the target prediction time step length, thus obtaining the prediction of the pedestrian's future trajectory.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] 1. This invention explicitly models and learns the movement intentions of pedestrians and the interactions between them. Pedestrians are divided into different groups according to spatial and temporal distance, and fully connected graphs are constructed within and between groups. The interactions between the movement intentions of different pedestrians are learned based on the message passing mechanism. By introducing the learning of pedestrian movement intentions and their interactions, better prediction results are achieved compared with previous methods that did not learn this.

[0051] 2. In order to effectively integrate pedestrian historical movement information and movement intention information, this invention proposes an iterative information fusion method based on neural networks, which can effectively integrate known information to improve the trajectory prediction compatibility among different pedestrians and make the predicted trajectory more consistent with the real movement situation.

[0052] 3. Unlike previous time series prediction methods, such as recurrent neural networks (RNN) and long short-term neural networks (LSTM), this invention does not use the method of sequentially inputting trajectory sequences. Instead, it flattens the trajectory sequences and inputs them into the network simultaneously. This can greatly improve the speed at which the model processes information and has wider applicability in scenarios with higher timeliness requirements. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the logic flow of the present invention.

[0054] Figure 2 This is an end-to-end architecture diagram of the present invention. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0056] This example uses pedestrian trajectory data recorded in the Stanford UAV dataset. The dataset contains 20 scenes with more than 11,000 different pedestrians. Each trajectory sequence is 8 seconds long, and the acquisition frequency is 0.4 frames / second, for a total of 20 frames. Each frame is a time step, meaning each trajectory sequence contains 20 trajectory points.

[0057] like Figure 1 and Figure 2 As shown, this embodiment discloses a pedestrian trajectory prediction method based on motion intent extraction, which includes the following steps:

[0058] 1) Use the historical feature encoder and the endpoint feature encoder to map the pedestrian's historical trajectory and endpoint position to the latent space to obtain the corresponding historical latent space feature representation and endpoint latent space feature representation;

[0059] A pedestrian's trajectory points refer to the sequence of 2D position coordinates of the pedestrian, represented as:

[0060]

[0061] In the formula, P i This refers to the sequence of movement trajectory points of the i-th pedestrian, where there are a total of N pedestrians. express It is a 2-dimensional vector. Represents a 2-dimensional vector. and Let be the X-axis and Y-axis components of pedestrian i at time t in the pixel coordinate system, with a total of T time steps. These T time steps are divided into two parts: historical time steps and future time steps. The historical time step is 8 frames long, and the future time step is 12 frames long. Pedestrian i's historical trajectory is the set of trajectory points from the historical time steps, and its future trajectory is the set of trajectory points from the future time steps. Therefore, pedestrian i's historical trajectory and future trajectory are represented as follows:

[0062]

[0063]

[0064] In the formula, This represents the set of trajectory points of pedestrian i over historical time steps. express It is a 2-dimensional vector. and It represents the X-axis and Y-axis components of pedestrian i at time t1 in the pixel coordinate system, with a total of 8 historical time steps. i pred This represents the set of trajectory points for pedestrian i at future time steps. express It is a 2-dimensional vector. and These are the X and Y axis components of pedestrian i in the pixel coordinate system at time t2, with a total of 12 future time steps. Predicting the trajectory points of these future time steps is the objective of this invention; the endpoint of pedestrian i's movement is taken as the trajectory point of the last step in these future time steps, expressed as:

[0065]

[0066] In the formula, d i This indicates the final position of pedestrian i's movement. and These are the X-axis and Y-axis components of pedestrian i at the 20th time step in the pixel coordinate system. Two independent multilayer perceptrons are used to map the historical trajectory and endpoint position of pedestrian i into the latent space, yielding the corresponding historical latent space feature representation and endpoint latent space feature representation, expressed as follows:

[0067]

[0068]

[0069] In the formula, and These are the historical latent space feature representations and the destination latent space feature representations for pedestrian i, respectively. The feature dimension, or vector dimension, is set to 256. The feature dimension is set to 16, f his and f des These represent the historical feature encoder and the endpoint feature encoder, respectively, f his It consists of two fully connected layers with feature dimensions of 512 and 256, respectively. des It consists of two fully connected layers with feature dimensions of 8 and 16, respectively.

[0070] 2) The historical latent space feature representation and the endpoint latent space feature representation are concatenated and input into a conditional variational autoencoder to predict the conditional probability distribution of the pedestrian's motion intention, outputting a feature representation of the pedestrian's motion intention. This includes the following steps:

[0071] 2.1) Concatenate the historical latent space feature representation of pedestrian i with the endpoint latent space feature representation to obtain the motion information feature representation, as follows:

[0072]

[0073] In the formula, Let [·||·] represent the motion information features of pedestrian i, where [·||·] represents the concatenation operation, which concatenates two vectors along the feature dimension. The feature dimension is 272.

