Group target intention discrimination method and device based on combination strategy deep learning

Abstract

The present application relates to the technical field of target intention discrimination, and provides a group target intention discrimination method and device based on a combination strategy deep learning, wherein the overall structure and processing flow of a discrimination model are redesigned, cluster division processing is additionally added, a discrimination flow from a local to a whole is realized, the model can process a scene where multiple targets exist at the same time, and an intention discrimination network based on Patching technology and an attention mechanism is designed according to the characteristics of ground target time sequence data, the internal relationship between different characteristic dimensions and time dimensions is captured through base models of different Patching divisions, the base models are fused at a decision-making level through the attention mechanism, the discrimination results of different perspectives are taken into account, the discrimination accuracy and robustness of the model are improved, and the discrimination model can process a group target scene.

Description

Technical Field

[0001] This invention belongs to the field of target intent discrimination technology, and relates to a group target intent discrimination method and device based on deep learning of combined strategies. Background Technology

[0002] Target intent discrimination extracts feature data from the behavior and state changes of the environment and the target, combines it with prior rule knowledge, and extracts and analyzes the inherent relationships between these data to determine the target's intent. Intent discrimination algorithms are widely used in fields such as human behavior prediction and autonomous driving, and the related technologies are relatively mature. Currently, intent discrimination technologies are mainly divided into two categories: one is based on designing models based on prior knowledge, adjusting relevant parameters manually or adaptively to form a deterministic discrimination model; the other is based on neural networks, which does not require manual design of discrimination logic, but directly learns potential information from data to automatically implement the intent discrimination model.

[0003] Traditional intent discrimination techniques largely rely on manually designed features and rules. Experts need to manually set the weights of each component based on prior task rules and experience. This design process is cumbersome and difficult to adapt to complex and changing environments, resulting in insufficient generalization ability and robustness of the model. Furthermore, most existing technologies only focus on single targets, exhibiting significant limitations in application scenarios. In real-world scenarios, targets act in clusters. Compared to single targets, the behavior patterns, intent classifications, and even the interactions between different clusters of targets differ greatly, making it impossible to simply apply single-target intent discrimination techniques. Therefore, how to achieve accurate intent discrimination for groups of targets has become a pressing technical problem. Summary of the Invention

[0004] To address the problems existing in the above-mentioned traditional methods, this invention proposes a group target intent discrimination method and a group target intent discrimination device based on combinatorial policy deep learning, which can achieve accurate discrimination of group target intent.

[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: On the one hand, a group target intent discrimination method based on combined strategy deep learning is provided, including the following steps: Acquire feature data of all targets in the current scene; the feature data is time-series data calculated based on the target's trajectory over a past period of time, and includes distance, azimuth, orientation angle, and deviation angle; The feature combinations obtained by pairing the feature data of all targets are input into multiple independent interaction intent discrimination networks for discrimination, and the target relationship matrix corresponding to the target relationship between any two targets is obtained. Based on the target relationship matrix, all targets are divided into clusters. The number of interaction intent discrimination networks is equal to the number of feature combinations. Targets in the same cluster are in a synchronous relationship. A target can only belong to one cluster and all targets in the same cluster have the same intent. Based on the cluster partitioning results, the feature data of each cluster are statistically calculated and then input into multiple independent action intent discrimination networks for discrimination to obtain the action intent of each cluster. The number of action intent discrimination networks is equal to the number of clusters. The input of an action intent discrimination network is the feature data of a single cluster. Both the interaction intent discrimination network and the action intent discrimination network are MTSC networks.

[0006] In one embodiment, the interaction intent discrimination network includes a patch partitioning unit and a base model; The patch partitioning unit uses the patching method to partition the input feature data from the time dimension and the feature dimension respectively, resulting in multi-granularity time-series data patching. The base model independently performs intent discrimination on each time series data patch, obtains the predicted latent vector corresponding to each time series data patch, and then performs weighted fusion based on the weight of each time series data patch to output the intent discrimination result. The base model uses a BiLSTM network as the backbone network and uses cross attention to calculate and determine the weight of each time series data patch based on the comprehensive calculation of each time series data patch and the corresponding predicted latent vector.

[0007] In one embodiment, the base model includes a feedforward neural network, a BiLSTM network, and a self-attention mechanism layer connected in sequence. The feedforward neural network is used to extract features from the input temporal data patching to obtain the input latent vector. The BiLSTM network is used to process the input latent vector into an intermediate latent vector. The self-attention mechanism layer is used to process the intermediate latent vector into a predicted latent vector.

[0008] In one embodiment, when calculating the original feature data using the target's trajectory, the trajectories of two targets are consistently used as input; if there is only one target, the origin is selected as the other target, and the calculated feature data includes the distance between the two targets. b Relative to the target a The azimuth, orientation, and deviation angles, as well as the target a Relative to the target b The azimuth, orientation, and deviation angles.

[0009] In one embodiment, in the step of obtaining feature data of all targets in the current scene, after calculating the original time-series data of all targets, the numerical data is preprocessed using min-max normalization, and the angular data is preprocessed using sin-cos encoding.

