Intelligent agent action prediction method based on multi-scale spatial perception

Abstract

The invention discloses an agent action prediction method based on multi-scale spatial perception, which comprises the steps of map spatial structure modeling, historical trajectory feature extraction, feature fusion and multi-modal action prediction. According to the map space structure, map feature extraction is conducted on two-dimensional vector map data in an application scene map through a multi-scale map convolutional neural network, and high-dimensional map feature information is obtained; the historical trajectory feature extraction adopts a convolutional neural network and a feature pyramid network to extract high-dimensional trajectory data features of all agents; according to feature fusion, correlation modeling fusion is carried out on high-dimensional map feature information and high-dimensional trajectory data feature information through a self-attention mechanism, and agent trajectory fusion features with direction information are obtained; according to the multi-modal trajectory prediction, intelligent agent trajectory fusion features are used for prediction regression and confidence scoring, multi-modal complete trajectory coordinates of action prediction and corresponding confidence scores are output, and a reasonable auxiliary decision is provided for intelligent agent actions.

Description

Technical Field

[0001] The present invention relates to the field of autonomous action prediction and intelligent behavior decision-making, and in particular to an intelligent agent action prediction method based on multi-scale spatial perception. Background Art

[0002] The construction of the intelligent model focuses on the physical model, perception model, behavior model, decision model and control model. The behavior model is the driving force that drives the physical model to effectively achieve specific goals. It is based on the extraction of perception features to form the data source and basis for decision-making and control. At present, the design focus of the behavior model is still based on dynamics, rules and expert experience data for prediction. The rule method for specific tasks requires huge manpower and resources to design, and cannot effectively resist knowledge blind spots and achieve innovative exploration, and is prone to fall into the dilemma of local suboptimality.

[0003] The current mainstream action prediction provides important geometric information for action prediction based on rasterized map data features, extracts spatial features formed by the relative positions of nodes in the map scene through convolutional neural networks (CNN), and extracts useful background information for motion prediction. Based on recurrent neural networks (RNNs), such as long short-term memory (LSTM), etc., by extracting the temporal features of the agent's historical trajectory data, linear or nonlinear methods are used to fit the temporal relationship of autonomous actions to generate future trajectory coordinates. This type of processing method inevitably loses some key information of high-definition maps and ignores the interactive relationship between historical trajectories with the same temporal features in the same scene field of view. The trajectory prediction method based on graph neural networks currently only considers the node interaction relationship between action entities, and does not integrate the complex and numerous map element node information in the map.

[0004] Therefore, the existing rule-based action prediction methods have technical problems such as complex design and lack of multi-task versatility, and the various types of neural network-related prediction methods currently available cannot meet the needs of covering all-factor interaction relationships in complex map environments. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent agent action prediction method based on multi-scale spatial perception, which realizes accurate and reasonable autonomous action by fully combining the powerful semantic representation ability of graph neural networks, and provides reliable and credible multimodal action prediction for intelligent agent simulation based on full-factor information fusion.

[0006] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0007] A method for predicting intelligent agent actions based on multi-scale spatial perception, including: map spatial structure modeling, historical trajectory feature extraction, feature fusion and multi-modal action prediction, wherein:

[0008] The map spatial structure uses a multi-scale graph convolutional neural network to extract map features from two-dimensional vector map data in the application scenario map to obtain high-dimensional map feature information;

[0009] The historical trajectory feature extraction uses a convolutional neural network and a feature pyramid network to extract high-dimensional trajectory data features of all intelligent agents in the application scenario map;

[0010] The feature fusion performs correlation modeling and fusion on the high-dimensional map feature information and high-dimensional trajectory data feature information obtained by map spatial structure modeling and historical trajectory feature extraction through a self-attention mechanism, so as to obtain the intelligent body trajectory fusion feature with direction information;

[0011] The multimodal trajectory prediction uses the intelligent agent trajectory fusion features to perform prediction regression and confidence scoring, outputs the multimodal complete trajectory coordinates and corresponding confidence scores of the action prediction, and provides auxiliary decision-making for the intelligent agent action.

[0012] Furthermore, the application scenario map should be able to provide map data in a two-dimensional bird's-eye view vector coordinate format, the map data containing historical behavior trajectory data of all intelligent agents in the application scenario map, and the lane centerline having lane coordinate information, including the initial and end position coordinates of the lane.