[0074] 2.2) The pedestrian motion information feature representation is input into a conditional variational autoencoder. The conditional variational autoencoder consists of an encoder and a decoder, both of which are composed of multiple fully connected layers. Here, the number of fully connected layers is set to 3, and the feature dimensions are 1024, 512, and 1024, respectively. The input is fed into the encoder to obtain the posterior probability distribution of the latent variable z. Here, the latent variable z is a random variable representing the pedestrian's final position. Then, the latent variable z is sampled and input into the decoder along with the historical latent space feature representation to obtain the conditional probability distribution of the pedestrian's movement intention. Finally, by solving the marginal distribution of the conditional probability distribution of pedestrian movement intention, we can obtain the conditional probability distribution of pedestrian i's movement intention based on historical movement information, which is expressed as:

[0075]

[0076] In the formula, This represents the predicted movement intention feature of pedestrian i. The feature dimension is set to 16. The conditional probability distribution representing the movement intention of pedestrian i based on historical movement information is used. The decoder will ultimately output a feature vector, which represents the movement intention feature of pedestrian i.

[0077] 3) Input the pedestrian's motion intention feature representation into a graph neural network, divide pedestrians into different groups according to distance and time, learn the interaction relationship of pedestrians' motion intentions within and between groups based on the graph neural network, and update the pedestrian's motion intention feature representation, including the following steps:

[0078] 3.1) Input the pedestrian's motion intention feature into a graph neural network, and divide pedestrians into different groups according to distance and time proximity. Pedestrian i and pedestrian j are represented as nodes v. i and v j With undirected edge e i,j By making connections, the pedestrians within the group construct a fully connected graph, represented as follows:

[0079] G b ={V b E b}

[0080]

[0081]

[0082] In the formula, G b Let N represent the b-th group, which is the b-th fully connected graph. There are a total of B fully connected graphs, each of which is a set of nodes and edges within a graph. b V represents the set of indices of the nodes contained in the b-th fully connected graph. b E is the set of nodes in the b-th fully connected graph. b It is the set of edges of the b-th fully connected graph. Here, a mask M is used. i,j To indicate whether pedestrian i and pedestrian j are in the same group, it is represented as:

[0083]

[0084] In the formula, M i,j This indicates whether pedestrian i and pedestrian j are in the same group. This represents pedestrian i at time step t. Let j be a pedestrian at time step s. If the minimum of the 2-norms of the two is... Greater than the set threshold d thresh , here will d thresh If set to 50 meters, then M i,j Setting it to 0 indicates that pedestrian i and pedestrian j are not in the same group; otherwise, M i,j A value of 1 indicates that the two are in the same group;

[0085] 3.2) Based on the self-attention network and graph message passing mechanism, the movement intentions of pedestrians in the same group are updated to learn the mutual influence of movement intentions among different pedestrians in the same group. The updated movement intention of pedestrian i is:

[0086]

[0087] In the formula, This represents the updated motion intention feature of pedestrian i. The feature dimension is set to 16, and σ is a non-linear activation function used to enhance the non-linear expressive power of the network. This indicates a weighted summation of the motion intention features of other pedestrians j within the same group. Let represent the set of other pedestrians in the same group as pedestrian i, and W be the weight matrix of the graph neural network used to learn the interaction relationships of movement intentions among different pedestrians in the group. α ij The influence weight of pedestrian j relative to pedestrian i is obtained through adaptive learning via a self-attention network.

[0088] 4) The updated pedestrian motion intention feature representation is concatenated with the historical latent space feature representation, and then input into the fusion network for feature fusion. Finally, the fused feature representation is input into the predictor to obtain the final future trajectory of the pedestrian, including the following steps:

[0089] 4.1) The updated pedestrian motion intention feature representation and the historical latent space feature representation are input into a fusion network for feature fusion. This makes the predicted trajectory more consistent with the actual motion situation and more compatible with the motion patterns of different pedestrians. For pedestrian i, the historical latent space feature representation and the updated motion intention feature representation are first concatenated together as the feature representation before fusion:

[0090]

[0091] In the formula, Let represent the pre-fusion feature representation of pedestrian i before input to the fusion network, and [·||·] represent the vector concatenation operation along the feature dimension. The feature dimension is set to 272.

