Target track classification-based intention recognition method and device

Abstract

The invention relates to the field of situation awareness, in particular to an intention recognition method and device based on target track classification. The method mainly comprises the steps that a feature vector set related to intention is acquired through track information, a global feature extraction network is built based on an inception network, and global features in the feature vector set are acquired through the global feature extraction network; constructing a local feature extraction network based on a BiGRU-Attention network, and extracting time sequence features in the sample set by using the local feature extraction network; according to the method, target tracks are classified by using global features and time sequence features, a cross entropy loss function is optimized by using local loss, the weight of a minority class is adjusted, and a result of intention recognition is output by using softmax. According to the method, the intention recognition hierarchy can be improved through multi-layer feature extraction, the feature extraction efficiency is improved, and the problem that the classification precision is not high is solved.

Description

Technical Field

[0001] The present invention relates to the field of situation awareness, and in particular to a method and device for intention recognition based on target track classification. Background Art

[0002] Intent recognition usually adopts template matching methods guided by domain expert knowledge, reasoning methods based on Bayesian networks, and methods based on neural networks. The template matching method refers to the process of identifying the other party's action intention based on the template formulated by the domain expert, and the target action intention is obtained by matching the template through the matching algorithm. The reasoning method based on the Bayesian network refers to the construction of a Bayesian network based on the target behavior sequence, and the network parameters are determined by the domain experts to realize the reasoning of the other party's action intention. In the case of relatively scarce data, the above two methods can accurately identify the target intention. However, the current scene target state data presents multi-dimensional and massive characteristics. Template matching and reasoning methods require a large amount of domain knowledge modeling in the early stage, so it is difficult to effectively handle the complex dependencies between characteristics and data. The neural network method can better adapt to modern scene data by extracting scene situation information related to intention without relying on expert selection of features.

[0003] In order to correctly judge the target's intention in a short time, it is necessary to analyze the track information on the ground, in the air or on the water. The target's intention can form a hierarchical structure of "parent intention-child intention-behavior". The current intention recognition method lacks hierarchy for track feature extraction; at the same time, the collected data set is unevenly distributed, which makes the classifier decision boundary biased towards the majority class, resulting in low accuracy of intention recognition.

[0004] In view of this, how to overcome the defects of the existing technology and solve the problem of lack of hierarchy in feature extraction and uneven distribution of data sets in existing intent recognition methods is a problem to be solved in this technical field. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention solves the problems of lack of hierarchy in feature extraction and uneven distribution of data sets in existing intent recognition methods.

[0006] The embodiment of the present invention adopts the following technical solution:

[0007] In the first aspect, the present invention provides an intent recognition method based on target track classification, specifically: obtaining a feature vector set related to intent through track information, building a global feature extraction network based on the inception network, and obtaining global features in the feature vector set through the global feature extraction network; building a local feature extraction network based on the BiGRU-Attention network, and using the local feature extraction network to extract time series features in the sample set; using global features and time series features to classify the target track, using focal loss to optimize the cross entropy loss function, adjusting the weights of the minority class, and using softmax to output the result of intent recognition. Preferably, the method of obtaining a feature vector set related to intent through track information specifically includes: obtaining an initial feature set related to intent in the track information, and forming the initial features in the initial feature set into a one-dimensional time series feature set; normalizing the time series feature set, and using a sliding window to convert the time series feature set into a two-dimensional feature vector set.

[0008] Preferably, the use of a sliding window to convert a time series feature set into a feature vector set in a two-dimensional form specifically includes: slicing the time series feature set using a sliding window of a specified time length to obtain a corresponding number of frame slices; and organizing the frame slices into a feature vector set in a two-dimensional form according to the sequence length of the frame slices and the dimension of the initial features.

[0009] Preferably, three convolutional layers and staggered connections are used to extract feature dimensions of different scales of the target, and batch normalization and ReLU activation function are used after each convolutional layer; the extracted feature dimensions are concatenated using a concat layer, and a residual connection structure is added to the data pooling layer to output the extracted target behavior feature information through a maximum pooling layer.

[0010] Preferably, the method of obtaining the global features in the feature vector set through the global feature extraction network specifically includes: inputting the feature vector set into the global feature extraction network, for each feature vector in the feature vector set, using each convolution layer to respectively extract features of different target scales, and merging the extracted features through the maximum pooling layer, and using the merged features as the global features of the feature vector.