[0013] Furthermore, the map space structure uses a multi-scale graph convolutional neural network to extract map features from two-dimensional vector map data in the application scenario to obtain high-dimensional map feature information, including:

[0014] 91-1) Use the MLP network to map the lane coordinate information, expressed as Among them l i represents the i-th lane node modeled, and Respectively represent the initial and end position coordinates of the center line of the i-th lane in the map data, v i Indicates the position of the center point of the i-th lane; MLP shape Indicates that the MLP network is used to extract the features of the lane node length; MLP loc Indicates that the central position features of lane nodes are extracted using the MPL network; all l in the application scenario map i Construct high-dimensional map data L, where l i is the i-th row of L;

[0015] (1-2) Based on the high-dimensional map data L, the GCN network is used to learn the association information between the nodes of the high-dimensional map data; wherein, the GCN network uses the GCN network to process the forward, backward, left neighbor and right neighbor according to the four connection types of lane nodes, expressed as: F = LW 0 +∑ j∈{q，h，z，y} A j LW j , where W 0 is the original information weight of the lane node, which is set to the unit matrix; j represents the th connection type, q, h, z, y represent the physical connection relationship of the lane node's forward, backward, left neighbor, and right neighbor in the map respectively, and A j and W j Represent the adjacency matrix and weight matrix of the jth connection type respectively;

[0016] (1-3) Select whether the lane node has ground markings as an additional weight for A j Restriction processing is performed, specifically expressed as follows: We get F = LW 0 +∑ j∈{q，h，z，y} A′ j LW j ; Wherein, I is the identification matrix related to ground signs and markings, that is, the connection relationship matrix of lane nodes without ground signs and markings;

[0017] (1-4) According to step (1-3), the selected A′ is obtained j After that, the final high-dimensional map feature information is expressed as: Where C is the number of expansion scale factors, k c is the cth scale factor.

[0018] Furthermore, the historical trajectory feature extraction uses a convolutional neural network and a feature pyramid network to extract high-dimensional trajectory data features of all agents in the application scenario map, including:

[0019] The CNN network is used to extract features from the historical behavior trajectory data of all intelligent agents in the scene to obtain the historical trajectory features T. At the same time, the feature pyramid FPN network is used to fuse the feature values ​​of different scales in the historical trajectory features T to output high-dimensional trajectory data features T′ with time series characteristics.

[0020] Furthermore, T′ is obtained by concatenating different node information in the historical trajectory feature T, which is expressed as T′=concat(t mi , Δ mi,mj , t mj ), where t mi is the feature information of the mi-th historical trajectory node in T feature variables, Δmi,mj =MLP(t mj -t mi ) indicates that MLP encoding is used to encode different historical trajectory nodes t mj ,t mi The process of extracting the position difference feature information is then performed through concat to obtain the high-dimensional trajectory data feature T′.

[0021] Furthermore, the feature fusion performs correlation modeling and fusion on the high-dimensional map feature information and high-dimensional trajectory data feature information obtained by map spatial structure modeling and historical trajectory feature extraction through a self-attention mechanism to obtain the intelligent agent trajectory fusion feature with direction information, including:

[0022] (3-1) The historical trajectory fusion information T′ obtained by extracting historical trajectory features is fused through the self-attention mechanism. The calculation process is expressed as: in d is the dimension of the key vector K, The representation layer is normalized and ReLU operation, and softmax represents the normalized exponential function; the query, key, and value vectors of the self-attention mechanism are represented as Q, K, and V, respectively, and the high-dimensional trajectory data feature T′ is obtained through three different linear layers, which is expressed as: Q = T′W Q , K = T′W K , V = T′W V ; Among them, W Q , W K , W V Represent the weight matrices of Q, K, and V respectively;

[0023] (3-2) Historical trajectory feature information The high-dimensional map feature information F is further integrated through the cross-attention mechanism, expressed as: where <-, ·> represents the inner product operation, i.e., the weight matrix of the cross attention operator is obtained by adding the query vector of the attention mechanism and the key vector K F The multiplication is then calculated by the value vector V F The weighting is performed to output the agent trajectory fusion feature Z with direction information.

[0024] Furthermore, the multimodal trajectory prediction uses the intelligent agent trajectory fusion features to perform prediction regression and confidence scoring, outputs the multimodal complete trajectory coordinates and corresponding confidence scores of the action prediction, and provides reasonable auxiliary decision-making for the intelligent agent action, including:

[0025] (4-1) The extracted agent trajectory fusion feature Z is regressed through the first MLP branch network to form the multimodal complete trajectory BEV coordinates, which is expressed as represents the T-th step BEV coordinate of the k-th action of the m-th agent, K is a configurable parameter, representing the prediction of K action trajectories;

[0026] (4-2) The multimodal complete trajectory BEV coordinates P obtained using (4-1) m，reg The feature Z is fused with the agent trajectory and the confidence score of K action predictions is formed through the second MLP branch network, which is represented by O m，cls =(c m，0 , C m，1 , …, c m，K-1 ), where C m，K-1 Represents the confidence score value of the K-th behavior trajectory prediction of the m-th agent.

