Artificial intelligence spatiotemporal data processing method based on large models

Abstract

The present application provides an artificial intelligence spatiotemporal data processing method based on a large model. The method obtains the spatiotemporal trajectory information of the target object, and then determines the velocity feature vector of the target object according to the first spatiotemporal trajectory point and the second spatiotemporal trajectory point sequence in the spatiotemporal trajectory information to form a velocity feature vector sequence corresponding to the spatiotemporal trajectory point sequence, and determines the acceleration feature vector of the target object according to the first spatiotemporal trajectory point, the second spatiotemporal trajectory point and the third spatiotemporal trajectory point in the spatiotemporal trajectory point sequence to form an acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence, thereby determining the predicted position range of the target object at the next time node of the current time node according to the velocity feature vector sequence, the acceleration feature vector sequence and the preset spatiotemporal trajectory prediction model, thereby achieving the technical effect of predicting the position range of the target object at the future time node.

Description

Technical Field

[0001] The present application relates to data processing technology, and in particular to an artificial intelligence spatiotemporal data processing method based on a large model. Background Art

[0002] With the rapid development of the Internet of Things, big data, and artificial intelligence technologies, the processing and analysis of spatiotemporal data are playing an increasingly important role in modern society. Spatiotemporal data, such as vehicle trajectories and crowd flow patterns, is not only highly dynamic and complex but also contains rich information and value.

[0003] Traditional spatiotemporal trajectory prediction methods are usually based on simple linear models or time series analysis methods. These methods have great limitations when dealing with nonlinear and non-stationary spatiotemporal data, and the prediction accuracy is low, which makes it difficult to meet the accuracy requirements in practical applications. Summary of the Invention

[0004] The present application provides an artificial intelligence spatiotemporal data processing method based on a large model, which is used to efficiently extract position change features in spatiotemporal trajectory data, thereby achieving the technical effect of predicting the position range of the target object at future time nodes.

[0005] In a first aspect, the present application provides an artificial intelligence spatiotemporal data processing method based on a large model, comprising:

[0006] Acquire spatiotemporal trajectory information of a target object, the spatiotemporal trajectory information including a spatiotemporal trajectory point sequence, the spatiotemporal trajectory point sequence including a plurality of spatiotemporal trajectory points of the target object, each spatiotemporal trajectory point including a position coordinate and a corresponding timestamp;

[0007] determining a velocity feature vector of the target object according to a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, wherein the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is a space-time trajectory point next to the first space-time trajectory point;

[0008] determining an acceleration feature vector of the target object based on the first, second, and third space-time trajectory points in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence, wherein the third space-time trajectory point is a space-time trajectory point subsequent to the second space-time trajectory point;

[0009] The predicted position range of the target object at the next time node of the current time node is determined according to the velocity feature vector sequence, the acceleration feature vector sequence and the preset space-time trajectory prediction model. The space-time trajectory point corresponding to the current time node is the last space-time trajectory point in the space-time trajectory point sequence. The preset space-time trajectory prediction model is a prediction model trained based on a deep space-time neural network model. The convolutional neural network included in the deep space-time neural network model is used to extract the position change characteristics between each space-time trajectory point in the space-time trajectory point sequence.

[0010] In this approach, the target object's velocity and acceleration feature vectors are determined by comprehensively considering its spatiotemporal trajectory information. This reflects the target object's motion state and provides a basis for subsequent position prediction. A deep spatiotemporal neural network model, particularly its convolutional neural network component, is then used to efficiently extract positional variation features from the spatiotemporal trajectory sequence, thereby predicting the target object's position range at future time points. This not only considers the target object's current velocity and acceleration, but also incorporates the complexity and diversity of its historical motion trajectory.

[0011] This solution, based on a large-scale model design, can process large amounts of spatiotemporal data and meet the needs of real-time applications. Through automated data processing, including the calculation of velocity and acceleration eigenvectors and the use of deep neural networks for prediction, the efficiency of data processing and prediction has been significantly improved.

[0012] Especially when dealing with large-scale spatiotemporal data such as complex traffic networks and crowd flows, the above solution can respond quickly and provide accurate prediction results, providing strong support for decision support, resource management and other aspects.