[0010] On the other hand, a group target intent discrimination device based on combinatorial strategy deep learning is also provided, including: The feature acquisition module is used to acquire feature data of all targets in the current scene. The feature data is time-series data calculated based on the target's trajectory over a period of time, and includes distance, azimuth, orientation angle, and deviation angle. The target relationship recognition module is used to input the feature combinations obtained by pairing the feature data of all targets into multiple independent interaction intent discrimination networks for discrimination, and obtain the target relationship matrix corresponding to the target relationship between any two targets. Based on the target relationship matrix, all targets are divided into clusters. The number of interaction intent discrimination networks is equal to the number of feature combinations. Targets in the same cluster are in a synchronous relationship. A target can only belong to one cluster and all targets in the same cluster have the same intent. The cluster intent discrimination module is used to statistically calculate the feature data of each cluster based on the cluster division results, and then input them into multiple independent action intent discrimination networks for discrimination to obtain the action intent of each cluster. The number of action intent discrimination networks is equal to the number of clusters. The input of an action intent discrimination network is the feature data of a single cluster. Both the interaction intent discrimination network and the action intent discrimination network are MTSC networks.

[0011] In one embodiment, the interaction intent discrimination network includes a patch partitioning unit and a base model; The Patch partitioning unit is used to partition the input feature data from the time dimension and the feature dimension respectively using the Patching method to obtain multi-granularity time-series data Patching. The base model is used to independently determine the intent of each time series data patch. After obtaining the predicted latent vectors corresponding to each time series data patch, it is weighted and fused according to the weights of each time series data patch to output the intent determination result. The base model uses a BiLSTM network as the backbone network and uses cross attention to calculate and determine the weights of each time series data patch based on the comprehensive calculation of each time series data patch and the corresponding predicted latent vector.

[0012] In one embodiment, when the feature acquisition module calculates the original feature data using the target's trajectory, it consistently uses the trajectories of two targets as input; if there is only one target, the origin is selected as the other target, and the calculated feature data includes the distance between the two targets. b Relative to the target a The azimuth, orientation, and deviation angles, as well as the target a Relative to the target b The azimuth, orientation, and deviation angles.

[0013] In one embodiment, after calculating all the original time-series data of the targets, the feature acquisition module preprocesses the numerical data using min-max normalization and the angular data using sin-cos encoding.

[0014] One of the above technical solutions has the following advantages and beneficial effects: The aforementioned group target intent discrimination method and device based on combined strategy deep learning redesigns the overall structure and processing flow of the discrimination model, adds cluster partitioning processing, and realizes a discrimination process from local to global. This enables the model to handle scenarios with multiple targets simultaneously. Furthermore, considering the characteristics of temporal data of ground targets, an intent discrimination network based on patching technology and attention mechanism is designed. By using base models with different patching partitions to capture the intrinsic relationship between different feature dimensions and time dimensions, and by using the attention mechanism to perform decision-level fusion of base models, the discrimination results from different perspectives are taken into account, improving the discrimination accuracy and robustness of the model and enabling the discrimination model to handle group target scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a group target intent discrimination method based on deep learning of combined strategies in one embodiment. Figure 2 This is a schematic diagram of the design of a group target intent discrimination model framework in one embodiment; Figure 3 This is a schematic diagram of a deep learning group target intent discrimination network in one embodiment; Figure 4Here are the ROC curves of each discrimination model on the interaction intent dataset in one embodiment; Figure 5 Here are the ROC curves of each discrimination model on the action intent dataset in one embodiment; Figure 6 This is a schematic diagram of the module structure of a group target intent discrimination device based on deep learning of a combination strategy in one embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. 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 invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0018] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] Intent discrimination techniques based on prior knowledge have made significant progress and are gradually maturing. These techniques mainly include template matching and Bayesian networks. Template matching first establishes a template database based on expert experience, then extracts intent feature information based on the target's actual actions. Finally, the matching degree between the feature information and the template is derived using the maximum similarity method; the template with the highest matching degree is the intent discrimination result. Bayesian networks originate from Bayes' theorem and are a product of the combination of probability theory and graph theory. A Bayesian network consists of nodes representing events and their states, and directed arcs representing the event propagation relationships. After constructing the network, it is necessary to determine the network parameters: for top-level events without parent nodes, the prior probability is determined; for events with parent nodes, the conditional probability is given. New events are detected and the network parameters are updated through backpropagation until a certain intent hypothesis exceeds a preset threshold; this intent is then considered the intent discrimination result.

[0021] In recent years, advancements in data acquisition and computing capabilities have enabled data-driven intent discrimination algorithms, including neural networks and deep learning. The goal of neural networks is to train an intent discrimination network with classification capabilities. First, scene situational information related to the intent is extracted, and after data preprocessing, a situational feature vector that the network can discriminate is formed. Then, the feature vector is input into the network to solve for the intent discrimination result. Deep learning networks are artificial neural networks with multiple implicit layers. The existence of multiple layers allows deep learning networks to learn more abstract contextual features. By utilizing layer-by-layer feature processing in deep learning networks, high-level features can be progressively extracted from low-level input features, resulting in more accurate discrimination.

[0022] However, existing intent discrimination techniques still have several drawbacks: First, intent discrimination is essentially a time-series task. In real-world scenarios, target behavior is highly complex, and the process of acquiring target data also involves a degree of uncertainty. Analyzing only a single moment or short segment cannot accurately capture continuous behaviors such as lane changes, acceleration, and deceleration, leading to limited and deceptive intent discrimination results. At the feature processing level, current techniques typically input multiple features directly into the model, ignoring the inherent physical meaning and correlation between these features. Since discrimination models tend to focus on more salient features, this input method causes them to overlook the intrinsic connections and deeper information between different features, thus failing to make accurate intent discrimination.

