Online track point identification method based on self-supervised learning model

Abstract

The invention provides an online track point recognition method based on a self-supervised learning model, and belongs to the technical field of deep learning, and the method comprises the steps: obtaining unlabeled sample data in a training set, carrying out the pre-training of the self-supervised learning model, and obtaining a mask encoder; transferring parameters of a mask encoder in the self-supervised learning model to an encoder of an end-to-end classification model, establishing an initial end-to-end classification model by connecting a classification head, training the initial end-to-end classification model through a small amount of labeled sample data in a training set, and obtaining the end-to-end classification model; and an encoder in the end-to-end classification model performs feature extraction on the parameter information of the online track points acquired in a preset time period, and a track point identification result is acquired through classification head prediction. According to the method, on the basis of the self-supervised learning model and the end-to-end classification model, online track points can be recognized, and complex track points are recognized while track point recognition efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an online trajectory point recognition method based on a self-supervised learning model. Background Art

[0002] With the popularization of smart devices, in order to achieve good human-computer interaction functions, smart devices all provide touch screens to receive online track points input by users.

[0003] At present, the recognition of online trajectory points is mainly based on manually extracting high-dimensional features based on high-order gradients or fitting the distribution of trajectory points based on Bezier curve parameters to extract trajectory point features for classification models to recognize. The traditional technology uses manual participation to extract trajectory point features, which not only makes the recognition efficiency of online trajectory points low, but also makes it difficult to recognize complex trajectory points, and it is also difficult to meet the needs of end-to-end classification prediction.

[0004] Therefore, an online trajectory point recognition method based on a self-supervised learning model is proposed. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, which is used to solve the problem that the recognition of online trajectory points in traditional technologies requires manual participation and has low recognition efficiency.

[0006] An embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, including an initial mask encoder and an initial decoder to obtain a self-supervised learning model;

[0007] Obtain unlabeled sample data in a training set to pre-train the self-supervised learning model, and obtain a mask encoder;

[0008] Migrating the parameters of the mask encoder in the self-supervised learning model to the encoder of the end-to-end classification model, and building an initial end-to-end classification model by connecting the classification head, and training the initial end-to-end classification model with a small amount of labeled sample data in the training set to obtain an end-to-end classification model;

[0009] The encoder in the end-to-end classification model extracts features from the parameter information of the online trajectory points, and predicts the features through the classification head to obtain the trajectory point recognition results.

[0010] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, wherein the steps of: obtaining a self-supervised learning model by constructing an initial mask encoder and an initial decoder; and comprising:

[0011] The initial mask encoder is constructed through the local enhancement feature extraction module; the initial decoder is constructed through the mask recovery module; the initial mask encoder is connected to the initial decoder to build a self-supervised learning model.

[0012] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, the steps of: obtaining unlabeled sample data in a training set to pre-train the self-supervised learning model, and obtaining a mask encoder; comprising:

[0013] Performing random masking processing on the unlabeled sample data according to a preset masking rate to obtain sample mask parameter information; optionally, including:

[0014] Get the parameter information P of the unlabeled sample data = (p1, ..., p s ), where p1 is the parameter information corresponding to the first trajectory point in the unlabeled sample data, p s is the parameter information corresponding to the s-th trajectory point in the unlabeled sample data. There are s trajectory points in the unlabeled sample data.

[0015] Randomly select mask parameter points in the parameter information of unlabeled sample data according to the preset mask rate Perform a mask operation, replace the mask parameter point with the standard parameter vector p0, and obtain the sample mask parameter information P ′ =(p1,…,p0,…,p0,…,p s ), where p k+1 is the parameter vector corresponding to the k+1th trajectory point in the parameter information of the unlabeled sample data, is the parameter vector corresponding to the k+k0th trajectory point in the parameter information of the unlabeled sample data, and k0 is the number of parameter points in the mask parameter point;

[0016] The local enhancement feature extraction module performs adaptive length enhancement on the local features of the sample mask parameter information and extracts the corresponding unlabeled sample feature vector;

[0017] The mask recovery module is used to recover the trajectory points according to the unlabeled sample feature vector, obtain the coordinate positions and stroke positions of the masked trajectory points in the unlabeled sample data, and construct an unlabeled recovery sample; the stroke position includes a starting point, a transition point, and an end point.