[0013] Furthermore, through continuous iterative training, the deep spatiotemporal neural network model can gradually optimize its weights, thereby improving prediction accuracy and stability. During training, forward propagation is performed using a preset spatiotemporal trajectory training set, and backpropagation is performed based on the error between the predicted and actual results, continuously adjusting the model parameters. Furthermore, this approach is not only applicable to the position prediction of a single target object but can also be extended to scenarios requiring simultaneous prediction of multiple targets. By appropriately adjusting the model parameters and training dataset, it is possible to predict the position of different target types (such as vehicles, pedestrians, and drones) in different environments (e.g., urban roads, rural roads, and complex terrain).

[0014] Optionally, determining the velocity feature vector of the target object according to the first space-time trajectory point and the second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence includes:

[0015] determining the velocity feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, and the timestamp of the second space-time trajectory point in the space-time trajectory point sequence;

[0016] The speed feature vector sequence corresponding to the space-time trajectory point sequence is generated according to the speed feature vector corresponding to each time node.

[0017] In the above scheme, the velocity feature vector is calculated using the position coordinates of the first and second spatiotemporal trajectory points and their respective timestamps. This calculation method, based directly on the actual position and time information in the spatiotemporal trajectory data, can reflect the target object's movement speed between adjacent time points, thus providing accurate basic data for subsequent analysis and prediction. After determining the velocity feature vector between each adjacent spatiotemporal trajectory point, the above scheme further concatenates these velocity feature vectors in chronological order to form a velocity feature vector sequence. This step ensures the continuity and integrity of the velocity features in the temporal dimension, helping subsequent models better understand the target object's motion pattern. Due to the accurate calculation of the velocity feature vectors and their sequences, subsequent spatiotemporal trajectory prediction models based on these features (such as deep spatiotemporal neural network models) receive more accurate and rich input information. This information enables the model to more accurately capture the target object's motion patterns, thereby providing a more accurate predicted position range when predicting the target object's position at the next time point. The velocity feature vectors and velocity feature vector sequences extracted by this scheme are not only applicable to specific target objects or scenes, but can also, to a certain extent, span different spatiotemporal trajectory datasets. This feature enhances the robustness and generalization capability of the entire processing system, enabling it to be more widely used in various spatiotemporal data processing and prediction tasks.

[0018] Optionally, determining the acceleration feature vector of the target object based on the first space-time trajectory point, the second space-time trajectory point, and the third space-time trajectory point in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence includes:

[0019] determining the acceleration feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, the timestamp of the second space-time trajectory point, the position coordinates of the third space-time trajectory point, and the timestamp of the third space-time trajectory point in the space-time trajectory point sequence;

[0020] The acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence is generated according to the acceleration feature vector corresponding to each time node.

[0021] Optionally, the deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, and the convolutional neural network layer is connected to the recurrent neural network layer; correspondingly, determining the predicted position range of the target object at the next time node of the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence, and the preset spatiotemporal trajectory prediction model includes:

[0022] Inputting the velocity feature vector sequence and the acceleration feature vector sequence into the convolutional neural network layer, so that the convolutional neural network extracts the behavioral feature values ​​of the spatiotemporal trajectory point sequence, and outputs a feature map corresponding to the spatiotemporal trajectory point sequence;

[0023] The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

[0024] In the above scheme, by inputting the sequence of velocity feature vectors and acceleration feature vectors into the convolutional neural network layer of the deep spatiotemporal neural network model, the behavioral feature values ​​in the spatiotemporal trajectory point sequence are efficiently extracted and the corresponding feature map is generated. Convolutional neural networks excel at capturing local features and automatically learning complex spatial hierarchical structures, which helps to more comprehensively understand the motion pattern of the target object. At the same time, the feature map and the spatiotemporal trajectory point sequence are further input into the recurrent neural network layer, achieving an effective integration of spatial features and temporal information, thereby improving the accuracy of prediction. Moreover, by combining the advantages of convolutional neural networks and recurrent neural networks, the deep spatiotemporal neural network model can not only capture the spatial relationship between spatiotemporal trajectory points, but also handle temporal dependencies. This enhanced spatiotemporal perception capability enables the model to more accurately understand the complex motion patterns of the target object in time and space, thereby improving the accuracy of the predicted position range.