[0023] In one embodiment, such as Figure 1 As shown, a group target intent discrimination method based on combined policy deep learning is provided, which may include the following processing steps S10 to S14: S10, acquire feature data of all targets in the current scene; the feature data is time-series data calculated based on the target's trajectory over a past period of time, and includes distance, azimuth, orientation angle and deviation angle; S12, the feature combinations obtained by pairing the feature data of all targets are input into multiple independent interaction intent discrimination networks for discrimination, and the target relationship matrix corresponding to the target relationship between any two targets is obtained. Based on the target relationship matrix, all targets are divided into clusters. The number of interaction intent discrimination networks is equal to the number of feature combinations. Targets in the same cluster are in a synchronous relationship. A target can only belong to one cluster and all targets in the same cluster have the same intent. S14. Based on the cluster division results, the feature data of each cluster are statistically calculated and then input into multiple independent action intent discrimination networks for discrimination to obtain the action intent of each cluster. The number of action intent discrimination networks is equal to the number of clusters. The input of an action intent discrimination network is the feature data of a single cluster. Both the interaction intent discrimination network and the action intent discrimination network are MTSC networks.

[0024] It is understood that this embodiment will utilize a target combination strategy to achieve intent discrimination in group target scenarios, employ deep learning technology with temporal analysis capabilities, and improve the discrimination model by combining the properties of target feature data to achieve accurate discrimination of group target intent.

[0025] In the task of target intent discrimination, intent represents the motivation and plan for a target to take action to achieve its purpose. In real-world scenarios, while it's impossible to directly obtain abstract information such as the target's intent, plan, and thought process, we can analyze observable target characteristics (such as changes in the target's actions or state) to determine the target's intent. This process of inferring the target's intent through analysis of its characteristics is called target intent discrimination. In other words, target intent discrimination establishes a mapping from target characteristic data to target intent.

[0026] This embodiment divides the intent space into two layers: the first is the interaction intent, which describes the interaction relationship between two targets, including no relationship, synchronous relationship, proximity relationship, and distance relationship; the second is the action intent, which describes the target's own behavior in the task, including specifying action, withdrawal, maneuvering, and stationing. The specific content of each intent and relationship is shown in Table 1 below.

[0027] Table 1

[0028] The feature data selected in this embodiment is time-series data of four features, namely distance, azimuth, orientation angle and deviation angle, calculated based on the trajectory of the target over a past period of time (the specific duration can be set according to the specific scenario, such as but not limited to the past 1 minute, past 3 minutes or past 5 minutes).

[0029] Furthermore, to unify the data format for both interactive and action intentions, this embodiment designs a universal feature calculation method. When calculating the original feature data using the target's trajectory, the trajectories of two targets are always used as input. If there is only one target, the origin is selected as the other target. The calculated feature data includes seven dimensions: the distance between the two targets. d ,Target b Relative to the target a azimuth α a Orientation angle θ a and deviation angle φ a and target a Relative to the target b azimuth α b Orientation angle θ b and deviation angle φ b .

[0030] The specific calculation method is as follows: First, obtain the target... a and target b exist T Coordinate trajectory within a given time period:

[0031] And the change in orientation angle:

[0032] in, Indicate target i In the j The coordinates at each moment. Indicate target i In the j The orientation angle at each moment. Note that both the coordinates and the orientation angle are relative to the scene origin (usually the friendly target). Furthermore, when determining the action intent, the origin needs to be considered as the target. a At this time, the target a exist T Coordinate trajectory within a given time period Changes in orientation angle All values ​​are 0. i Possible values a and b , j Values ​​can range from 1 to T .

[0033] Then, based on the target's coordinate trajectory and changes in orientation angle, the characteristics at each moment are calculated step by step; for the first... t Target distance at time 1 d t It can be calculated using the following formula:

[0034] Then, the angular features are calculated for the target. b Relative to the target a Orientation angle θ at It can be by Obtain directly:

[0035] And the goal b Relative to the target a azimuth α at and deviation angle φ at Additional calculations are required:

[0036] Similarly, the goal a Relative to the target b azimuth α bt Orientation angle θ bt and deviation angle φ bt It can be calculated using the following formula:

[0037] Furthermore, after obtaining the original time-series data of the target, preprocessing is required to enable the neural network to better process this data due to the different dimensions of each feature. For numerical data (such as target distance), the preprocessing operation can use minimum-maximum normalization to map it to the range [0,1], as shown in the following formula:

[0038] in, v These are the original numerical feature values ​​(such as target spacing) in the original time-series data of the target. min is the minimum value of this feature value, and max is the maximum value of this feature value. These are the eigenvalues ​​obtained after normalization.

[0039] For angular data (including azimuth, orientation, and deviation angles), sin-cos encoding is used, and the calculation method is as follows:

[0040] in, r The original angular features (such as) in the original time series data representing the target. α at , θ at , φ at , α bt , θ bt and φ bt ), r s andr c These represent the corresponding features obtained after encoding each of the original angular features.