[0092] 4.2) The obtained pre-fusion feature representations are input into the fusion network for feature fusion. A total of K rounds of feature fusion are performed. The feature representation of the (k+1)th round is obtained by updating the feature representation of the kth round. The iterative process is as follows:

[0093]

[0094] In the formula, Let i be the feature representation of pedestrian i in the (k+1)th round. Let i be the feature representation of pedestrian i in the k-th round. Let g(·) be the feature representation of pedestrian j in the k-th round, and g(·) be a transform encoder used to process the feature representation of the k-th round. The remapping process consists of three fully connected layers with feature dimensions of 256, 128, and 64, used to remap the features from the previous round. Perform remapping, f s (·,·) is a similarity network used to calculate the feature similarity between pedestrian i and pedestrian j. The result is a scalar, and the calculation process is as follows:

[0095]

[0096] In the formula, ReLU(·) represents the ReLU activation function, which maps the result to a value between [0,1]. θ(·) and φ(·) are both transform encoders consisting of three fully connected layers with feature dimensions of 256, 128, and 64, respectively, mapping the features of pedestrian i and pedestrian j to the learnable latent space. [·||·] represents the vector concatenation operation along the feature dimension direction. This represents a single fully connected layer used to map the concatenated feature representations to a scalar.

[0097] 4.3) After K rounds of iteration, the final fused feature representation is obtained. It integrates both its own historical motion information and the motion intention information of other pedestrians in the surrounding environment, and finally represents the fused features. The input is fed into a predictor consisting of three fully connected layers with feature dimensions of 1024, 512 and 256 respectively. Finally, it passes through a fully connected layer to map the feature dimension to 24, that is, 12*2 represents the prediction of the trajectory for the next 12 frames.

[0098] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for predicting pedestrian trajectories based on motion intent extraction, characterized in that, Includes the following steps: 1) Use the historical feature encoder and the endpoint feature encoder to map the pedestrian's historical trajectory and endpoint position to the latent space to obtain the corresponding historical latent space feature representation and endpoint latent space feature representation; 2) The historical latent space feature representation and the endpoint latent space feature representation are concatenated together and input into the conditional variational autoencoder to predict the conditional probability distribution of the pedestrian's motion intention, and the output is the feature representation of the pedestrian's motion intention. 3) Input the pedestrian's motion intention feature representation into a graph neural network, divide pedestrians into different groups according to distance and time, learn the interaction relationship of motion intention between pedestrians within and between groups based on the graph neural network, and update the pedestrian's motion intention feature representation, including the following steps: 3.1) Input the pedestrian's motion intention feature into the graph neural network, and divide the pedestrians into different groups according to distance and time, and represent pedestrian i and pedestrian j as nodes. and With undirected edges By making connections, the pedestrians within the group construct a fully connected graph, represented as follows: ； ； ； In the formula, Let represent the b-th group, which is the b-th fully connected graph. There are a total of B fully connected graphs, each of which is a set of nodes and edges within a graph. Let represent the set of indices of the nodes contained in the b-th fully connected graph. It is the set of nodes in the b-th fully connected graph. It is the set of edges of the b-th fully connected graph; a mask is used here. To indicate whether pedestrian i and pedestrian j are in the same group, it is represented as: ； In the formula, This indicates the number of steps in the historical timeline. This indicates whether pedestrian i and pedestrian j are in the same group. This represents pedestrian i at time step t. Let j be a pedestrian at time step s. If the minimum of the 2-norms of the two is... Greater than the set threshold ,but A value of 0 indicates that pedestrian i and pedestrian j are not in the same group; otherwise... A value of 1 indicates that the two are in the same group; 3.2) Based on the self-attention network and graph message passing mechanism, the movement intentions of pedestrians in the same group are updated to learn the mutual influence of movement intentions among different pedestrians in the same group. The updated movement intention of pedestrian i is: ； In the formula, This represents the updated motion intention feature of pedestrian i. It is a non-linear activation function used to enhance the non-linear expressive power of the network; This indicates a weighted summation of the motion intention features of other pedestrians j within the same group. Let i represent the set of other pedestrians in the same group as pedestrian i. This is the weight matrix of the graph neural network, used to learn the interaction relationships of motion intentions between different pedestrians within a group; The influence weight of pedestrian j relative to pedestrian i is obtained through adaptive learning via a self-attention network; 4) The updated pedestrian motion intention feature representation is concatenated with the historical latent space feature representation and input into the fusion network for feature fusion. Finally, the fused feature representation is input into the predictor to obtain the final future trajectory of the pedestrian.