[0025] Furthermore, using a deep spatiotemporal neural network model to process the velocity and acceleration feature vector sequences can deeply explore the potential information in the data and reduce the impact of noise and interference on the prediction results. At the same time, through continuous learning and iterative optimization, the model can gradually adapt to the changes in different target objects and scenes, improving the accuracy and robustness of the prediction. In addition, for target objects with complex motion patterns, such as variable speed motion and curved motion, this method can more accurately capture and predict changes in their motion state by introducing acceleration feature vector sequences and a deep spatiotemporal neural network model. This makes the method more adaptable and more practical in practical applications.

[0026] It is worth noting that although the deep spatiotemporal neural network model may require higher computing resources and time during the training phase, once the model training is completed and deployed in actual applications, its efficient feature extraction and prediction capabilities can significantly improve the processing speed and optimize the utilization of computing resources, which is especially important for scenarios that require real-time processing of large amounts of spatiotemporal trajectory data.

[0027] Optionally, before extracting the behavior feature value of the spatiotemporal trajectory point sequence to output a feature map corresponding to the spatiotemporal trajectory point sequence, the method further includes:

[0028] Step a: When training the convolutional neural network, randomly assigning an initial weight value to the expansion offset of the convolution kernel in the convolutional neural network;

[0029] Step b: inputting a preset spatiotemporal trajectory training set into the convolutional neural network and performing forward conduction through the convolutional neural network to output a training predicted position range, wherein the preset spatiotemporal trajectory training set includes a labeled spatiotemporal trajectory information set and a labeled actual spatiotemporal trajectory point corresponding to each labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set, and the labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set includes a labeled velocity feature vector sequence and a labeled acceleration feature vector sequence corresponding to the labeled spatiotemporal trajectory point sequence;

[0030] Step c: determining a position prediction error based on the training predicted position range and the actual space-time trajectory point, wherein the position prediction error is used to represent the distance between the coordinates of the center point of the training predicted position range and the position coordinates of the actual space-time trajectory point;

[0031] Step d, calculating the characteristic gradient of the current weight value of the extended offset of the convolution kernel according to the position prediction error, and updating the next iterative weight value of the convolution kernel;

[0032] Repeat steps b to d until the determined position prediction error is less than a preset position prediction error threshold or the current number of iterations reaches the maximum number of iterations, and then determine the latest weight value as the weight of the extended offset of the convolution kernel.

[0033] In the above scheme, by randomly assigning initial weights to the convolution kernel's expansion offset, the model's randomness and diversity are increased, helping the model better learn the inherent patterns of the data during subsequent training, thereby improving prediction accuracy. Then, by repeatedly performing the iterative training process from steps b to d, the convolution kernel's weights are continuously adjusted based on the error between the predicted and actual positions. This process enables the model to gradually reduce prediction errors, improving prediction accuracy and robustness. Furthermore, the convolution kernel's expansion offset is continuously adjusted and optimized during training, enabling it to more accurately capture the positional variation characteristics of the spatiotemporal trajectory point sequence. This optimization enhances the convolutional neural network's ability to process complex spatiotemporal data and improves the efficiency and accuracy of feature extraction. The optimized convolutional neural network can more accurately extract behavioral features from the velocity and acceleration feature vector sequences and generate feature maps. These feature maps provide rich input information for subsequent recurrent neural network layers, helping to improve overall prediction performance. Then, the feature gradient of the position prediction error with respect to the current weight value is calculated using gradient descent, and the weight values ​​are updated accordingly, accelerating model convergence. This method effectively reduces oscillation during training and improves training efficiency. Training automatically terminates when the position prediction error falls below a preset threshold or when the current iteration count reaches the maximum. This precondition avoids overfitting, which can be caused by excessive training, while also conserving computing resources.

[0034] In addition, by training with a preset spatiotemporal trajectory training set containing a set of annotated spatiotemporal trajectory information and its corresponding annotated actual spatiotemporal trajectory points, the model can learn the characteristics of various complex spatiotemporal data. This training method improves the model's adaptability and generalization ability to complex spatiotemporal data.

[0035] In summary, the training process in the above scheme significantly improves the overall performance and prediction accuracy of the artificial intelligence spatiotemporal data processing method based on large models through iterative optimization of the weight values ​​of the convolutional neural network, enhanced feature extraction capabilities, accelerated training speed, and adaptation to complex spatiotemporal data.