[0041] Ultimately, the target feature data obtained after preprocessing contains 13 dimensions at each time step:

[0042] Since the target features change over time, the input data will contain target features from multiple time points. Therefore, a feature matrix can be used. X To represent input data:

[0043] in, T This means that the data contains a total of T At that moment, x t Representing the t The feature vector after preprocessing at each time step t =1,2,…, T , The target spacing is obtained after normalization.

[0044] Overall framework design: Most existing intent discrimination technologies focus on single-target scenarios, analyzing the characteristics of a single target to determine its intent. However, in real-world ground scenarios, multiple targets often exist simultaneously, and the number of targets may even change dynamically. Furthermore, different targets may interact with each other, such as assembling, disbanding, or acting in groups. These interactions can also affect the target's intent. Therefore, when discerning intent in multi-target scenarios, it is necessary to consider not only the characteristics of the target itself but also the mutual influence between different targets.

[0045] To achieve this, more global target features need to be introduced when performing intent discrimination. A simple approach is to use the features of all targets as input. However, since the number of targets in a scene is uncertain, the dimension of the input features obtained by merging will also change with the number of targets. Conventional discrimination models cannot cope with this situation of changing input dimensions. Therefore, it is necessary to design a special discrimination model for intent discrimination tasks in multi-target scenes.

[0046] To meet the specific requirements of the intent discrimination task, this embodiment redesigns the model framework and processing flow to enable it to handle intent discrimination tasks in multi-objective scenarios. The improved model framework is as follows: Figure 2As shown, it can be mainly divided into two modules: a target relationship identification module focusing on global relationships and a cluster intent discrimination module focusing on local targets. These two modules correspond to the target's interaction intent and action intent, respectively.

[0047] Suppose a containing n In scenarios involving individual ground targets, the method proposed in this embodiment is used for intent determination.

[0048] First, it is necessary to calculate the feature data of all targets in the current scene (e.g., ...). Figure 2 The target features (1 to n) are input into the target relationship recognition module. These features are paired and then input into an independent interaction intent discrimination network (hereinafter referred to as discrimination network A) to determine the corresponding relationships between targets. n The goal is obtained through pairing. The features are combined and input into In an independent discriminant network A, discriminant network A is a trained MTSC (Multivariate time series classification) network whose input is fixed as the feature data of two targets, and whose output is the inter-target relationship between the two targets.

[0049] Repeat this operation for all feature combinations of all targets to obtain the inter-target relationships between any two targets. Summarize these relationship results into a dataset of size [size missing]. The target relationship matrix is ​​used to group targets that are synchronized into the same cluster. A target can only belong to one cluster, and all targets within the same cluster have the same intent. This grouping process is a logical correction behavior based on a graph structure. Thus, by analyzing the action intent of the cluster, the intent of each target within the cluster can be obtained. n The goals can ultimately be divided into m Clusters, of which , No. i The number of targets in each cluster is , .

[0050] After the clusters are divided, the cluster intent determination module determines the action intent of each cluster.

[0051] First, it is necessary to obtain the feature data of each cluster. This can be calculated by taking the average, median, or other methods from the features of each target in the cluster. m The feature data of each cluster are respectively input into mThe intention of each action is determined in an independent action intent discrimination network (hereinafter referred to as discrimination network B). Discrimination network B is also an MTSC network. Its input is the feature data of a single cluster, and its output is the action intent of that cluster. Through discrimination network B, the action intent of each cluster can be obtained, which is also the action intent of each target in the cluster. Thus, the interaction intent and action intent of all targets can be obtained.

[0052] The model framework first fixes the feature dimension of the actual input network by pairing targets, thus solving the problem of variable input dimension caused by uncertain number of targets. Then, it divides the targets into clusters based on the target relationship matrix obtained by the target relationship recognition module, grouping targets with the same intent into the same cluster. Then, it judges the action intent on a cluster-by-cluster basis, so that the judgment of action intent only needs to consider the features of the cluster itself, simplifying the computation of subsequent processes.

[0053] Meanwhile, this model framework divides the target's intent into two levels: interaction intent and action intent. These correspond to the global relationship between all targets and the local situation of a single target, respectively. This expands and clarifies the division of the intent space. Compared with traditional techniques that only consider the target itself, this model framework design can provide more information and enhance the interpretability of the judgment results.

[0054] Furthermore, the main discrimination processes in the target relationship recognition module and the cluster intent discrimination module can run in parallel. Even when the number of targets is large, the running time of the entire model will not be greatly affected, which meets the real-time requirements of practical applications. The discrimination network A and discrimination network B adopt a modular design and are decoupled from the model framework. They can be replaced with other MTSC algorithms at will, or the network can be optimized separately according to actual usage requirements, which makes them highly scalable.

[0055] The aforementioned group target intent discrimination method based on combined strategy deep learning redesigns the overall structure and processing flow of the discrimination model, adds cluster partitioning processing, and realizes a discrimination process from local to global. This enables the model to handle scenarios with multiple targets simultaneously. Furthermore, considering the characteristics of ground target time-series data, an intent discrimination network based on patching technology and attention mechanism is designed. By using base models with different patching partitions to capture the intrinsic relationship between different feature dimensions and time dimensions, and by using the attention mechanism to perform decision-level fusion of base models, the discrimination results from different perspectives are taken into account, improving the model's discrimination accuracy and robustness, and enabling the discrimination model to handle group target scenarios.