[0027] Furthermore, in the MLP training process of multimodal trajectory prediction, according to the multimodal complete trajectory BEV coordinates P m，reg and confidence score O m，cls , combined with the true value of the trajectory of each agent in the existing data set, the classification loss and regression loss in the loss function use the maximum edge loss and smooth L1 loss respectively, that is, L = L cls +L reg ,in The loss function trains the model; specifically, M is the number of agents, ∈ represents the boundary value of the loss function, which is 0.2; is the true value of the BEV coordinate of agent m at step t, reg(x) = ∑ i d(x i ), x i is the i-th element of x, d(x i ) is the smooth L1 loss.

[0028] An action prediction device comprises a processor, a memory and a computer program stored in the memory; when the processor executes the computer program, the intelligent agent action prediction method based on multi-scale spatial perception is implemented.

[0029] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the intelligent agent action prediction method based on multi-scale spatial perception is implemented.

[0030] Compared with the prior art, the present invention has the following technical features:

[0031] 1. The present invention constructs a multi-scale graph neural network for feature extraction of map element information with long-distance dependencies and directional information, overcoming the problem of unacceptable and unexplainable predictions caused by the lack of highly correlated map node semantic information in existing action prediction methods.

[0032] 2. The present invention constructs a data fusion network to capture the complete information interaction between map features and intelligent agents, specifically including the relevant fusion of all-factor node data such as target intelligent agent information, dynamic historical information, static environmental information, etc., to achieve effective modeling of spatial topological structure for multimodal action prediction. On the basis of complete spatial information, it can obtain K (K is a configurable parameter) action trajectory predictions that are explainable, acceptable and closer to real empirical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the action prediction framework diagram of the present invention;

[0034] Figure 2 This is a network architecture diagram of the map space structure modeling module of the present invention;

[0035] Figure 3 This is a network architecture diagram of the historical trajectory feature extraction module of the present invention;

[0036] Figure 4 A network architecture diagram of feature fusion of the present invention;

[0037] Figure 5 A network architecture diagram of the multimodal action prediction of the present invention;

[0038] Figure 6 It is a schematic diagram of the application process of an embodiment of the present invention;

[0039] Figure 7 This is a visualization diagram of the simulation results of the present invention. DETAILED DESCRIPTION

[0040] The present invention provides an intelligent agent action prediction method based on multi-scale spatial perception, and the inventive method can be directly applied to application scenarios with two-dimensional vector map data. In order to be more general, the scheme of the present invention is implemented using the open source Argoverse dataset, which can provide high-precision and diverse autonomous driving scenes and visual data. The Argoverse dataset can provide lane-level precision map information, including center lines, intersections, and crosswalks, etc. Its motion prediction dataset contains about 320,000 scene sequences. These scenes are all set with a main intelligent agent, and are shot following the trajectory perspective of the main intelligent agent. The total duration is 5 seconds, and the sampling rate is 10Hz, that is, 10 frames per second. The task of motion prediction is to predict the motion trajectory of the last 3 seconds of the given main intelligent agent given the motion information and high-precision map information of the previous two seconds; trajectory prediction involves complex interactions between different intelligent agents, and is generally shot in places with complex traffic conditions. The size of the data set can meet the training of complex models.

[0041] like Figure 1 As shown, the present invention provides an agent action prediction method based on multi-scale spatial perception, including map space structure modeling, historical trajectory feature extraction, feature fusion and multi-modal action prediction, wherein:

[0042] The map spatial structure is constructed by using a multi-scale graph convolutional neural network (GCN) to extract map features from two-dimensional vector map data in the application scenario to obtain high-dimensional map feature information;

[0043] The historical trajectory feature extraction uses a convolutional neural network (CNN) and a feature pyramid (FPN) network with high parallel computing efficiency to extract high-dimensional trajectory data features of all intelligent agents in the scene;

[0044] The feature fusion performs correlation modeling and fusion on the high-dimensional map feature information and high-dimensional trajectory data feature information obtained by map spatial structure modeling and historical trajectory feature extraction through a self-attention mechanism, and obtains the intelligent agent trajectory fusion feature with direction information;

[0045] The multimodal trajectory prediction uses the intelligent agent trajectory fusion features to perform prediction regression and confidence scoring, outputs the multimodal complete trajectory BEV coordinates and corresponding confidence scores of the action prediction, and provides reasonable auxiliary decision-making for the intelligent agent action.