[0036] Optionally, before extracting the behavior feature value of the spatiotemporal trajectory point sequence to output a feature map corresponding to the spatiotemporal trajectory point sequence, the method further includes:

[0037] The value of the input data at the expanded position is determined based on the velocity feature vector sequence and the acceleration feature vector sequence, wherein the value of the input data at the expanded position is used to determine the feature map, and the input data includes the velocity feature vector sequence and the acceleration feature vector sequence.

[0038] In this approach, by calculating the input data in the velocity and acceleration feature vector sequences at expanded locations, a richer and more detailed feature representation can be generated. This expansion not only considers the original data point but also encompasses information within its neighborhood, thereby enhancing the data representation capability. The values ​​of the input data at these expanded locations directly influence the subsequent feature map generation. By carefully calculating the values ​​at these expanded locations, the feature map can more accurately reflect the complex changes in the spatiotemporal trajectory point sequence, providing stronger support for subsequent prediction tasks.

[0039] The expanded input data can capture more spatiotemporal trajectory details. For example, at locations where there is a sudden change in velocity or acceleration, the expanded data can more accurately reflect this change, thus avoiding the loss of important information during feature extraction. By calculating the values ​​at the expanded locations, the error propagation caused by inaccurate data representation can be reduced. In the subsequent convolutional neural network and recurrent neural network processing, these accurate feature maps will help improve the accuracy and reliability of predictions. In addition, the expanded input data can contain more redundant information, which helps the model remain robust in the face of noise or missing data. Even if some input data contains errors, the model can compensate and correct them using information from other expanded locations.

[0040] It is worth noting that although expanding the input data increases the complexity of data processing, in the subsequent feature extraction and prediction process, due to the improved quality of the feature map, the model may require fewer iterations to achieve the same prediction accuracy, which optimizes the computational efficiency to a certain extent.

[0041] In summary, the value of the input data at the expanded position is determined according to the velocity eigenvector sequence and the acceleration eigenvector sequence. By enhancing data representation capabilities, improving prediction accuracy, enhancing model generalization capabilities, and optimizing computational efficiency, the proposed method provides strong support for artificial intelligence spatiotemporal data processing methods based on large models.

[0042] Optionally, inputting the feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence into the recurrent neural network layer to output the predicted position range includes:

[0043] Inputting the feature graph corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence into the recurrent neural network layer to determine the predicted position point, wherein the recurrent neural network layer is used to capture the dependency between the position information in the feature graph and the time information in the spatiotemporal trajectory point sequence;

[0044] The predicted position range of the next time node is determined according to the predicted position point and the preset position extension range.

[0045] In the above scheme, by simultaneously inputting the feature map (containing location information) and the timestamp sequence (containing time information) into the recurrent neural network layer, this layer can accurately capture and fuse the dependency between the location information in the feature map and the time information in the spatiotemporal trajectory point sequence. This spatiotemporal fusion capability is the key to accurately predicting the future location of the target object.

[0046] It’s worth noting that recurrent neural networks are particularly adept at processing temporal dependencies in sequential data. They can combine previous state information with current input to dynamically model how the target object’s position changes over time. This is particularly important for processing spatiotemporal trajectory data with complex dynamic characteristics.

[0047] In the recurrent neural network layer, not only the position features in the feature map are considered, but also the time features in the timestamp sequence are integrated, realizing the deep fusion of multi-dimensional features. This multi-feature fusion strategy helps to reduce information loss and improve prediction accuracy.

[0048] Furthermore, through comprehensive analysis of feature maps and timestamp sequences, the recurrent neural network can generate more accurate predicted locations. Furthermore, by combining the predicted location range with a pre-set location extension range, the uncertainty of a single prediction point can be reduced to a certain extent, improving the reliability of the prediction.

[0049] Furthermore, recurrent neural networks can process spatiotemporal trajectory data with irregular time intervals, without requiring additional time alignment or interpolation. This feature makes the method applicable to a wider range of practical application scenarios and improves the applicability of the model.