[0018] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, wherein the steps include: a local enhancement feature extraction module adaptively enhances the local features of the sample mask parameter information in length and extracts the corresponding unlabeled sample feature vector; comprising:

[0019] A local enhanced feature extraction module, including a feature extraction unit and a deformable feature extraction unit;

[0020] The sample mask parameter information is processed by the multi-head attention mechanism layer in the feature extraction unit to construct a query vector matrix, a key vector matrix and a value vector matrix, thereby obtaining a corresponding attention value; the perception layer performs nonlinear feature extraction on the sample mask parameter information according to the attention value through the GELU activation function to obtain a sample mask feature vector;

[0021] The deformable attention mechanism layer in the deformable feature extraction unit processes the sample mask feature vector, obtains the vector attention value in the dynamic window, performs weighted summation based on the sample mask feature vector and the vector attention value, obtains the weighted attention vector, performs a normalization layer operation, and obtains the standard attention vector; the multilayer perceptron performs dynamic feature sampling on the standard feature vector to obtain the unlabeled sample feature vector.

[0022] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, the steps of: the deformable attention mechanism layer in the deformable feature extraction unit processes the sample mask feature vector to obtain the vector attention value in the dynamic window, and performs weighted summation according to the sample mask feature vector and the vector attention value to obtain a weighted attention vector, performs a normalization layer operation to obtain a standard attention vector; the multilayer perceptron performs dynamic feature sampling on the standard feature vector to obtain an unlabeled sample feature vector; including:

[0023] Performing a weighted summation on the sample mask feature vector and the vector attention value in the dynamic window to obtain a weighted attention vector;

[0024]

[0025] Among them, T is the weighted attention vector obtained by processing the sample mask feature vector, H l is the trainable weight parameter of the lth attention head in the deformable attention mechanism layer, L i is the vector attention value corresponding to the i-th feature vector in the sample mask feature vector, V i is the vector value corresponding to the i-th feature vector in the sample mask feature vector, x i is the trajectory position of the i-th eigenvector, Δx i is the trajectory offset position of the i-th feature vector, n is the number of attention heads in the deformable attention mechanism layer, and m is the number of feature vectors in the dynamic window of the sample mask feature vector;

[0026] Performing a normalization layer operation according to the weighted attention vector to obtain a standard attention vector;

[0027]

[0028] Among them, LN(T i ) is the i-th attention vector element T in the weighted attention vector i The standard attention vector element obtained by normalization, μ i is the weighted attention vector T i The corresponding mean, σ i is the weighted attention vector T i The corresponding standard deviation, ∈ is a constant, α is the first normalization vector parameter of the normalization layer, and β is the second normalization vector parameter of the normalization layer;

[0029] The standard attention vector is obtained by traversing and normalizing the attention vector elements in the weighted attention vector.

[0030] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, the steps of: obtaining unlabeled sample data in a training set to pre-train the self-supervised learning model, and obtaining a mask encoder; thereafter, further comprising:

[0031] Set up the trajectory discriminator;

[0032] The unlabeled restored sample is identified by the trajectory identifier to determine whether the unlabeled restored sample has abnormal trajectory data. When it is determined that the unlabeled restored sample has abnormal trajectory data, the control parameters of the mask encoder are optimized.

[0033] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, the steps of: migrating the parameters of the mask encoder in the self-supervised learning model to the encoder of the end-to-end classification model, and building an initial end-to-end classification model by connecting the classification head, and training the initial end-to-end classification model with a small amount of labeled sample data in the training set to obtain an end-to-end classification model; including:

[0034] Obtaining labeled sample data in the training set;

[0035] Inputting the labeled samples into the initial end-to-end classification model, and the initial end-to-end classification model obtains corresponding label recognition results;

[0036] The label recognition result is compared with the standard recognition result corresponding to the labeled sample data, the classification head of the initial end-to-end classification model is trained, the control parameters of the classification head are optimized, and the end-to-end classification model is obtained.

[0037] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising:

[0038] When the initial end-to-end classification model is trained using the labeled sample data in the training set, there is no need to train the encoder in the initial end-to-end classification model.

[0039] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising:

[0040] The training set contains a large amount of unlabeled sample data and a small amount of labeled sample data;

[0041] The labeled sample data corresponds to a standard recognition result, wherein the standard recognition result is manually annotated based on the labeled sample data.