[0046] 1. Map spatial structure modeling

[0047] The process of constructing the map space structure mainly includes: using a multi-layer perceptron (MLP) network layer, a multi-scale graph convolutional neural network (GCN) and a linear layer to model lane nodes and ground sign information to obtain high-dimensional map feature information. The network structure is as follows: Figure 2shown.

[0048] (1-1) First, the MLP network layer is used to pre-process the attributes of the lane centerline, intersection, traffic signs, etc. provided by the application scenario map; the application scenario map should be able to provide map data in a two-dimensional bird's-eye view (BEV) vector coordinate format, and the map data contains the historical behavior trajectory data of all intelligent agents in the application scenario map. The lane centerline has lane coordinate information, including the initial and end position coordinates of the lane, expressed as v start (x, y) and v end (x, y); the lane coordinate information is mapped using the MLP network, expressed as

[0049] Among them, l i represents the i-th lane node modeled, and Respectively represent the initial and end position coordinates of the center line of the i-th lane in the map data, v i represents the position of the center point of the i-th lane, which is calculated by the mean of the initial and final position coordinates of the lane centerline. i Construct high-dimensional map data L, where l i is the i-th row of L. MLP shape Indicates that the MLP network is used to extract the features of the lane node length; MLP loc It means that the central position features of lane nodes are extracted using the MPL network.

[0050] (1-2) Based on the high-dimensional map data L obtained in step (1-1), the GCN network is used to learn the association information between the nodes of the high-dimensional map data; wherein the GCN network uses the GCN network to process the forward, backward, left neighbor and right neighbor respectively according to the four connection types of lane nodes, which is expressed as: F = LW 0 +∑ j∈{q，h，z，y} A j LW j , where W 0 is the original information weight of the lane node, which is generally set to the unit matrix; j represents the jth connection type, q, h, z, y represent the physical connection relationship of the lane node's forward, backward, left neighbor, and right neighbor in the map, respectively, and A j and W j Represent the adjacency matrix and weight matrix of the jth connection type respectively; where A q and A h The left and right neighbors are obtained by moving the unit matrix one step to the upper right and lower left. The left and right neighbors refer to the node connection relationship that can change lanes with each other without violating traffic rules, which is expressed as A z and A y, by setting the position of the node that can cross lanes to 1, the connection type and direction of the lane nodes are encoded by the above GCN network structure modeling.

[0051] (1-3) According to the lane node connection type and direction obtained in step (1-2), this scheme selects the ground sign and line type coupled with F as information extension, and performs lane node adjacency matrix A i Perform constraint design.

[0052] In fact, the selection of favorable coupling information is related to the characteristics of the dataset and the target task. In order to preserve generality, this scheme selects whether the lane node has ground signs and markings as an additional weight for A. j Restriction processing is performed to achieve the encoding of more geometric information, which is specifically expressed as follows: Perform a preliminary screening of the connection relationship; obtain F = LW 0 +∑ j∈{q，h，z，y} A′ j LW j ; Wherein, I is the identification matrix related to ground signs and markings, that is, the connection relationship matrix of lane nodes without ground signs and markings.

[0053] (1-4) According to step (1-3), the selected A′ is obtained j Finally, in order to further capture the long-distance forward dependency relationship between lane nodes to meet the action prediction with time span, this scheme expands the lane information of the forward relationship; the final high-dimensional map feature information is expressed as:

[0054] Where C is the expansion scale factor, which can be flexibly set according to the task. In this scheme, it is set to 5; k c is the cth scale factor. The scale factor of the present invention adopts k c =2 c , through the scale factor, the selected adjacency matrix A′ j According to k c Perform matrix power operations; for example, when c = 5, k c =2 5 =32, W j＝q，c That is W q or W c .

[0055] The present invention aggregates the forward lane node information of longer distance into the current node through multi-scale operation technology, and the larger the scale factor (such as the lane node with a scale factor of 5 as mentioned above), the smaller the aggregation effect, and then outputs high-dimensional map feature information F through a linear layer, with a dimension of N×128, where N is the number of lane nodes in the scene.

[0056] 2. Historical trajectory feature extraction

[0057] Historical trajectory feature extraction mainly uses convolutional neural network (CNN) and feature pyramid (FPN) network to extract high-dimensional trajectory data features of all agents in the scene. The network structure is as follows: Figure 3 shown.