[0050] Furthermore, due to its ability to capture complex spatiotemporal dependencies, this approach excels at processing spatiotemporal trajectory data in complex scenarios. Whether it involves rapid changes in the target object's speed, sudden turns in its trajectory, or other irregular motion patterns, it can achieve relatively accurate prediction results.

[0051] It's worth noting that the above solution adopts a modular design concept, separating feature extraction (convolutional neural network layer) and time series modeling (recurrent neural network layer) into two independent modules. This design not only allows each module to be independently optimized and improved, but also facilitates integration and expansion with other models or algorithms.

[0052] In addition, by adjusting the size of the preset position extension range, the size and accuracy of the predicted position range can be flexibly controlled according to actual needs. This customizable feature makes this method more flexible and practical.

[0053] In summary, the technical effect of inputting the feature map and timestamp sequence corresponding to the spatiotemporal trajectory point sequence into the recurrent neural network layer to output the predicted position range is mainly reflected in the accurate capture of spatiotemporal dependencies, improved prediction accuracy, strong adaptability and strong scalability.

[0054] In a second aspect, the present application provides an artificial intelligence spatiotemporal data processing device based on a large model, comprising:

[0055] an acquisition module, configured to acquire spatiotemporal trajectory information of a target object, wherein the spatiotemporal trajectory information includes a spatiotemporal trajectory point sequence, wherein the spatiotemporal trajectory point sequence includes a plurality of spatiotemporal trajectory points of the target object, and each spatiotemporal trajectory point includes a position coordinate and a corresponding timestamp;

[0056] a processing module, configured to determine a velocity feature vector of the target object based on a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence, so as to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, wherein the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is a space-time trajectory point next to the first space-time trajectory point;

[0057] The processing module is further configured to determine an acceleration feature vector of the target object based on the first space-time trajectory point, the second space-time trajectory point, and a third space-time trajectory point in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence, wherein the third space-time trajectory point is a space-time trajectory point subsequent to the second space-time trajectory point;

[0058] The processing module is also used to determine the predicted position range of the target object at the next time node of the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence and the preset space-time trajectory prediction model. The space-time trajectory point corresponding to the current time node is the last space-time trajectory point in the space-time trajectory point sequence. The preset space-time trajectory prediction model is a prediction model trained based on a deep space-time neural network model. The convolutional neural network included in the deep space-time neural network model is used to extract the position change characteristics between each space-time trajectory point in the space-time trajectory point sequence.

[0059] Optionally, the processing module is specifically configured to:

[0060] The determining the velocity feature vector of the target object according to the first space-time trajectory point and the second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence includes:

[0061] determining the velocity feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, and the timestamp of the second space-time trajectory point in the space-time trajectory point sequence;

[0062] The speed feature vector sequence corresponding to the space-time trajectory point sequence is generated according to the speed feature vector corresponding to each time node.

[0063] Optionally, the processing module is specifically configured to:

[0064] determining the acceleration feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, the timestamp of the second space-time trajectory point, the position coordinates of the third space-time trajectory point, and the timestamp of the third space-time trajectory point in the space-time trajectory point sequence;

[0065] The acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence is generated according to the acceleration feature vector corresponding to each time node.

[0066] Optionally, the deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, and the convolutional neural network layer is connected to the recurrent neural network layer; optionally, the processing module is specifically used to:

[0067] Inputting the velocity feature vector sequence and the acceleration feature vector sequence into the convolutional neural network layer, so that the convolutional neural network extracts the behavioral feature values ​​of the spatiotemporal trajectory point sequence, and outputs a feature map corresponding to the spatiotemporal trajectory point sequence;

[0068] The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

[0069] Optionally, the processing module is specifically configured to:

[0070] Step a: When training the convolutional neural network, randomly assigning an initial weight value to the expansion offset of the convolution kernel in the convolutional neural network;

[0071] Step b: inputting a preset spatiotemporal trajectory training set into the convolutional neural network and performing forward conduction through the convolutional neural network to output a training predicted position range, wherein the preset spatiotemporal trajectory training set includes a labeled spatiotemporal trajectory information set and a labeled actual spatiotemporal trajectory point corresponding to each labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set, and the labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set includes a labeled velocity feature vector sequence and a labeled acceleration feature vector sequence corresponding to the labeled spatiotemporal trajectory point sequence;

[0072] Step c: determining a position prediction error based on the training predicted position range and the actual space-time trajectory point, wherein the position prediction error is used to represent the distance between the coordinates of the center point of the training predicted position range and the position coordinates of the actual space-time trajectory point;

[0073] Step d, calculating the characteristic gradient of the current weight value of the extended offset of the convolution kernel according to the position prediction error, and updating the next iterative weight value of the convolution kernel;

[0074] Repeat steps b to d until the determined position prediction error is less than a preset position prediction error threshold or the current number of iterations reaches the maximum number of iterations, and then determine the latest weight value as the weight of the extended offset of the convolution kernel.