[0042] Preferably, the present invention provides an online trajectory point recognition method based on a self-supervised learning model, the steps of: extracting features from the parameter information of the online trajectory point through an encoder in an end-to-end classification model, and predicting through a classification head to obtain a trajectory point recognition result; including:

[0043] Extracting features from parameter information of the online trajectory points through an encoder to obtain a trajectory feature vector;

[0044] The Transformer layer in the classification head decodes and predicts the trajectory feature vector to obtain a pre-classified high-dimensional feature vector, performs downsampling operation through the pooling layer to obtain a pre-classified low-dimensional feature vector, performs nonlinear activation according to the pre-classified low-dimensional feature vector through the fully connected layer, and outputs the trajectory point recognition result.

[0045] Compared with traditional technologies, the beneficial effects of the present invention are: an online trajectory point recognition method based on a self-supervised learning model, which does not require human participation in the extraction of trajectory point features. Based on the self-supervised learning model and the end-to-end classification model, the online trajectory point recognition can be realized, which solves the problem of relying on high-order gradients to manually extract high-dimensional features or fitting trajectory points based on Bezier curve parameters for recognition in traditional technologies. While improving the efficiency of trajectory point recognition, the self-supervised learning model and the end-to-end classification model trained by the training set can also realize the recognition of complex trajectory points, thereby meeting the needs of end-to-end classification prediction.

[0046] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0047] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of the process of an online trajectory point recognition method based on a self-supervised learning model provided by the present invention. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0050] Embodiment 1:

[0051] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model. Figure 1 ,include:

[0052] Obtain parameter information of online trajectory points within a preset time period;

[0053] Obtain a self-supervised learning model by constructing an initial mask encoder and an initial decoder;

[0054] Obtain unlabeled sample data in the training set to pre-train the self-supervised learning model and obtain the mask encoder;

[0055] The parameters of the mask encoder in the self-supervised learning model are transferred to the encoder of the end-to-end classification model, and the initial end-to-end classification model is built by connecting the classification head. The initial end-to-end classification model is trained with a small amount of labeled sample data in the training set to obtain the end-to-end classification model.

[0056] The encoder in the end-to-end classification model is used to extract features from the parameter information of the online trajectory points, and the classification head is used to make predictions to obtain the trajectory point recognition results.

[0057] In the above embodiments, parameter information of online trajectory points is collected within a preset time period, feature extraction is performed on the parameter information of the online trajectory points through an encoder in an end-to-end classification model, and prediction is performed through a classification head to obtain trajectory point recognition results.

[0058] In the above embodiments, the initial mask encoder and the initial decoder in the self-supervised learning model are trained by using a large amount of unlabeled sample data in the training set to obtain the mask encoder.

[0059] In the above embodiments, the parameters of the mask encoder in the self-supervised learning model are transferred to the encoder of the end-to-end classification model, and the initial end-to-end classification model is built by connecting the classification head.

[0060] In the above embodiments, the classification head in the initial end-to-end classification model is trained by using a small amount of labeled sample data in the training set to obtain the end-to-end classification model.

[0061] The beneficial effects of the above technology are: the self-supervised learning model is pre-trained through a large amount of unlabeled sample data to obtain the mask encoder, and then the parameters of the mask encoder are transferred to the encoder of the end-to-end classification model, the classification head is connected, the initial end-to-end classification model is constructed, and fine-tuning training is performed through a small amount of labeled sample data to obtain the end-to-end classification model; the end-to-end classification model obtains the trajectory point recognition results according to the parameter information of the online trajectory points; compared with traditional technologies, the above method does not require manual participation in the extraction of trajectory point features, and can realize the recognition of online trajectory points based on the self-supervised learning model and the end-to-end classification model, which solves the problem of relying on high-order gradients to manually extract high-dimensional features or fitting trajectory points based on Bezier curve parameters for recognition in traditional technologies. While improving the efficiency of trajectory point recognition, the self-supervised learning model and the end-to-end classification model trained by the training set can also realize the recognition of complex trajectory points, thereby meeting the needs of end-to-end classification prediction.

[0062] Embodiment 2:

[0063] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising: obtaining a self-supervised learning model by constructing an initial mask encoder and an initial decoder; comprising:

[0064] The initial mask encoder is constructed through the local enhancement feature extraction module; the initial decoder is constructed through the mask recovery module; the initial mask encoder is connected to the initial decoder to build a self-supervised learning model.

[0065] In the above embodiments, the self-supervised learning model is constructed by connecting the initial mask encoder and the initial decoder. Specifically, the initial mask encoder is constructed by the local enhancement feature extraction module, and the initial decoder is constructed by the mask recovery module.