2. The pedestrian trajectory prediction method based on motion intent extraction according to claim 1, characterized in that, In step 1), the pedestrian's historical trajectory refers to the pedestrian's 2D position coordinate sequence, represented as: ； In the formula, It refers to the first The sequence of movement trajectory points of N pedestrians. express It is a 2-dimensional vector. Represents a 2-dimensional vector. and Let be the X-axis and Y-axis components of pedestrian i at time t in the pixel coordinate system, with a total of T time steps. These T time steps are divided into two parts: historical time steps and future time steps. The historical trajectory of pedestrian i is the set of trajectory points from the historical time steps, and the future trajectory is the set of trajectory points from the future time steps. Therefore, the historical trajectory and future trajectory of pedestrian i are represented as follows: ； ； In the formula, This represents the set of trajectory points of pedestrian i over historical time steps. express It is a 2-dimensional vector. and Is it a pedestrian i in The X-axis and Y-axis components in the pixel coordinate system at time t have a total of A historical step; from arrive Indicates a future time step. This represents the set of trajectory points for pedestrian i at future time steps. express It is a 2-dimensional vector. and Is it a pedestrian i in The X-axis and Y-axis components in the pixel coordinate system at time t have a total of A future time step; pedestrian The endpoint of the motion is taken as the trajectory point of the last step in the future time step, and the endpoint is represented as: ； In the formula, This indicates the final position of pedestrian i's movement. This indicates the last time step in the future. and Is it a pedestrian i in The X and Y components of the pedestrian image in the pixel coordinate system at any given time; using two independent multilayer perceptrons to image the pedestrian image. The historical trajectory and endpoint position are mapped to the latent space, resulting in the corresponding historical latent space feature representation and endpoint latent space feature representation, as follows: ； ； In the formula, and These are the historical latent space feature representations and the destination latent space feature representations for pedestrian i, respectively. and These represent the historical feature encoder and the endpoint feature encoder, respectively, both of which are composed of multiple fully connected layers.

3. The pedestrian trajectory prediction method based on motion intent extraction according to claim 2, characterized in that, Step 2) includes the following steps: 2.1) Concatenate the historical latent space feature representation of pedestrian i with the endpoint latent space feature representation to obtain the motion information feature representation, as follows: ； In the formula, Let i be the motion information feature representation of pedestrian i. This represents a concatenation operation, which concatenates two vectors together along the feature dimension. 2.2) The motion information feature representation of the pedestrian is input into a conditional variational autoencoder, which consists of an encoder and a decoder, both of which are composed of multiple fully connected layers; Input to encoder to obtain latent variables posterior probability distribution Here are hidden variables It is a random variable representing the pedestrian's final destination; then, for the latent variables... The samples are then input into the decoder along with the historical latent space feature representation to obtain the conditional probability distribution of the pedestrian's movement intention. Finally, by solving the marginal distribution of the conditional probability distribution of pedestrian movement intention, we can obtain the conditional probability distribution of pedestrian i's movement intention based on historical movement information, which is expressed as: ； In the formula, This represents the predicted movement intention feature of pedestrian i. The conditional probability distribution representing the movement intention of pedestrian i based on historical movement information is used. The decoder will ultimately output a feature vector, which represents the movement intention feature of pedestrian i. .

4. The pedestrian trajectory prediction method based on motion intent extraction according to claim 3, characterized in that, Step 4) includes the following steps: 4.1) The updated pedestrian motion intent feature representation and the historical latent space feature representation are input into a fusion network for feature fusion; for pedestrian i, the historical latent space feature representation and the updated motion intent feature representation are first concatenated together as the feature representation before fusion: ； In the formula, This represents the pre-fusion feature representation of pedestrian i before it is input into the fusion network. This represents a vector concatenation operation along the feature dimension. 4.2) The obtained pre-fusion feature representations are input into the fusion network for feature fusion. A total of K rounds of feature fusion are performed. The feature representation of the (k+1)th round is obtained by updating the feature representation of the kth round. The iterative process is as follows: ； In the formula, Let i be the feature representation of pedestrian i in the (k+1)th round. Let i be the feature representation of pedestrian i in the k-th round. Let j be the feature representation of pedestrian j in the kth round. It is a transform encoder used for feature representation in the k-th round. Perform remapping. This is a similarity network used to calculate the feature similarity between pedestrian i and pedestrian j. The result is a scalar, and the calculation process is as follows: ； In the formula, This represents the ReLU activation function, used to map the result to a value between [0,1]. and Both are transform encoders consisting of three fully connected layers, which map the feature representations of pedestrian i and pedestrian j in the k-th round to a learnable latent space, respectively. This represents a vector concatenation operation along the feature dimension. This represents a single fully connected layer used to map the concatenated feature representations to a scalar. 4.3) After K rounds of iteration, the final fused feature representation is obtained. This integrates both the historical information of its own movement and the movement intention information of other pedestrians in the surrounding environment, ultimately resulting in a fused feature representation. The input is fed into a predictor consisting of three fully connected layers. Finally, it passes through a fully connected layer to map the feature dimension to the target prediction time step length, thus obtaining the prediction of the pedestrian's future trajectory.

Images