[0075] Optionally, the processing module is specifically configured to:

[0076] The value of the input data at the expanded position is determined based on the velocity feature vector sequence and the acceleration feature vector sequence, wherein the value of the input data at the expanded position is used to determine the feature map, and the input data includes the velocity feature vector sequence and the acceleration feature vector sequence.

[0077] Optionally, the processing module is specifically configured to:

[0078] Inputting the feature graph corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence into the recurrent neural network layer to determine the predicted position point, wherein the recurrent neural network layer is used to capture the dependency between the position information in the feature graph and the time information in the spatiotemporal trajectory point sequence;

[0079] The predicted position range of the next time node is determined according to the predicted position point and the preset position extension range.

[0080] In a third aspect, the present application provides an electronic device, comprising:

[0081] processor; and,

[0082] a memory for storing executable instructions of the processor;

[0083] The processor is configured to perform any possible method described in the first aspect by executing the executable instructions.

[0084] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement any possible method described in the first aspect.

[0085] The large-model-based artificial intelligence spatiotemporal data processing method provided in the present application obtains the spatiotemporal trajectory information of the target object, and then determines the velocity feature vector of the target object based on the first spatiotemporal trajectory point and the second spatiotemporal trajectory point sequence in the spatiotemporal trajectory information to form a velocity feature vector sequence corresponding to the spatiotemporal trajectory point sequence, and determines the acceleration feature vector of the target object based on the first spatiotemporal trajectory point, the second spatiotemporal trajectory point and the third spatiotemporal trajectory point in the spatiotemporal trajectory point sequence to form an acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence, thereby determining the predicted position range of the target object at the next time node of the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence and the preset spatiotemporal trajectory prediction model, thereby achieving efficient extraction of position change features in the spatiotemporal trajectory point sequence, thereby predicting the position range of the target object at future time nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0087] Figure 1 This is a flowchart of a method for processing spatiotemporal data using artificial intelligence based on a large model according to an exemplary embodiment of the present application;

[0088] Figure 2 This is a flowchart of an artificial intelligence spatiotemporal data processing method based on a large model according to another exemplary embodiment of the present application;

[0089] Figure 3 1 is a schematic structural diagram of an artificial intelligence spatiotemporal data processing device based on a large model according to an exemplary embodiment of the present application;

[0090] Figure 4 It is a structural diagram of an electronic device according to an exemplary embodiment of the present application.

[0091] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0092] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0093] To address the above issues, the embodiments provided in this application obtain the spatiotemporal trajectory information of the target object as input data. The spatiotemporal trajectory information includes a series of spatiotemporal trajectory points, each of which records not only the location coordinates of the target object but also the corresponding timestamp information. This information provides the basis for subsequent velocity and acceleration feature extraction.

[0094] Next, feature extraction is performed on the spatiotemporal trajectory information. The specific steps include:

[0095] Velocity feature vector extraction: Based on any two adjacent space-time trajectory points (the first and second space-time trajectory points) in the space-time trajectory point sequence, the velocity feature vector is calculated using a specific formula using the position coordinates and timestamp information between the two points. This process is repeated to form a sequence of velocity feature vectors corresponding to the entire space-time trajectory point sequence.

[0096] Acceleration feature vector extraction: Furthermore, considering three consecutive spacetime trajectory points (the first, second, and third spacetime trajectory points) in the spacetime trajectory sequence, we use their position coordinates and timestamp information to calculate the acceleration feature vector using a specific formula. This process is repeated to form a sequence of acceleration feature vectors.

[0097] These feature vector sequences accurately reflect the motion state of the target object and provide key input for subsequent position prediction.