[0066] Embodiment 3:

[0067] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising: step: obtaining unlabeled sample data in a training set to pre-train the self-supervised learning model and obtaining a mask encoder; comprising:

[0068] Perform random masking on the unlabeled sample data according to the preset masking rate to obtain sample mask parameter information; optional, including:

[0069] Get the parameter information of unlabeled sample data P = (p1, ..., ps ), where p1 is the parameter information corresponding to the first trajectory point in the unlabeled sample data, p s is the parameter information corresponding to the s-th trajectory point in the unlabeled sample data. There are s trajectory points in the unlabeled sample data.

[0070] Randomly select mask parameter points in the parameter information of unlabeled sample data according to the preset mask rate Perform a mask operation, replace the mask parameter point with the standard parameter vector p0, and obtain the sample mask parameter information P ′ =(p1,…,p0,…,p0,…,p s ), where p k+1 is the parameter vector corresponding to the k+1th trajectory point in the parameter information of the unlabeled sample data, is the parameter vector corresponding to the k+k0th trajectory point in the parameter information of the unlabeled sample data, and k0 is the number of parameter points in the mask parameter point;

[0071] The local enhancement feature extraction module adaptively enhances the local features of the sample mask parameter information and extracts the corresponding unlabeled sample feature vector;

[0072] The mask recovery module is used to recover the trajectory points according to the unlabeled sample feature vector, obtain the coordinate position and stroke position of the masked trajectory points in the unlabeled sample data, and construct the unlabeled recovery sample; the stroke position includes the starting point, transition point and end point.

[0073] In the above embodiments, the parameter information of the unlabeled sample data in the training set is randomly masked according to a preset mask rate to obtain the sample mask parameter information. The local enhancement feature extraction module adaptively enhances the local features of the sample mask parameter information and extracts the corresponding unlabeled sample feature vector. The trajectory points are restored through the mask recovery module to construct an unlabeled recovery sample.

[0074] In the above embodiment, mask parameter points in the parameter information of the unlabeled sample data are randomly selected according to a preset mask rate and replaced with standard parameter vectors to obtain sample mask parameter information.

[0075] In the above embodiments, the mask encoder obtained through training can realize adaptive feature extraction, which is more efficient and effective, and more in line with the self-regulation and self-learning characteristics of the deep learning model.

[0076] The beneficial effects of the above technology are: the parameter information of the unlabeled sample data is randomly masked according to the preset mask rate, thereby realizing the random masking of the unlabeled sample data by the self-supervised learning model, and the feature extraction of the sample mask parameter information by the mask encoder is realized through the local enhancement feature extraction module, and the corresponding unlabeled recovery sample is obtained through the mask recovery module, thereby realizing the training of the self-supervised learning model according to the unlabeled sample data and obtaining the mask encoder.

[0077] Embodiment 4:

[0078] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising the steps of: a local enhancement feature extraction module adaptively enhances the local features of sample mask parameter information in length and extracts the corresponding unlabeled sample feature vector; comprising:

[0079] A local enhanced feature extraction module, including a feature extraction unit and a deformable feature extraction unit;

[0080] The sample mask parameter information is processed through the multi-head attention mechanism layer in the feature extraction unit to construct the query vector matrix, key vector matrix and value vector matrix, and then the corresponding attention value is obtained; the perception layer performs nonlinear feature extraction on the sample mask parameter information according to the attention value through the GELU activation function to obtain the sample mask feature vector;

[0081] The deformable attention mechanism layer in the deformable feature extraction unit processes the sample mask feature vector, obtains the vector attention value in the dynamic window, and performs weighted summation based on the sample mask feature vector and the vector attention value to obtain the weighted attention vector, performs a normalization layer operation to obtain the standard attention vector; the multilayer perceptron performs dynamic feature sampling on the standard feature vector to obtain the unlabeled sample feature vector.

[0082] In the above embodiments, the multi-head attention mechanism layer in the feature extraction unit calculates the attention value by constructing a query vector matrix, a key vector matrix and a value vector matrix, and the perception layer performs nonlinear feature extraction on the mask parameter information through the GELU activation function to obtain the sample mask feature vector; the standard feature vector is processed through the deformable attention mechanism layer, the normalization layer and the multi-layer perceptron in the deformable feature extraction unit to obtain the unlabeled sample feature vector.

[0083] In the above embodiments, the local enhancement feature extraction module includes a plurality of feature extraction units and a deformable feature extraction unit.

[0084] In one embodiment, the local enhanced feature extraction module can be implemented as “several feature extraction units+a deformable feature extraction unit+several feature extraction units”.