[0058] Firstly, a CNN network with high parallel computing efficiency is used to extract features from the historical behavior trajectory data of all agents in the scene; the CNN network uses three groups of one-dimensional convolutional networks with a convolution kernel size of 3. Each group of convolutions consists of two residual networks, the first residual block has a stride of 2, and the second residual block has a stride of 1; after extraction by the CNN network, the historical trajectory feature T is obtained, and the feature pyramid (FPN) network is used to fuse the feature values ​​of different scales in the historical trajectory feature T, and the high-dimensional trajectory data feature T′ with a dimension of 128 and time series characteristics is output, which is:

[0059] T′ is obtained by concatenating different node information in the historical trajectory feature T, expressed as T′=concat(t mi , Δ mi,mj , t mj ), where t mi is the feature information of the mi-th historical trajectory node in T feature variables, Δ mi,mj =MLP(t mj -t mi ) indicates that MLP encoding is used to encode different historical trajectory nodes t mj ,t mi The process of extracting the position difference feature information is then performed through concat to obtain the high-dimensional trajectory data feature T′.

[0060] In the feature extraction process, the historical action trajectory data of different intelligent agents in the scene are padded with 0 data to obtain point displacement coordinate data with the same duration (taking 2 seconds as an example), and then the same number of 0 / 1 masks are connected after the displacement coordinate data to indicate whether the coordinate point is filled data. Thus, a historical behavior trajectory tensor T with a dimension of M×20×3 is formed for the next feature extraction, where M is the number of all intelligent agents in the scene, 20 represents the number of trajectory collection points within 2 seconds, and 3 represents the (x, y) coordinates and mask data of each point displacement coordinate.

[0061] By extracting historical trajectory features, a high-dimensional trajectory data feature T with time characteristics is output, with a dimension of 128, that is, the output dimension is M×20×128.

[0062] 3. Feature Fusion

[0063] Feature fusion encodes the trajectory position difference information through the MLP network layer, then fuses the historical trajectory through the self-attention network, and finally fuses the extracted vector map high-dimensional feature information F with the high-dimensional trajectory data feature T′ through the cross-attention module to obtain the intelligent agent trajectory fusion feature Z; the network structure is as follows Figure 4 shown.

[0064] (3-1) The historical trajectory fusion information T′ obtained by extracting historical trajectory features is fused through the self-attention mechanism. The calculation process is expressed as: in d is the dimension of the key vector K. Dividing by the square root of d can prevent the inner product from being too large. Represents layer normalization and ReLU operation, softmax represents the normalized exponential function; the self-attention module outputs the fused historical trajectory feature information

[0065] The query, key, and value vectors of the self-attention mechanism are denoted as Q, K, and V, respectively, and are obtained by using the high-dimensional trajectory data feature T′ through three different linear layers, expressed as: Q = T′W Q , K = T′W K , V = T′W V ; Among them, W Q , W K , W V They represent the weight matrices of Q, K, and V respectively, and their initial values ​​are generally obtained by random initialization.

[0066] (3-2) The fused historical trajectory feature information obtained in step (3-1) The high-dimensional map feature information F obtained in step 1 is further fused through the cross-attention mechanism, which can encode the relative position information of different historical trajectory nodes and lane nodes. The key point is to uniformly encode the two types of node information, trajectory and map, through the weight matrix, and to shield the unrelated node information through the mask attention mechanism. The specific calculation process is expressed as:

[0067] where <·, ·> represents the inner product operation, i.e., the weight matrix of the cross attention operator is obtained by adding the query vector of the attention mechanism and the key vector K F The multiplication is then calculated by the value vector V F The weighted output is the agent trajectory fusion feature Z with direction information, with a dimension of M×128. The query, key, and value vectors of the cross attention mechanism are respectively K F =FW K , V F =FWV get.

[0068] 4. Multimodal trajectory prediction

[0069] The multimodal trajectory prediction module uses the agent trajectory fusion feature Z obtained in step 3 to perform trajectory prediction regression and confidence scoring, and outputs the predicted multimodal complete trajectory BEV coordinate information and scoring information to provide reasonable auxiliary decision-making for the agent's actions; that is, the network finally outputs the predicted agent coordinates and the corresponding scores. In actual application, the prediction results with higher confidence scores can be selected as decision-making references.

[0070] The multimodal trajectory prediction is composed of two trained MLP branch architectures, each of which is composed of two groups of linear layers, group normalization and ReLU modules; the first MLP branch network is used to output the multimodal complete trajectory BEV coordinates of action prediction, with input and output dimensions of M×128 and M×K×60 respectively, and the other MLP branch network performs confidence scoring for the action prediction results, with input and output dimensions of M×128 and M×K×1 respectively, and performs confidence scoring for the action prediction results under each modality, K is a configurable parameter, representing K action trajectory predictions, and the network structure is as follows: Figure 5 shown.