[0056] Regarding the design of discriminant networks: Compared to other time series analysis tasks, ground target intent recognition tasks have two problems: First, the time series data of ground targets are often obtained through remote sensing images and other means, with a long actual collection time interval and fewer time dimensions in the data obtained; second, the features of ground targets have clear physical meanings, and there are certain correlations between different features, which can affect the analysis results.

[0057] Conventional time series analysis algorithms cannot effectively address these issues. Therefore, to effectively identify the intent of ground targets, this embodiment proposes an intent discrimination algorithm based on the Patching method and a BiLSTM (Bi-directional Long Short-Term Memory) network, such as... Figure 3 As shown, to address the issue of limited time-dimensional data and correlation among features, this model utilizes a patching method. It partitions the input feature data into patching segments based on both the time and feature dimensions. Different patching data are processed using independent feature extraction and temporal prediction networks. Finally, an attention mechanism is employed to fuse the results from each patching segment at the decision level, further improving the network's accuracy and robustness. Figure 3 In the middle, x n and x n+1 These represent the nth and (n+1)th input data points of the time series data, respectively.

[0058] Patching methods divide raw time-series data into time or feature dimensions, enabling the network to extract information from the raw data from different spatiotemporal perspectives. This approach is widely used in time series analysis algorithms. The multi-granularity patching method used in this embodiment is improved to address the characteristics of ground target time-series data. In the time dimension, it divides the data by different time frames and time intervals to compensate for the limited amount of raw target time-series data. In the feature dimension, it divides the data based on the correlation between different features, such as different coordinate systems, the target's own position (e.g., coordinates and azimuth), and the relationships between targets (e.g., orientation angle and its angle), enhancing the network's ability to mine deeper information from the time-series data. Finally, the two division methods are combined to obtain multi-granularity time-series data patching.

[0059] In one embodiment, the interaction intent discrimination network includes a patch partitioning unit and a base model. The patch partitioning unit uses the patching method to partition the input feature data from both the temporal and feature dimensions, resulting in multi-granularity temporal data patches. The base model independently performs intent discrimination on each temporal data patch, obtaining the predicted latent vectors corresponding to each patch, and then weights and fuses them according to their respective weights to output the intent discrimination result. The base model uses a BiLSTM network as the backbone network and uses cross-attention to calculate and determine the weights of each temporal data patch based on the combined calculation of the predicted latent vectors.

[0060] It is understandable that after each patching is processed by the base model to obtain the corresponding prediction results, these prediction results need to be fused at the decision level. This embodiment calculates the weights of each patching based on a cross-attention mechanism. The attention mechanism enables the neural network model to distinguish the importance of each vector in the sequence and assign different weights, making the learning of the neural network model more flexible and greatly improving its interpretability. This embodiment uses cross-attention to integrate each patching with its corresponding prediction results (i.e., predicted latent vectors). H P The final weights are determined using the following method:

[0061]

[0062]

[0063]

[0064] Where Linear() represents a linear layer. H P These are the predicted latent vectors output for each patching. P t Encoding each patching in the time dimension. P f For each patching, encode it in the feature dimension, then query. Q Depend on H P A linear transformation yields a size of patch_num × hidden_size, where patch_num is the number of patches and hidden_size is the dimension of the hidden layer; key K Sum V respectively byP t and P f After linear projection size alignment and stitching, the key is obtained. K Size and query Q Same, value V The size is patch_num×1, and the weights of each patch are calculated according to the cross-attention formula. Weight p Concatenate() is a concatenation operation. This is the scaling factor.

[0065] Specifically, for the input time-series data of ground targets:

[0066] in, Representing the t The time step n Data for each feature, which contains a total of [number] features. T Each time step and N Each feature is processed using a multi-granularity patching method to obtain... m Patching data For the first i Patching data :

[0067] It can be seen that, for A subset, for A subset, corresponding also by X It consists of subsets of .

[0068] In one embodiment, such as Figure 3 As shown, the base model consists of a feedforward neural network, a BiLSTM network, and a self-attention mechanism layer connected in sequence. The feedforward neural network is used to extract features from the input temporal data patching to obtain the input latent vector. The BiLSTM network is used to process the input latent vector into an intermediate latent vector, and the self-attention mechanism layer is used to process the intermediate latent vector into a predicted latent vector.

[0069] It is understandable that after obtaining the patching data, each patching data needs to be input into an independent base model for discrimination. In the base model, the input patching data first undergoes feature extraction through a feed-forward network (FFN) to obtain the input latent vector. H I :

[0070] in, ReLU As the activation function, since the number of features is different for each patching, we can specify here: if the input patching data... The feature dimension is of length Then the first one obtained after FFN input latent vectors H Ii The feature dimension length is min( ,64). FFN consists of ReLU activation layers and linear layers. Optionally, the number of base models can be equal to or less than the number of patches (e.g., all or part of the patches are processed serially).