[0085] In the above embodiments, the feature extraction unit is built based on the Transformer structure, and the local enhancement feature extraction module is built based on the Deformable Transformer structure.

[0086] In the above embodiment, the multi-head attention mechanism layer in the feature extraction unit processes the sample mask parameter information to construct the query vector matrix W q , key vector matrix W k Sum value vector matrix W v , further calculate the corresponding attention query vector Q, attention key vector K and attention value vector V;

[0087]

[0088] Among them, LN(P ′ ) is the value after normalizing the sample mask parameter information;

[0089] Calculate the attention value S according to the attention query vector Q, the attention key vector K and the attention value vector V;

[0090]

[0091] Among them, K T is the transpose of the attention key vector K, and L is the dimension of the attention query vector Q.

[0092] The beneficial effect of the above technology is that: through the feature extraction unit and the deformable feature extraction unit in the local enhancement feature extraction module, the feature extraction of the sample mask parameter information by the initial mask encoder is realized, wherein the feature extraction unit is used to realize the extraction of global basic features, and the deformable feature extraction unit is used to realize the extraction of local features.

[0093] Embodiment 5:

[0094] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising the following steps: a deformable attention mechanism layer in a deformable feature extraction unit processes a sample mask feature vector to obtain a vector attention value in a dynamic window, and performs weighted summation according to the sample mask feature vector and the vector attention value to obtain a weighted attention vector, and performs a normalization layer operation to obtain a standard attention vector; a multilayer perceptron performs dynamic feature sampling on the standard feature vector to obtain an unlabeled sample feature vector; comprising:

[0095] Perform weighted summation based on the sample mask feature vector and the vector attention value in the dynamic window to obtain a weighted attention vector;

[0096]

[0097] Among them, T is the weighted attention vector obtained by processing the sample mask feature vector, H l is the trainable weight parameter of the lth attention head in the deformable attention mechanism layer, L i is the vector attention value corresponding to the i-th feature vector in the sample mask feature vector, V i is the vector value corresponding to the i-th feature vector in the sample mask feature vector, x i is the trajectory position of the i-th eigenvector, Δx i is the trajectory offset position of the i-th feature vector, n is the number of attention heads in the deformable attention mechanism layer, and m is the number of feature vectors in the dynamic window of the sample mask feature vector;

[0098] Perform a normalization layer operation according to the weighted attention vector to obtain the standard attention vector;

[0099]

[0100] Among them, LN(T i ) is the i-th attention vector element T in the weighted attention vector i The standard attention vector element obtained by normalization, μ i is the weighted attention vector T i The corresponding mean, σ i is the weighted attention vector T i The corresponding standard deviation, ∈ is a constant, α is the first normalization vector parameter of the normalization layer, and β is the second normalization vector parameter of the normalization layer;

[0101] The standard attention vector is obtained by traversing and normalizing the attention vector elements in the weighted attention vector.

[0102] In the above embodiments, the deformable attention mechanism layer obtains the weighted attention vector of the feature vector based on the vector attention value and vector value of each feature vector in the sample mask feature vector, as well as the trajectory position and trajectory offset position corresponding to the feature vector, combined with the trainable weight parameters of the attention head.

[0103] In the above embodiments, the deformable attention mechanism layer maps the sample mask feature vector to obtain the query vector, key vector and value vector, and adjusts the attention weight distribution by calculating the weights through vector interaction, wherein the key vector points to the position of the feature point near each vector in the query vector.

[0104] In the above embodiments, the normalization layer performs normalization processing on each attention vector element in the weighted attention vector, obtains the corresponding standard attention vector element, and constructs the standard attention vector corresponding to the weighted attention vector.

[0105] The beneficial effects of the above technology are: the deformable attention mechanism layer can achieve more flexible adjustment of attention weights, so that while obtaining the weighted attention vector, the amount of calculation and complexity can be effectively reduced, and the normalization layer is used for normalization processing, which facilitates the subsequent steps. The multi-layer perceptron performs processing according to the standard attention vector, thereby realizing the dynamic sampling mechanism of the deformable feature extraction unit, which is used to deal with the complex changes of trajectory points in the unlabeled sample data.

[0106] Embodiment 6:

[0107] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising the following steps: obtaining unlabeled sample data in a training set to pre-train the self-supervised learning model and obtaining a mask encoder; and then, further comprising:

[0108] Set up the trajectory discriminator;

[0109] The unlabeled restored samples are identified by the trajectory identifier to determine whether the unlabeled restored samples have abnormal trajectory data. When it is determined that the unlabeled restored samples have abnormal trajectory data, the control parameters of the mask encoder are optimized.