[0071] (4-1) The extracted agent trajectory fusion feature Z is regressed through the first MLP branch network to form the multimodal complete trajectory BEV coordinates, which is expressed as represents the BEV coordinates of the Tth step of the kth action of the mth agent. T is generally 3 seconds. Since the sampling frequency is 10 Hz, the predicted coordinate value of each trajectory is 30 here. The coordinate value is expressed in two-dimensional coordinates (x, y), so the output is the 60-dimensional multimodal complete trajectory BEV coordinate P m，reg ;

[0072] (4-2) The multimodal complete trajectory BEV coordinates P obtained using (4-1) m，reg The feature Z is fused with the agent trajectory and the confidence score of K action predictions is formed through the second MLP branch network, which is represented by O m，cls =(c m，0 , C m，1 , …, c m，K-1 ), where C m，K-1 Represents the confidence score value of the K-th behavior trajectory prediction of the m-th agent.

[0073] In the MLP training process of multimodal trajectory prediction, the multimodal complete trajectory BEV coordinates P obtained in step (4-1) and step (4-2) are m，reg and confidence score Om，cls , combined with the true value of the trajectory of each agent in the existing data set, the classification loss and regression loss in the loss function use the maximum edge loss and smooth L1 loss respectively, that is, L = L cls +L reg ,in The loss function trains the model; specifically, M is the number of agents, ∈ represents the boundary value of the loss function, which is a non-negative real number, and the value in this embodiment is 0.2; is the true value of the BEV coordinate of agent m at step t, reg(x) = ∑ i d(x i ), x i is the fth element of x, d(x i ) is the smooth L1 loss, defined as:

[0074] Based on map data and scene data, the present invention performs end-to-end model training on multi-scale GCN networks, CNN networks, feature fusion networks and multimodal trajectory prediction networks, so the trained model framework is convenient for algorithm migration and application. In order to be more general, the data source format required by the method of the present invention is simple and does not require specific preprocessing. The deployment and application of the model can be realized by focusing on combining two-dimensional vector map data and trajectory data in the scene.

[0075] Figure 6 A practical application flow chart of the present invention is given. After loading the data set or the map data in the actual scene, the method generally selects the main intelligent agent as the research object, collects the historical trajectory data with time series characteristics of all intelligent agents in the scene and the map vector data in the current scene; uses the CNN network to extract the historical behavior trajectory features, uses the multi-scale GCN to perform feature learning on the vector map data, and fuses the learned historical behavior trajectory features and map element features through a fusion network with the attention mechanism as the core. The fused feature information is input into the action prediction network to perform multimodal prediction and evaluation on the future action trajectory of the intelligent agent, and the number of predicted trajectories can be flexibly set according to the application scenario.

[0076] The intelligent agent action prediction method based on multi-scale spatial perception provided by the present invention extracts features from the historical trajectory of the intelligent agent in the target simulation data and the scene map corresponding to different network structures respectively. The map data feature extraction directly uses the vector map data form, and does not require rasterization processing, thereby avoiding the loss of map element information in this scene, and can establish a structured map element graph based on the complete vector map data. When a multi-scale GCN network is used to aggregate information of multiple forward lane nodes, the characteristics of large displacement of intelligent agents with fast modeling speed can be accurately modeled, and feature extraction of lane nodes with long distance dependencies can be achieved. The trajectory prediction output of the intelligent agent is further updated, and the confidence score is calculated based on the updated trajectory data, so that the intelligent agent can realize multi-modal behavior strategy exploration and improve the exploration efficiency of the intelligent agent.

[0077] Example:

[0078] See attached Figure 6 This embodiment uses the public Argoverse trajectory prediction benchmark to verify and visualize this technology. The Argoverse dataset contains 324,557 scenes, each scene lasts 5 seconds, the sampling rate is 10 Hz, and the trajectory of each tracked object is represented by its centroid trajectory. The first 2 seconds of data in the training set and validation set in the dataset are used as input data, and the last 3 seconds of data are used as the model prediction true value. The map file provides the relevant BEV vector coordinate map data within the 5 seconds scene, which can meet the map data collection application of the present invention.

[0079] Firstly, the original data of the two-dimensional initial and ending positions (BEV) of the lane centerline and ground sign elements provided in the Argoverse dataset map file are used to model the lane nodes and ground sign information through a multi-layer perceptron (MLP) network layer, a multi-scale graph convolutional neural network (GCN) and a linear layer to obtain high-dimensional map feature information with a dimension of N×128, where N is the number of lane nodes in the scene.