[0071] The extracted latent vectors are input into the base model for time series analysis. In this embodiment, BiLSTM is used as the backbone network of the base model. BiLSTM is a typical RNN (Recurrent Neural Network) model, which can effectively handle time series data-related problems and consists of multiple LSTM cells. Input latent vectors H I The intermediate hidden vector is obtained after processing by a BiLSTM network. H M :

[0072] in, H Mi Indicates the first There are several intermediate latent vectors. Then, a self-attention layer (SelfAtten) is used to process the RNN output, adaptively assigning information weights, focusing on key content, and improving the model's ability to model sequence information. Matmul is matrix multiplication, and Scale is scaling operation; intermediate latent vectors H M The predicted latent vector is obtained after processing by the SelfAtten self-attention mechanism layer. H P :

[0073] in, H Pi Indicates the first Each predicts a hidden vector. Based on the output characteristics of the BiLSTM network, the three dimensions of the output are batch size, time step size, and hidden layer size. This indicates that the predicted latent vector is taken from the last time step as the output of the base model.

[0074] After obtaining the predicted latent vectors output by the base model from each input latent vector, we can use cross-attention weights (i.e., patch attention) to perform weighted fusion as described above to obtain the final intent discrimination result output by the discriminative network. Y :

[0075] In this process, the output dimension of the linear layer is equivalent to the intent category that needs to be determined, thus obtaining the intent determination result. Y Then, this can be converted into the probability of each intent through the Softmax (normalization exponent) function, or the intent label with the highest probability can be output through the argmax function.

[0076] In some implementations, one experimental example is also provided: The experimental platform was built using Python programming. The experimental data was generated by the simulation system. Users interactively specified the action routes and methods of each target, and the program automatically calculated the time-series data of various features of that target within a specified time range, and labeled the interaction intent and action intent according to the user's settings. Through this method, this example obtained interaction intent datasets and action intent datasets for offline training of two discriminative models, respectively. Each sample contains 13 dimensions of time-series feature data, and each time-series data point contains 5 sampling points, corresponding to 15 minutes of real-world data.

[0077] Next, we will compare and analyze the intent discrimination model (denoted as Ours) proposed in this paper with other existing intent discrimination algorithms. In this experiment, we selected 13 intent discrimination algorithms, including the traditional MTSC algorithm, RNN class, TCN (Temporal Convolutional Networks) class and Transformer class (as shown in the model column in Table 2), and trained and tested them under the same experimental environment and settings. The results are shown in Table 2 below.

[0078] Table 2

[0079] As can be seen, all three mainstream MTSC algorithms have good discrimination accuracy. For RNN-type algorithms, it was observed that improving the network itself, such as adding convolutional layers and attention mechanisms, did not significantly improve the model's performance. Similarly, for Transformer-type algorithms, using patching techniques to process the data actually reduced the discrimination accuracy. This is due to the characteristics of the intent discrimination data itself. Since intent discrimination data is time-series data, its time dimension is relatively short. Most commonly used UCR datasets contain dozens or hundreds of sampling points, while the interaction intent data and action intent data used in the experiment only have 5 sampling points. Networks designed for conventional time-series data are difficult to be effective on the short data used for intent discrimination, resulting in a performance degradation of these networks after improvement.

[0080] The model proposed in this paper is specifically designed for the characteristics of intent discrimination data. It combines patching technology with RNN networks to enhance its ability to mine short-term time-series data. It achieves the highest accuracy in both interaction intent discrimination and action intent discrimination tasks, at 95.52% and 95.26% respectively, improving upon other methods by at least 4.97% and 1.89%. Furthermore, it achieves the best results in all three evaluation metrics: precision, recall, and F1 score. Figure 4 and Figure 5 The ROC curves of each model are shown. It can be seen that the model proposed in this paper has the highest AUC value in both interaction intent discrimination and action intent discrimination, further demonstrating the accuracy of the model. The Random Classifier is a randomized classifier.

[0081] This paper innovatively proposes a general-purpose multi-target intent discrimination framework for ground target intent discrimination tasks, addressing the characteristics of multi-target scenarios and data inherent in such tasks. The framework divides multi-target intent discrimination into target relationship recognition and cluster relationship discrimination. Based on a pairwise combination strategy and employing modular and parallel design, it can handle target groups of arbitrary size. Furthermore, it clarifies the intent space of ground targets, dividing target intent into the target's own action intent and the interaction intent between targets, achieving a hierarchical and progressive discrimination process from local to global, consistent with the discrimination logic of real-world scenarios and providing more information for decision-making. Simultaneously, based on the characteristics of ground target temporal data, a novel intent discrimination network is proposed, using a BiLSTM network as the base model. Patching technology is innovatively introduced into the RNN network, enabling the base model to analyze the target's temporal data from multiple different temporal and spatial perspectives, and a cross-attention mechanism is used to complete decision-level fusion. Finally, the experimental section compares the proposed method with existing intent discrimination models, analyzing their discrimination performance on interaction intent and action intent. Experimental results show that the proposed method can achieve accurate discrimination of group target intents.

[0082] In summary, this paper first proposes a group target intent discrimination architecture based on a combination strategy, which solves the problem of uncertain input dimensions caused by the uncertainty of the number of targets. At the same time, the framework adopts a parallel design and is decoupled from the discrimination network, so that it can realize the intent discrimination of group targets without affecting the performance of the discrimination model.

[0083] Then, deep learning technology was used to improve the discriminative model. By using an end-to-end learning approach, discriminative high-level features were automatically mined from the raw data. This avoided the problem that traditional methods required manual design of features and rules and were difficult to adapt to changes in the scene. It reduced the difficulty of model design and improved the generalization ability of the model.