[0110] In the above embodiments, a trajectory identifier is set to identify the unlabeled restored samples to determine whether the restored trajectory point parameters contain abnormal trajectory data. When it is determined that abnormal trajectory data exists, the control parameters of the mask encoder are optimized.

[0111] In the above embodiments, the mask encoder is optimized by using a trajectory identifier to improve the feature extraction performance of the mask encoder.

[0112] In the above embodiments, the trajectory discriminator promotes the mask encoder and decoder to further optimize parameters to improve the restoration effect of unlabeled restoration samples.

[0113] In the above embodiments, the trajectory discriminator is committed to distinguishing the feature distribution of real sample trajectory points and mask-restored sample trajectory points, and the mask encoder and decoder are committed to restoring the mask sample effect to be close to the real sample to confuse the trajectory discriminator. The two are iteratively trained through adversarial training to further optimize the parameters of the local enhancement feature extraction module and the mask recovery module of the mask encoder.

[0114] The beneficial effect of the above technology is that by setting the trajectory discriminator, the abnormal trajectory data in the unlabeled recovery sample can be judged, and the mask encoder can be optimized to improve the feature extraction performance of the mask encoder.

[0115] Embodiment 7:

[0116] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising the steps of: migrating the parameters of a mask encoder in the self-supervised learning model to an encoder of an end-to-end classification model, and building an initial end-to-end classification model by connecting a classification head, and training the initial end-to-end classification model with a small amount of labeled sample data in a training set to obtain an end-to-end classification model; comprising:

[0117] Get labeled sample data in the training set;

[0118] Input the labeled samples into the initial end-to-end classification model, and the initial end-to-end classification model obtains the corresponding label recognition results;

[0119] The label recognition result is compared with the standard recognition result corresponding to the labeled sample data, the classification head of the initial end-to-end classification model is trained, the control parameters of the classification head are optimized, and the end-to-end classification model is obtained.

[0120] In the above embodiments, the labeled sample data is input into the initial end-to-end classification model, and the corresponding label recognition result is output; the label recognition result is judged against the standard recognition result corresponding to the labeled sample data, the classification head of the initial end-to-end classification model is trained, and the control parameters of the classification head are optimized.

[0121] The beneficial effects of the above technology are: through labeled sample data, supervised fine-tuning training of the classification head in the initial end-to-end classification model is achieved, effectively improving the trajectory point recognition performance of the classification head.

[0122] Embodiment 8:

[0123] An embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising:

[0124] When the initial end-to-end classification model is trained using labeled sample data in the training set, there is no need to train the encoder in the initial end-to-end classification model.

[0125] In the above embodiments, the encoder in the initial end-to-end classification model is built based on the parameter migration of the mask encoder in the self-supervised learning model, and no further training is required.

[0126] The beneficial effect of the above technology is that while realizing the convenient construction of the encoder, the mask encoder in the self-supervised learning model has a high feature extraction performance after being trained with unlabeled sample data. When the initial end-to-end classification model is trained, only the classification head is trained, thereby effectively avoiding repeated training of the encoder.

[0127] Embodiment 9:

[0128] An embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising:

[0129] The training set contains a large amount of unlabeled sample data and a small amount of labeled sample data;

[0130] The labeled sample data corresponds to a standard recognition result, where the standard recognition result is manually labeled based on the labeled sample data.

[0131] In the above embodiments, a large amount of unlabeled sample data in the training set is used to train the initial mask encoder in the self-supervised learning model, and a small amount of labeled sample data is used to train the classification head of the initial end-to-end classification model.

[0132] The beneficial effects of the above technology are: through a large amount of unlabeled sample data, large-scale training of the initial mask encoder is achieved, which effectively improves the feature extraction performance of the mask encoder; and a small amount of labeled sample data is used to perform supervised fine-tuning training on the classification head, which improves the classification accuracy while avoiding overfitting of training.

[0133] Embodiment 10:

[0134] The embodiment of the present invention provides an online trajectory point recognition method based on a self-supervised learning model, comprising the steps of: extracting features of parameter information of online trajectory points through an encoder in an end-to-end classification model, and performing prediction through a classification head to obtain trajectory point recognition results; comprising:

[0135] The encoder is used to extract the feature information of the online trajectory points to obtain the trajectory feature vector;

[0136] The Transformer layer in the classification head decodes and predicts the trajectory feature vector to obtain the pre-classified high-dimensional feature vector, performs downsampling operation through the pooling layer to obtain the pre-classified low-dimensional feature vector, and performs nonlinear activation according to the pre-classified low-dimensional feature vector through the fully connected layer to output the trajectory point recognition result.