[0011] Preferably, the local feature extraction network is constructed based on the BiGRU-Attention network, specifically comprising: connecting the BiGRU layer and the Attention layer into a BiGRU-Attention network; and connecting the feature layer and the normalized index layer after the BiGRU-Attention network to form a local feature extraction network.

[0012] Preferably, the use of focal loss to improve the cross entropy loss function of the BiGRU-Attention network specifically includes calculating using the following formula:

[0013] FL softmax =-α c (1-p c ) γ log(p c );

[0014] Among them, α c represents the weight, p c represents the probability of the cth class in the output of the softmax activation function, the modulation factor (1-p c ) γ , γ ≥ 0 is used to distinguish between difficult and easy samples.

[0015] Preferably, the extracting the temporal features in the sample set using the local feature extraction network specifically includes calculating using the following formula:

[0016]

[0017] in, and are the states of the forward and backward hidden layers at time t respectively; and are the weights of the forward and backward hidden layer states at time t; b t is the bias of the hidden layer state at time t.

[0018] In a second aspect, the present invention provides an intention recognition device based on target track classification, specifically comprising: comprising at least one processor and a memory, wherein the at least one processor and the memory are connected via a data bus, and the memory stores instructions that can be executed by the at least one processor, and after the instructions are executed by the processor, they are used to complete the intention recognition method based on target track classification in the first aspect.

[0019] On the third aspect, the present invention also provides a non-volatile computer storage medium, wherein the computer storage medium stores computer executable instructions, and the computer executable instructions are executed by one or more processors to complete the method described in the first aspect.

[0020] In a fourth aspect, a chip is provided, comprising: a processor and an interface, for calling and running a computer program stored in the memory from a memory, and executing the method of the first aspect.

[0021] According to a fifth aspect, a computer program product comprising instructions is provided. When the instructions are executed on a computer or a processor, the computer or the processor executes the method according to the first aspect.

[0022] Compared with the prior art, the beneficial effect of the present invention is that the complete global features in the track information are first extracted through the data pooling layer improved based on the inception module, and then the important time series features are extracted using the BiGRU-Attention network. The multi-layer feature extraction not only improves the hierarchy of intent recognition but also speeds up the feature extraction efficiency. In the optimization scheme of the present invention, the cross entropy loss function of the model is improved by introducing focal loss, which improves the problem of low classification accuracy caused by the imbalance of various sample categories in the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 A flow chart of a method for intention recognition based on target track classification provided by an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of the recognition process of the intention recognition method provided by the present invention;

[0026] Figure 3 A flow chart of another method for intention recognition based on target track classification provided by an embodiment of the present invention;

[0027] Figure 4 A flow chart of another method for intention recognition based on target track classification provided by an embodiment of the present invention;

[0028] Figure 5 A flow chart of another method for intention recognition based on target track classification provided by an embodiment of the present invention;

[0029] Figure 6 A schematic diagram of the structure of a global feature extraction network in the intent recognition method provided by the present invention;

[0030] Figure 7 A schematic diagram of the structure of a local feature extraction network in the intent recognition method provided by the present invention;

[0031] Figure 8 A schematic diagram of the structure of an intention recognition device based on target track classification provided by an embodiment of the present invention;

[0032] The reference numerals are as follows:

[0033] 11: processor; 12: memory. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0035] The present invention is an architecture of a specific functional system, so the specific embodiments mainly illustrate the functional logical relationship between the various structural modules, and do not limit the specific software and hardware implementation methods.

[0036] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0037] Embodiment 1:

[0038] In order to address the problems in current intent recognition methods such as the lack of hierarchy in time series feature extraction and the low classification accuracy of intelligent models caused by unbalanced data sets, this embodiment provides an intent recognition method based on target track classification. This method uses a deep learning model combination method to improve the model's multi-level feature extraction capability and classification accuracy.

[0039] like Figure 1 As shown, the specific steps of the intention recognition method based on target track classification provided by the embodiment of the present invention are as follows:

[0040] Step 101: Obtain a feature vector set related to the intent through the track information, build a global feature extraction network based on the inception network, and obtain global features in the feature vector set through the global feature extraction network.

[0041] Intention recognition is to obtain the other party's action intention by integrating and reasoning the information indicators collected by various sensors, so as to assist the party in making action decisions. In order to correctly judge the intention of the ground, air or water target in a short time, it is necessary to analyze the target track information.