[0084] The intent discrimination model used in this paper adopts BiLSTM as the base model. As a typical recurrent neural network model, it introduces a memory and forgetting mechanism similar to the human brain. By adding reset gate and update gate, it effectively solves the gradient vanishing and gradient explosion problems that exist in the training process. It has strong temporal analysis capabilities and can make a judgment on the intent of the target by integrating data from multiple time points, which is in line with the temporal characteristics of the intent discrimination task.

[0085] At the feature processing level, considering the characteristics of time-series data used for intent discrimination, patching techniques and cross-attention mechanisms are introduced to improve the model. By performing multi-granularity patching in the time and feature dimensions, the model can analyze the input data from different perspectives, enhancing its ability to uncover relationships between different features and deep features. Finally, cross-attention weights are used to comprehensively evaluate the discrimination results of each patching, thereby improving the accuracy and robustness of the final discrimination result.

[0086] It should be understood that, although Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0087] In one embodiment, such as Figure 6 As shown, a group target intent discrimination device based on combined strategy deep learning is provided, which may include a feature acquisition module 11, a target relationship recognition module 13, and a cluster intent discrimination module 15. The feature acquisition module 11 is used to acquire feature data of all targets in the current scene; the feature data is time-series data calculated based on the targets' trajectories over a past period, and includes distance, azimuth, orientation angle, and deviation angle. The target relationship recognition module 13 is used to input the feature combinations obtained by pairwise pairing of all target feature data into multiple independent interactive intent discrimination networks for discrimination, obtaining the target relationship matrix corresponding to the target relationship between any two targets, and dividing all targets into clusters according to the target relationship matrix; the number of interactive intent discrimination networks is equal to the number of feature combinations, targets in the same cluster are in a synchronous relationship, a target can only belong to one cluster, and all targets in the same cluster have the same intent. The cluster intent discrimination module 15 is used to statistically calculate the feature data of each cluster based on the cluster division results, and then input them into multiple independent action intent discrimination networks for discrimination to obtain the action intent of each cluster. The number of action intent discrimination networks is equal to the number of clusters. The input of an action intent discrimination network is the feature data of a single cluster. Both the interaction intent discrimination network and the action intent discrimination network are MTSC networks.

[0088] The aforementioned group target intent discrimination device based on combined strategy deep learning redesigns the overall structure and processing flow of the discrimination model, adds cluster partitioning processing, and realizes a discrimination process from local to global. This enables the model to handle scenarios with multiple targets simultaneously. Furthermore, considering the characteristics of ground target time-series data, an intent discrimination network based on patching technology and attention mechanism is designed. By using base models with different patching partitions to capture the intrinsic relationship between different feature dimensions and time dimensions, and by using the attention mechanism to perform decision-level fusion of base models, the discrimination results from different perspectives are taken into account, improving the model's discrimination accuracy and robustness, and enabling the discrimination model to handle group target scenarios.

[0089] In one embodiment, when the feature acquisition module 11 calculates the original feature data using the trajectory of the target, it consistently uses the trajectories of two targets as input. If there is only one target, the origin is selected as the other target, and the calculated feature data includes the distance between the two targets. b Relative to the target a The azimuth, orientation, and deviation angles, as well as the target a Relative to the target b The azimuth, orientation, and deviation angles.

[0090] In one embodiment, after calculating all the original time-series data of the targets, the feature acquisition module 11 preprocesses the numerical data using min-max normalization and the angular data using sin-cos encoding.

[0091] In one embodiment, the interaction intent discrimination network includes a patch partitioning unit and a base model. The patch partitioning unit uses a patching method to partition the input feature data from both the temporal and feature dimensions, resulting in multi-granularity temporal data patches. The base model independently discriminates intent for each temporal data patch, obtaining the predicted latent vectors corresponding to each patch, and then weights and fuses them according to their respective weights to output the intent discrimination result. The base model uses a BiLSTM network as its backbone and employs cross-attention to calculate and determine the weights of each temporal data patch based on the combined calculation of the predicted latent vectors.

[0092] It is understood that the specific limitations of the aforementioned group target intent discrimination device based on combinatorial strategy deep learning can be found in the corresponding limitations of the group target intent discrimination method based on combinatorial strategy deep learning mentioned above, and will not be repeated here.

[0093] Each module in the aforementioned group target intent discrimination device based on combined strategy deep learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of a device with data processing capabilities, or stored in software within the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of computer devices already existing in the art.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus DRAM (RDRAM), and interface DRAM (DRDRAM), etc.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and all such modifications and improvements fall within the scope of protection of the present invention.

Claims

1. A method for group target intent discrimination based on combinatorial strategy deep learning, characterized in that, Including the following steps: Obtain feature data of all targets in the current scene; Feature data is time-series data calculated based on the target's trajectory over a past period of time. Feature data includes distance, azimuth, orientation angle, and deviation angle. The feature combinations obtained by pairing the feature data of all targets are input into multiple independent interaction intent discrimination networks for discrimination, and the target relationship matrix corresponding to the target relationship between any two targets is obtained. Based on the target relationship matrix, all targets are divided into clusters. The number of interaction intent discrimination networks is equal to the number of feature combinations. Targets in the same cluster are in a synchronous relationship. A target can only belong to one cluster and all targets in the same cluster have the same intent. Based on the cluster division results, the feature data of each cluster are statistically calculated and then input into multiple independent action intent discrimination networks for discrimination to obtain the action intent of each cluster. The number of action intent discriminator networks is equal to the number of clusters. The input of an action intent discriminator network is the feature data of a single cluster. Both the interaction intent discriminator network and the action intent discriminator network are MTSC networks.