[0137] In the above embodiments, the encoder performs feature extraction on the parameter information of the online trajectory point to obtain the trajectory feature vector, the Transformer layer of the classification head decodes and predicts the trajectory feature vector to obtain the pre-classified high-dimensional feature vector, and performs downsampling operation through the pooling layer to obtain the pre-classified low-dimensional feature vector, and performs nonlinear activation through the fully connected layer to output the trajectory point recognition result.

[0138] The beneficial effect of the above technology is that through the encoder, the parameter information feature extraction of the online trajectory point is realized, and through the Transformer layer, pooling layer and fully connected layer in the classification head, the trajectory point recognition result is obtained, thereby realizing the extraction of the online trajectory point parameter features while effectively improving the accuracy of trajectory point recognition.

[0139] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An online trajectory point recognition method based on a self-supervised learning model, characterized in that: include: Obtain parameter information of online trajectory points within a preset time period; Obtain a self-supervised learning model by constructing an initial mask encoder and an initial decoder; Obtain unlabeled sample data in a training set to pre-train the self-supervised learning model, and obtain a mask encoder; Migrating the parameters of the mask encoder in the self-supervised learning model to the encoder of the end-to-end classification model, and building an initial end-to-end classification model by connecting the classification head, and training the initial end-to-end classification model with a small amount of labeled sample data in the training set to obtain an end-to-end classification model; The encoder in the end-to-end classification model extracts features from the parameter information of the online trajectory points, and predicts the features through the classification head to obtain the trajectory point recognition results.

2. The online trajectory point recognition method based on the self-supervised learning model according to claim 1, characterized in that: The step of obtaining a self-supervised learning model by constructing an initial mask encoder and an initial decoder includes: The initial mask encoder is constructed through the local enhancement feature extraction module; the initial decoder is constructed through the mask recovery module; the initial mask encoder is connected to the initial decoder to build a self-supervised learning model.

3. The online trajectory point recognition method based on the self-supervised learning model according to claim 2 is characterized in that: The step of obtaining unlabeled sample data in a training set to pre-train the self-supervised learning model and obtaining a mask encoder includes: Performing random masking processing on the unlabeled sample data according to a preset masking rate to obtain sample mask parameter information; optionally, including: Get the parameter information P of the unlabeled sample data = (p1, ..., p s ), where p1 is the parameter information corresponding to the first trajectory point in the unlabeled sample data, p s is the parameter information corresponding to the s-th trajectory point in the unlabeled sample data. There are s trajectory points in the unlabeled sample data. Randomly select mask parameter points in the parameter information of unlabeled sample data according to the preset mask rate Perform a mask operation, replace the mask parameter point with the standard parameter vector p0, and obtain the sample mask parameter information P ′ =(p1,…,p0,…,p0,…,p s ), where p k+1 is the parameter vector corresponding to the k+1th trajectory point in the parameter information of the unlabeled sample data, is the parameter vector corresponding to the k+k0th trajectory point in the parameter information of the unlabeled sample data, and k0 is the number of parameter points in the mask parameter point; The local enhancement feature extraction module performs adaptive length enhancement on the local features of the sample mask parameter information and extracts the corresponding unlabeled sample feature vector; The mask recovery module is used to recover the trajectory points according to the unlabeled sample feature vector, obtain the coordinate positions and stroke positions of the masked trajectory points in the unlabeled sample data, and construct an unlabeled recovery sample; the stroke position includes a starting point, a transition point, and an end point.

4. The online trajectory point recognition method based on the self-supervised learning model according to claim 3 is characterized in that: The step: the local enhancement feature extraction module performs adaptive length enhancement on the local features of the sample mask parameter information and extracts the corresponding unlabeled sample feature vector; includes: A local enhanced feature extraction module, including a feature extraction unit and a deformable feature extraction unit; The sample mask parameter information is processed by the multi-head attention mechanism layer in the feature extraction unit to construct a query vector matrix, a key vector matrix and a value vector matrix, thereby obtaining a corresponding attention value; the perception layer performs nonlinear feature extraction on the sample mask parameter information according to the attention value through the GELU activation function to obtain a sample mask feature vector; The deformable attention mechanism layer in the deformable feature extraction unit processes the sample mask feature vector, obtains the vector attention value in the dynamic window, performs weighted summation based on the sample mask feature vector and the vector attention value, obtains the weighted attention vector, performs a normalization layer operation, and obtains the standard attention vector; the multilayer perceptron performs dynamic feature sampling on the standard feature vector to obtain the unlabeled sample feature vector.