[0042] In actual implementation, the target's track information can be collected through radar detection or signal acquisition methods. The track information usually includes target status information such as longitude, latitude, altitude, speed, and azimuth. After collecting the track information, the target status information related to the intention can be selected, and the extracted initial features can be used to form an initial feature set. The initial feature set is then preprocessed to obtain a one-dimensional time series feature set, and the time series feature set is converted into two-dimensional data using a sliding window, thereby obtaining a feature vector set that can be used for a deep learning model, so that the subsequent feature extraction network and intent recognition network can be trained and predicted.

[0043] After obtaining the feature vector set, the data in the initial feature set is subjected to global feature extraction based on the improved inception network. In order to facilitate the training of the feature extraction network and the intent recognition network, the track information in the sample set can be divided into a training set and a test set. The training set consists of samples with intent labels, and the test set can be unlabeled. In actual implementation, the ratio of the training set to the test set can be 7:3.

[0044] Step 102: Build a local feature extraction network based on the BiGRU-Attention network, and use the local feature extraction network to extract temporal features in the sample set.

[0045] Since there is a hierarchical structure of "parent intention-child intention-behavior" in the target's intention, we can use the long-term and short-term interdependence of this hierarchy to extract temporal features related to behavioral intentions from historical track information to improve the hierarchy of feature extraction.

[0046] The feature vector set obtained in step 101 includes the historical track information that has been collected. In this embodiment, the bidirectional gated recurrent unit (Bidirectional Gated Recurrent Unit, abbreviated as BiGRU) layer and the attention layer are connected to extract local features in the feature vector set. In the general scenario of this embodiment, the time series features are used as local features.

[0047] Step 103: Use global features and temporal features to classify the target track, use focal loss to optimize the cross entropy loss function, adjust the weight of the minority class, and use softmax to output the result of intent recognition.

[0048] Because intent recognition is actually a classification task, the intent recognition network can use various deep learning networks for classification.

[0049] The training set is input into the BiGRU-Attention network, and the test set is input into the intent recognition network. The parameters of the BiGRU-Attention network are optimized according to the prediction structure of the intent recognition network. The intent recognition network is iteratively trained according to the optimized sample sharing degree until the specified stopping condition is reached, and the trained intent recognition network is obtained.

[0050] The features in the feature vector collection are directly obtained from the original historical track information, and there may be a problem of low classification accuracy caused by the imbalance of each sample category. Therefore, this embodiment also introduces a loss function for the classification problem, namely, the focal loss function, and improves the cross entropy loss function in the BiGRU-Attention network. The contribution of the sample is obtained through the improved cross entropy loss function, thereby improving the balance of the sample category.

[0051] On the other hand, after the target tracks are classified, the target track categories with a smaller proportion are called minority classes. In order to improve the recognition accuracy, the weights of the minority classes need to be adjusted.

[0052] After completing the training and optimization, a set of feature vectors to be identified is generated based on the track information to be identified, and a test set to be identified is generated using the global feature extraction network. The test set to be identified is input into the trained intent recognition network to predict the intent recognition result corresponding to the track information to be identified.

[0053] After the steps 101 to 103 provided in this embodiment, the intention recognition of the target track information is completed. In the specific implementation, the overall process of the intention recognition corresponding to the steps 101 to 103 is as follows: Figure 2 shown.

[0054] like Figure 3 As shown, the following method can be used to obtain a set of feature vectors related to the intention through the track information.

[0055] Step 201: Acquire an initial feature set related to the intention in the track information, and combine the initial features in the initial feature set into a one-dimensional time series feature set.

[0056] After completing the track information collection, according to the intention to be identified, the required initial features are selected from the collected track information, and the non-numerical features are encoded to construct the initial feature set. In order to avoid interference caused by abnormal data, appropriate methods can be used to clean or otherwise process the data that is obviously abnormal, missing, and repeated.

[0057] After completing the above processing, we can obtain a one-dimensional time series feature set that can represent the target action process:

[0058] (X,Y)=([x1,x2,...,x t ] i ,y i )∈R N×W . Where: R represents the real number space; N is the sequence length of the time series feature set; W is the feature dimension, that is, the number of features related to the intent; x t is the t-th feature data, represented by the specified encoding; [x1,x2,...,x t ] i The feature vector representing the i-th track information represents each initial feature corresponding to the i-th track information in the initial feature set; y i is a label indicating the intention of the i-th track information.