2. The group target intent discrimination method based on combinatorial strategy deep learning according to claim 1, characterized in that, The interaction intent discrimination network consists of patch partitioning units and a base model; The patch partitioning unit uses the patching method to partition the input feature data from the time dimension and the feature dimension respectively, resulting in multi-granularity time-series data patching. The base model independently performs intent discrimination on each time series data patching, obtains the predicted latent vectors corresponding to each time series data patching, and then performs weighted fusion based on the weights of each time series data patching to output the intent discrimination result; The base model uses a BiLSTM network as the backbone network, and uses cross attention to calculate and determine the weight of each time series data patching based on the combined calculation of each time series data patching and the corresponding predicted latent vector.

3. The group target intent discrimination method based on combined strategy deep learning according to claim 2, characterized in that, The base model consists of a feedforward neural network, a BiLSTM network, and a self-attention mechanism layer connected in sequence. The feedforward neural network is used to extract features from the input temporal data patching to obtain the input latent vector. The BiLSTM network is used to process the input latent vector into an intermediate latent vector. The self-attention mechanism layer is used to process the intermediate latent vector into a predicted latent vector.

4. The group target intent discrimination method based on combined strategy deep learning according to any one of claims 1 to 3, characterized in that, When using the trajectory of the target to calculate the original feature data, the trajectories of two targets are always used as input; If there is only one target, the origin is selected as the other target. The calculated feature data includes the distance between the two targets. b Relative to the target a The azimuth, orientation, and deviation angles, as well as the target a Relative to the target b The azimuth, orientation, and deviation angles.

5. The group target intent discrimination method based on combined strategy deep learning according to claim 4, characterized in that, In the step of obtaining feature data of all targets in the current scene, after calculating the original time-series data of all targets, the numerical data is preprocessed using min-max normalization, and the angular data is preprocessed using sin-cos encoding.

6. A group target intent discrimination device based on combinatorial strategy deep learning, characterized in that, include: The feature acquisition module is used to acquire feature data of all targets in the current scene; Feature data is time-series data calculated based on the target's trajectory over a past period of time. Feature data includes distance, azimuth, orientation angle, and deviation angle. The target relationship recognition module is used to input the feature combinations obtained by pairing the feature data of all targets into multiple independent interaction intent discrimination networks for discrimination, and obtain the target relationship matrix corresponding to the target relationship between any two targets. Based on the target relationship matrix, all targets are divided into clusters. The number of interaction intent discrimination networks is equal to the number of feature combinations. Targets in the same cluster are in a synchronous relationship. A target can only belong to one cluster and all targets in the same cluster have the same intent. The cluster intent discrimination module is used to statistically calculate the feature data of each cluster based on the cluster division results, and then input them into multiple independent action intent discrimination networks for discrimination to obtain the action intent of each cluster. The number of action intent discriminator networks is equal to the number of clusters. The input of an action intent discriminator network is the feature data of a single cluster. Both the interaction intent discriminator network and the action intent discriminator network are MTSC networks.

7. The group target intent discrimination device based on combined strategy deep learning according to claim 6, characterized in that, The interaction intent discrimination network consists of patch partitioning units and a base model; The Patch partitioning unit is used to partition the input feature data from the time dimension and the feature dimension respectively using the Patching method to obtain multi-granular time-series data Patching. The base model is used to independently determine the intent of each time series data patching. After obtaining the predicted latent vectors corresponding to each time series data patching, the model is weighted and fused according to the weights of each time series data patching to output the intent determination result. The base model uses a BiLSTM network as the backbone network, and uses cross attention to calculate and determine the weight of each time series data patching based on the combined calculation of each time series data patching and the corresponding predicted latent vector.

8. The group target intent discrimination device based on combined strategy deep learning according to claim 7, characterized in that, The base model consists of a feedforward neural network, a BiLSTM network, and a self-attention mechanism layer connected in sequence. The feedforward neural network is used to extract features from the input temporal data patching to obtain the input latent vector. The BiLSTM network is used to process the input latent vector into an intermediate latent vector. The self-attention mechanism layer is used to process the intermediate latent vector into a predicted latent vector.

9. The group target intent discrimination device based on combined strategy deep learning according to any one of claims 6 to 8, characterized in that, When the feature acquisition module calculates the original feature data using the trajectory of the target, it always uses the trajectories of two targets as input. If there is only one target, the origin is selected as the other target. The calculated feature data includes the distance between the two targets. b Relative to the target a The azimuth, orientation, and deviation angles, as well as the target a Relative to the target b The azimuth, orientation, and deviation angles.

10. The group target intent discrimination device based on combined strategy deep learning according to claim 9, characterized in that, After calculating the original time-series data of all targets, the feature acquisition module preprocesses the numerical data using min-max normalization and the angular data using sin-cos encoding.

Images