[0080] The first 2s action data of all agents in the scene in the Argoverse dataset is used as the historical trajectory input data, and the data that does not meet the 2s time requirement is filled. The present invention adopts the method of 0 data filling to obtain 20 point displacement coordinate data with the same duration-sampling frequency of 10Hz. The same number of 0 / 1 masks are connected after the displacement coordinate data to indicate whether the coordinate point is filled data, thereby forming a historical behavior trajectory tensor with a dimension of M×20×3 for the next feature extraction, where the dimension M is the number of all agents in the scene, 20 represents the number of trajectory collection points within 2s for each agent, and 3 represents the x, y coordinates and mask data of each displacement point coordinate. Through the historical trajectory feature extraction network designed by the present invention, historical trajectory data with time series characteristics is output, and the dimension is 128, that is, the output historical trajectory data M×20×128.

[0081] Based on the above high-dimensional map and historical motion trajectory feature information, the correlation information is fused and modeled through the attention mechanism to obtain the position coordinates and driving direction information related to the intelligent body trajectory. Finally, the fused high-dimensional features are used for predictive regression and confidence scoring, and the multimodal complete trajectory coordinates and scoring information of the main intelligent body are output to provide reasonable auxiliary decision-making for the intelligent body's actions.

[0082] from Figure 7 From the visualization of the simulation results of this scheme, it can be seen that the data set scenes selected by this scheme are mostly at intersections, where vehicles turn left or right, change lanes, and other places with complex traffic. This provides more complex training conditions for the action prediction of the intelligent agent, so that the model can combine the map element feature information to reliably predict the trajectory in scenes such as complex road conditions and real-time obstacles. The present invention adopts an end-to-end model training mechanism, which is easy to deploy and apply in a variety of simulation platforms. The map data format corresponding to the algorithm model is simple and does not require specific preprocessing. It can achieve trajectory prediction modeling with the support of basic data such as historical behavior trajectory coordinates with time series characteristics, vector lane node center lines, and ground signs, and output flexibly configurable multi-modal behavior trajectories and evaluation results. The evaluation results of the corresponding relevant trajectories can provide a good basis for the evaluation of the autonomous behavior capabilities of intelligent models in industrial applications, and have good application advantages.

[0083] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for predicting intelligent agent actions based on multi-scale spatial perception, characterized in that: Includes: map spatial structure modeling, historical trajectory feature extraction, feature fusion, and multimodal action prediction, including: The map spatial structure uses a multi-scale graph convolutional neural network to extract map features from two-dimensional vector map data in the application scenario map to obtain high-dimensional map feature information; The historical trajectory feature extraction uses a convolutional neural network and a feature pyramid network to extract high-dimensional trajectory data features of all intelligent agents in the application scenario map; The feature fusion performs correlation modeling and fusion on the high-dimensional map feature information and high-dimensional trajectory data feature information obtained by map spatial structure modeling and historical trajectory feature extraction through a self-attention mechanism, so as to obtain the intelligent body trajectory fusion feature with direction information; The multimodal trajectory prediction uses the intelligent agent trajectory fusion features to perform prediction regression and confidence scoring, outputs the multimodal complete trajectory coordinates of the action prediction and the corresponding confidence score, and provides auxiliary decision-making for the intelligent agent action.

2. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 1, characterized in that: The application scenario map should be able to provide map data in a two-dimensional bird's-eye view vector coordinate format. The map data contains historical behavior trajectory data of all intelligent agents in the application scenario map. The lane centerline has lane coordinate information, including the initial and end position coordinates of the lane.

3. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 1, characterized in that: The map spatial structure uses a multi-scale graph convolutional neural network to extract map features from two-dimensional vector map data in the application scenario to obtain high-dimensional map feature information, including: (1-1) The lane coordinate information is mapped using the MLP network, expressed as Among them l i represents the i-th lane node modeled, and Respectively represent the initial and end position coordinates of the center line of the i-th lane in the map data, v i Indicates the position of the center point of the i-th lane; MLP shape Indicates that the MLP network is used to extract the features of the lane node length; MLP loc Indicates that the central position features of lane nodes are extracted using the MPL network; all l in the application scenario map i Construct high-dimensional map data L, where l i is the i-th row of L; (1-2) Based on the high-dimensional map data L, the GCN network is used to learn the association information between the nodes of the high-dimensional map data; wherein, the GCN network uses the GCN network to process the forward, backward, left neighbor and right neighbor according to the four connection types of lane nodes, which can be expressed as: F = LW0 + ∑ j∈{q,h,z,y} A j LW j , where W0 is the original information weight of the lane node, which is set to the unit matrix; j represents the jth connection type, q, h, z, y represent the physical connection relationship of the lane node's forward, backward, left neighbor, and right neighbor in the map, respectively, and A j and W j Represent the adjacency matrix and weight matrix of the j-th connection type respectively; (1-3) Select whether the lane node has ground markings as an additional weight for A j Restriction processing is performed, specifically expressed as follows: We get F = LW0 + ∑ j∈{q,h,z,y} A′ j LW j ; Wherein, I is the identification matrix related to ground signs and markings, that is, the connection relationship matrix of lane nodes without ground signs and markings; (1-4) According to step (1-3), the selected A′ is obtained j After that, the final high-dimensional map feature information is expressed as: Where C is the number of expansion scale factors, k c is the cth scale factor.

4. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 1, characterized in that: The historical trajectory feature extraction uses a convolutional neural network and a feature pyramid network to extract high-dimensional trajectory data features of all agents in the application scenario map, including: The CNN network is used to extract the features of the historical behavior trajectory data of all intelligent agents in the scene to obtain the historical trajectory features T. At the same time, the feature pyramid FPN network is used to fuse the feature values ​​of different scales in the historical trajectory features T to output high-dimensional trajectory data features T with time series characteristics. ′ .

5. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 4, characterized in that: High-dimensional trajectory data features T ′ The different node information in the historical trajectory feature T is obtained by concatenating, which is expressed as T′=concat(t mi ,Δ mi,mj ,t mj ), where t mi is the feature information of the mi-th historical trajectory node in T feature variables, Δ mi,mj =MLP(t mj -t mi ) indicates that MLP encoding is used to encode different historical trajectory nodes t mj ,t mi The process of extracting the position difference feature information is then performed through concat to obtain the high-dimensional trajectory data feature T ′ .

6. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 1, characterized in that: The feature fusion performs correlation modeling and fusion on the high-dimensional map feature information and high-dimensional trajectory data feature information obtained by map spatial structure modeling and historical trajectory feature extraction through a self-attention mechanism to obtain the intelligent agent trajectory fusion feature with direction information, including: (3-1) The historical trajectory fusion information T obtained by extracting historical trajectory features ′ The feature fusion is performed through the self-attention mechanism, and the calculation process is expressed as: in d is the dimension of the key vector K, represents layer normalization and ReLU operation, and softmax represents the normalized exponential function; the query, key, and value vectors of the self-attention mechanism are represented as Q, K, and V respectively, using the high-dimensional trajectory data feature T ′ After three different linear layers, we get: Q = T′W Q ,K=T′W K ,V=T′W V ; Among them, W Q ,W K ,W V Represent the weight matrices of Q, K, and V respectively; (3-2) Historical trajectory feature information The high-dimensional map feature information F is further integrated through the cross-attention mechanism, expressed as: where <·,·> represents the inner product operation, i.e., the weight matrix of the cross attention operator is obtained by adding the query vector of the attention mechanism and the key vector K F The multiplication is then calculated by the value vector V F The weighting is performed to output the agent trajectory fusion feature Z with direction information.

7. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 1, characterized in that: The multimodal trajectory prediction uses the intelligent agent trajectory fusion features for prediction regression and confidence scoring, outputs the multimodal complete trajectory coordinates and corresponding confidence scores of the action prediction, and provides reasonable auxiliary decision-making for the intelligent agent action, including: (4-1) The extracted agent trajectory fusion feature Z is regressed through the first MLP branch network to form the multimodal complete trajectory BEV coordinates, which is expressed as represents the T-th step BEV coordinate of the k-th action of the m-th agent, K is a configurable parameter, representing the prediction of K action trajectories; (4-2) The multimodal complete trajectory BEV coordinates P obtained using (4-1) m,reg The feature Z is fused with the agent trajectory and the confidence score of K action predictions is formed through the second MLP branch network, which is represented by O m,cls =(c m,0 ,c m,1 ,…,c m,K-1 ), where c m,K-1 Represents the confidence score value of the K-th behavior trajectory prediction of the m-th agent.

8. The method for predicting intelligent agent actions based on multi-scale spatial perception according to claim 1, characterized in that: In the MLP training process of multimodal trajectory prediction, according to the multimodal complete trajectory BEV coordinates P m,reg and confidence score O m,cls , combined with the true value of the trajectory of each agent in the existing data set, the classification loss and regression loss in the loss function use the maximum edge loss and smooth L1 loss respectively, that is, L = L cls +L reg ,in The loss function trains the model; specifically, M is the number of agents, ∈ represents the boundary value of the loss function, which is 0.2; is the true value of the BEV coordinate of agent m at step t, reg(x) = ∑ i d(x i ), x i is the i-th element of x, d(x i ) is the smooth L1 loss.

9. An action prediction device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that: When the processor executes the computer program, it implements the intelligent agent action prediction method based on multi-scale spatial perception according to any one of claims 1-8.

10. A computer-readable storage medium, wherein a computer program is stored in the medium; characterized in that: When the computer program is executed by a processor, the intelligent agent action prediction method based on multi-scale spatial perception according to any one of claims 1-8 is implemented.

Images