[0059] Step 202: Normalize the time series feature set and use a sliding window to convert the time series feature set into a two-dimensional feature vector set.

[0060] In a multi-target scenario, the different feature state data of different targets are quite different, resulting in the time series feature set containing feature data of multiple different dimensions, which needs to be normalized to eliminate the dimension effect. In actual implementation, the Z-score normalization method can be selected to normalize the data to a distribution with a mean of 0 and a variance of 1.

[0061] In this embodiment, in order to use the neural network structure for intent recognition, it is also necessary to convert the one-dimensional time series feature set into a two-dimensional feature vector set and input it into the neural network structure, and then extract features from the two-dimensional feature vector set through the neural network structure. The output of each layer in the neural network structure serves as the input of the next layer.

[0062] After step 201 and step 202 provided in this embodiment, a set of two-dimensional feature vectors that can be processed by a subsequent neural network structure can be obtained.

[0063] like Figure 4 As shown, the time series feature set can be converted into a two-dimensional feature vector set using a sliding window in the following manner.

[0064] Step 301: Slice the temporal feature set using a sliding window of a specified time length to obtain a corresponding number of frame slices.

[0065] Set a sliding window of length T, split the feature data in the one-dimensional time series feature set in units of T / 2, and form a frame slice with each T feature data split. After splitting in this way, each frame slice has 50% feature data overlap with other frame slices, which can ensure that the subsequent feature extraction networks such as the inception network and BiGRU-Attention network can capture more behavioral details.

[0066] Step 302: According to the sequence length of the frame slices and the dimension of the initial features, the frame slices are organized into a two-dimensional feature vector set.

[0067] After slicing, the features in the one-dimensional feature sequence are represented as a frame slice sequence of size N×W, and the two-dimensional frame slice sequence is a feature vector set, where N is the sequence length of the one-dimensional feature sequence and W is the feature dimension.

[0068] After step 301 and step 302 provided in this embodiment, the one-dimensional feature sequence can be converted into a two-dimensional feature vector set.

[0069] like Figure 5 As shown, the following method can be used to build a global feature extraction network based on the inception network. Figure 6 The global feature extraction network shown.

[0070] Step 401: Use three convolutional layers and staggered connections to extract feature dimensions of different scales of the target, and use batch normalization and ReLU activation function after each convolutional layer.

[0071] In the global feature extraction network, convolutional layers are used for feature extraction, and the input feature vector set will pass through multiple small convolutional layers to obtain temporal features and local features of different scales. In order to improve the network's feature extraction rate and the ability to extract intent feature information, batch normalization (BN) and nonlinear activation function (ReLU) are also required after each convolutional layer.

[0072] Step 402: Use the concat layer to concatenate the extracted feature dimensions, add a residual connection structure to the data pooling layer, and output the extracted target behavior feature information through the maximum pooling layer.

[0073] In order to extract features of different target scales, three behavior feature perception modules are used in the global feature extraction network of this embodiment. The target scale of the small convolution layer in each behavior feature perception module is different. Figure 6In the example, the sizes of the three small convolutional layers are 3*3*8, 3*3*16, and 3*3*16. Two convolutional layers of different scales and different numbers of levels are connected in parallel, and the feature dimensions extracted by different convolutional layers are concatenated together using the concat layer, and then the concatenated data is sent to the maximum pooling layer for pooling.

[0074] Furthermore, the above structure contains two small convolutional layers connected in series. In order to avoid the overfitting problem caused by the deepening of the number of layers during training, it is also necessary to add a residual connection structure to the global feature extraction network to output the extracted target behavior feature information through the maximum pooling layer.

[0075] After step 401-step 402 provided in this embodiment, a global feature extraction network can be constructed based on the improved Inception network.

[0076] After building the global feature extraction network, the feature vector set is input into the global feature extraction network. For each feature vector in the feature vector set, each convolution layer is used to extract features of different target scales respectively, and the extracted features are merged through the maximum pooling layer, and the merged features are used as the global features of the feature vector. Specifically: each feature vector X=[x1,x2,...,x t ] i ∈R N×W As the input pooling network layer F, each behavior feature perception module in the global feature extraction network learns the feature as max(0,X·W c +b c )), where W c represents the weight, b c Indicates the offset.