5. The online trajectory point recognition method based on the self-supervised learning model according to claim 4 is characterized in that: The steps include: the deformable attention mechanism layer in the deformable feature extraction unit processes the sample mask feature vector, obtains the vector attention value in the dynamic window, performs weighted summation according to the sample mask feature vector and the vector attention value, obtains the weighted attention vector, performs a normalization layer operation, and obtains the standard attention vector; the multi-layer perceptron performs dynamic feature sampling on the standard feature vector to obtain the unlabeled sample feature vector; including: Performing a weighted summation on the sample mask feature vector and the vector attention value in the dynamic window to obtain a weighted attention vector; Among them, T is the weighted attention vector obtained by processing the sample mask feature vector, H l is the trainable weight parameter of the lth attention head in the deformable attention mechanism layer, L i is the vector attention value corresponding to the i-th feature vector in the sample mask feature vector, V i is the vector value corresponding to the i-th feature vector in the sample mask feature vector, x i is the trajectory position of the i-th eigenvector, Δx i is the trajectory offset position of the i-th feature vector, n is the number of attention heads in the deformable attention mechanism layer, and m is the number of feature vectors in the dynamic window of the sample mask feature vector; Performing a normalization layer operation according to the weighted attention vector to obtain a standard attention vector; Among them, LN(T i ) is the i-th attention vector element T in the weighted attention vector i The standard attention vector element obtained by normalization, μ i is the weighted attention vector T i The corresponding mean, σ i is the weighted attention vector T i The corresponding standard deviation, ∈ is a constant, α is the first normalization vector parameter of the normalization layer, and β is the second normalization vector parameter of the normalization layer; The standard attention vector is obtained by traversing and normalizing the attention vector elements in the weighted attention vector.

6. The online trajectory point recognition method based on the self-supervised learning model according to claim 3 is characterized in that: The step: obtaining unlabeled sample data in a training set to pre-train the self-supervised learning model, and obtaining a mask encoder; After that, it also includes: Set up the trajectory discriminator; The unlabeled restored sample is identified by the trajectory identifier to determine whether the unlabeled restored sample has abnormal trajectory data. When it is determined that the unlabeled restored sample has abnormal trajectory data, the control parameters of the mask encoder are optimized.

7. The online trajectory point recognition method based on the self-supervised learning model according to claim 1, characterized in that: The step of migrating the parameters of the mask encoder in the self-supervised learning model to the encoder of the end-to-end classification model, and building an initial end-to-end classification model by connecting the classification head, and training the initial end-to-end classification model with a small amount of labeled sample data in the training set to obtain an end-to-end classification model; includes: Obtaining labeled sample data in the training set; Inputting the labeled samples into the initial end-to-end classification model, and the initial end-to-end classification model obtains corresponding label recognition results; The label recognition result is compared with the standard recognition result corresponding to the labeled sample data, the classification head of the initial end-to-end classification model is trained, the control parameters of the classification head are optimized, and the end-to-end classification model is obtained.

8. The online trajectory point recognition method based on the self-supervised learning model according to claim 7, characterized in that: include: When the initial end-to-end classification model is trained using the labeled sample data in the training set, there is no need to train the encoder in the initial end-to-end classification model.

9. The online trajectory point recognition method based on the self-supervised learning model according to claim 7, characterized in that: include: The training set contains a large amount of unlabeled sample data and a small amount of labeled sample data; The labeled sample data corresponds to a standard recognition result, wherein the standard recognition result is manually annotated based on the labeled sample data.

10. The online trajectory point recognition method based on the self-supervised learning model according to claim 1, characterized in that: The step of extracting features from the parameter information of the online trajectory points through an encoder in an end-to-end classification model and predicting through a classification head to obtain a trajectory point recognition result comprises: Extracting features from parameter information of the online trajectory points through an encoder to obtain a trajectory feature vector; The Transformer layer in the classification head decodes and predicts the trajectory feature vector to obtain a pre-classified high-dimensional feature vector, performs downsampling operation through the pooling layer to obtain a pre-classified low-dimensional feature vector, performs nonlinear activation according to the pre-classified low-dimensional feature vector through the fully connected layer, and outputs the trajectory point recognition result.

Images