[0077] In order to accurately identify the target action intention, it is necessary to express the forward and backward dependencies based on high-dimensional features, and carry the bidirectional dependency information for the next step of learning. Figure 7 As shown in the figure, the BiGRU layer and the Attention layer are connected to form a BiGRU-Attention network, and the cross entropy loss function of the BiGRU-Attention network is improved using focal loss to adjust the weight of the minority class and avoid the problem of uneven distribution of classes in the sample set; the feature (Dense) layer and the normalized exponent (softmax) layer are connected after the BiGRU-Attention network to output the result of intent recognition to form a local feature extraction network.

[0078] In the constructed local feature extraction network, the BiGRU network has the global perception ability to extract long-term dependencies from both the forward and reverse directions. The calculation formula is as follows:

[0079]

[0080] in, and are the states of the forward and backward hidden layers at time t respectively; and are the weights of the forward and backward hidden layer states at time t; b t is the bias of the hidden layer state at time t.

[0081] In order to further improve the ability to express the intention process, after using the local feature extraction network to extract the long and short time series features of the intention, it is necessary to assign different weights to the extracted time series features to represent the contribution of the feature to the recognition of the intention. The feature representation obtained by the intention feature aggregation module is: Based on the attention mechanism, for each h i , calculate h i The degree of relevance to the object of interest is used as a weight to integrate and concatenate the feature representations to obtain the final representation

[0082] However, the initial feature data has a certain degree of category imbalance. In order to improve the problem of low classification accuracy caused by the imbalance of each sample category, the cross entropy loss function is improved by using focal loss. The calculation formula is as follows:

[0083] FL softmax =-α c (1-p c ) γ log(p c );

[0084] Among them, FL softmax represents focal loss, α c represents the weight, p c represents the probability of the cth class in the output of the softmax activation function, the modulation factor (1-p c ) γ , γ ≥ 0 is used to distinguish between difficult and easy samples.

[0085] Because action intention recognition is actually a classification task, softmax is selected to express the output in the form of probability. The calculation formula is as follows:

[0086] y i =max(softmax(r i ));

[0087] Among them, yi represents the output category, that is, the encoding of the target intention; max means taking the classification with the highest probability as the recognition result.

[0088] The intention recognition method based on target track classification provided in this embodiment has the following advantages:

[0089] (1) Through the improved data pooling layer and BiGRU-Attention network structure based on the inception module, not only the behavioral characteristics and timing information of the target track at different levels are extracted, but also the efficiency of feature extraction is improved.

[0090] (2) By introducing focal loss to improve the cross entropy loss function, the problem of low classification accuracy caused by the imbalance of sample categories in the data is improved, and the recognition accuracy of intentions, especially the target intentions that the party needs to pay more attention to, is improved.

[0091] Embodiment 2:

[0092] Based on the method for intention recognition based on target track classification provided in the above embodiment 1, the present invention also provides an intention recognition device based on target track classification that can be used to implement the above method, such as Figure 8 , which is a schematic diagram of the device architecture of an embodiment of the present invention. The device for identifying intentions based on target track classification in this embodiment includes one or more processors 11 and a memory 12. Figure 8 A processor 11 is taken as an example.

[0093] The processor 11 and the memory 12 may be connected via a bus or other means. Figure 8 The example of connecting through bus is taken in the following.

[0094] The memory 12 is a non-volatile computer-readable storage medium for the method for identifying intentions based on target track classification, and can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as the method for identifying intentions based on target track classification in Example 1. The processor 11 executes various functional applications and data processing of the device for identifying intentions based on target track classification by running the non-volatile software programs, instructions and modules stored in the memory 12, that is, implementing the method for identifying intentions based on target track classification in Example 1.

[0095] The memory 12 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 12 may optionally include a memory remotely arranged relative to the processor 11, and these remote memories may be connected to the processor 11 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0096] The program instructions / modules are stored in the memory 12, and when executed by one or more processors 11, the intention recognition method based on target track classification in the above embodiment 1 is executed, for example, the above described Figure 1 , Figures 3 to 5 The steps shown.

[0097] An embodiment of the present invention further provides a non-volatile computer storage medium, wherein the computer storage medium stores computer executable instructions, and the computer executable instructions are executed by one or more processors, such as Figure 8 A processor 11 in the embodiment may enable the one or more processors to execute the intention recognition method based on target track classification in embodiment 1, for example, to execute the above described Figure 1 , Figures 3 to 5 The steps shown can also be implemented Figure 2 , Figure 6 and Figure 7 The various modules and units described.

[0098] It is worth noting that the information interaction, execution process, etc. between the modules and units within the above-mentioned devices and systems are based on the same concept as the processing method embodiment of the present invention. The specific contents can be found in the description of the method embodiment of the present invention and will not be repeated here.

[0099] A person skilled in the art may understand that all or part of the steps in the various methods of the embodiments may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intention recognition based on target track classification, characterized in that: include: Obtain a set of feature vectors related to intent through track information, build a global feature extraction network based on the inception network, and obtain global features in the feature vector set through the global feature extraction network; Build a local feature extraction network based on the BiGRU-Attention network, and use the local feature extraction network to extract the temporal features in the sample set; Use global features and temporal features to classify the target track, use focal loss to optimize the cross entropy loss function, adjust the weight of the minority class, and use softmax to output the result of intent recognition.

2. The method for intention recognition based on target track classification according to claim 1, characterized in that: The step of obtaining a set of feature vectors related to the intention through the track information specifically includes: Obtain an initial feature set related to the intention in the track information, and combine the initial features in the initial feature set into a one-dimensional time series feature set; The time series feature set is normalized and converted into a two-dimensional feature vector set using a sliding window.

3. The method for intention recognition based on target track classification according to claim 2, characterized in that: The use of a sliding window to convert a time series feature set into a two-dimensional feature vector set specifically includes: Slice the temporal feature set using a sliding window of a specified time length to obtain a corresponding number of frame slices; According to the sequence length of the frame slices and the dimension of the initial features, the frame slices are organized into a set of feature vectors in a two-dimensional form.

4. The method for intention recognition based on target track classification according to claim 1, characterized in that: The global feature extraction network based on the inception network is constructed, specifically including: Use 3 convolutional layers and staggered connections to extract feature dimensions of different scales of the target, and use batch normalization and ReLU activation function after each convolutional layer; The concat layer is used to concatenate the extracted feature dimensions, and a residual connection structure is added to the data pooling layer to output the extracted target behavior feature information through the maximum pooling layer.

5. The method for intention recognition based on target track classification according to claim 4, characterized in that: The step of obtaining global features in a feature vector set through a global feature extraction network specifically includes: The feature vector set is input into the global feature extraction network. For each feature vector in the feature vector set, each convolutional layer is used to extract features of different target scales respectively, and the extracted features are merged through the maximum pooling layer, and the merged features are used as the global features of the feature vector.

6. The method for intention recognition based on target track classification according to claim 1, characterized in that: The local feature extraction network is constructed based on the BiGRU-Attention network, specifically including: Connect the BiGRU layer and the Attention layer to form a BiGRU-Attention network; The feature layer and the normalized index layer are connected after the BiGRU-Attention network to form a local feature extraction network.

7. The method for intention recognition based on target track classification according to claim 1, characterized in that: The use of focal loss to optimize the cross entropy loss function specifically includes calculating using the following formula: FL softmax =-a c (1-p c ) γ log(p c ); Among them, α c represents the weight, p c represents the probability of the cth class in the output of the softmax activation function, the modulation factor (1-p c ) γ , γ ≥ 0 is used to distinguish between difficult and easy samples.

8. The method for intention recognition based on target track classification according to claim 1, characterized in that: The method of extracting the time series features in the sample set using the local feature extraction network specifically includes calculating using the following formula: in, and are the states of the forward and backward hidden layers at time t respectively; and are the weights of the forward and backward hidden layer states at time t; b t is the bias of the hidden layer state at time t.

9. An intention recognition device based on target track classification, characterized in that: It includes at least one processor and a memory, wherein the at least one processor and the memory are connected via a data bus, and the memory stores instructions that can be executed by the at least one processor, and after the instructions are executed by the processor, they are used to complete the intention recognition method based on target track classification as described in any one of claims 1-8.

10. A non-volatile computer storage medium, characterized in that: The computer storage medium stores computer executable instructions, which are executed by one or more processors to complete the intention recognition method based on target track classification as described in any one of claims 1-8.

Images