[0098] The extracted velocity feature vector sequence and acceleration feature vector sequence are then input into a preset spatiotemporal trajectory prediction model built based on a deep spatiotemporal neural network model. The prediction model includes at least one convolutional neural network layer and at least one recurrent neural network layer:

[0099] Convolutional neural network layer: This layer extracts the positional variation features in the spatiotemporal trajectory sequence and generates a feature map through convolution operations. During training, the accuracy and efficiency of feature extraction are optimized by continuously adjusting the convolution kernel's expansion offset and weight values.

[0100] Recurrent Neural Network Layer: This layer takes the feature map and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence as input, captures the dependency between the location information in the feature map and the time information in the spatiotemporal trajectory point sequence, and outputs the predicted location of the target object at a future time point. The predicted location range is determined by combining this with the preset location expansion range.

[0101] To improve the accuracy and stability of predictions, the deep spatiotemporal neural network model in this application is continuously optimized through iterative training. The training process includes:

[0102] Use the preset spatiotemporal trajectory training set for forward conduction and output the training prediction position range.

[0103] The position prediction error is calculated based on the error between the predicted position range and the actual space-time trajectory point.

[0104] The gradient descent method is used to adjust the model parameters, including the expansion offset weight value of the convolution kernel, until the position prediction error is less than the preset threshold or the maximum number of iterations is reached.

[0105] In addition, during the actual prediction process, the model continuously obtains the actual spatiotemporal trajectory points of the target object, calculates the actual position prediction error, and adds newly annotated spatiotemporal trajectory information to the training set as needed to continuously iterate and optimize the model.

[0106] Through the above solution, the specific technical effects that can be achieved by the technical solution provided by this application include:

[0107] Improve prediction accuracy: By comprehensively considering the speed and acceleration characteristics of the target object and combining it with a deep spatiotemporal neural network model, high-precision position prediction can be achieved.

[0108] Enhanced robustness and generalization capabilities: The model is continuously optimized through iterative training and can handle different types of target objects and complex and changing scenarios.

[0109] Adaptable to large-scale spatiotemporal data processing: Based on large-scale model design, it can efficiently process large-scale spatiotemporal trajectory data and meet application scenarios with high real-time requirements.

[0110] Modular design: facilitates integration and expansion with other models or algorithms, improving the flexibility and customizability of the system.

[0111] Figure 1This is a flow chart of a method for planning a weeding task path based on data processing according to an exemplary embodiment of the present application. Figure 1 As shown, the method provided in this embodiment includes:

[0112] S101: Acquire the spatiotemporal trajectory information of the target object.

[0113] In this step, the spatiotemporal trajectory information of the target object is obtained, wherein the spatiotemporal trajectory information includes a spatiotemporal trajectory point sequence, the spatiotemporal trajectory point sequence includes multiple spatiotemporal trajectory points of the target object, and each spatiotemporal trajectory point includes a position coordinate and a corresponding timestamp.

[0114] Specifically, in practical applications, we first need to collect the target object's spatiotemporal trajectory information. This information usually exists in the form of a dataset, containing the target object's location coordinates (such as longitude and latitude) at different time points and the corresponding timestamps. To process this data, we first need to define a data structure to store this information, such as using a list or array to store a sequence of spatiotemporal trajectory points, each of which contains location coordinates and a timestamp.

[0115] S102 : Determine a velocity feature vector of the target object according to the first space-time trajectory point and the second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence.

[0116] A velocity feature vector of the target object is determined according to a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, where the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is the space-time trajectory point next to the first space-time trajectory point.

[0117] Specifically, the velocity feature vector of the target object is calculated based on adjacent spatiotemporal trajectory points in the spatiotemporal trajectory point sequence. The velocity feature vector can be obtained by dividing the position difference between two points by the time difference. This calculation is performed for each pair of adjacent points through the entire spatiotemporal trajectory point sequence to generate a sequence of velocity feature vectors.

[0118] In a possible implementation, the velocity feature vector may be determined by using Formula 1 and the first and second space-time trajectory points in the space-time trajectory point sequence: , Formula 1 is:

[0119] in, is the position coordinate of the first space-time trajectory point, is the timestamp of the first space-time trajectory point, is the position coordinate of the second space-time trajectory point, is the timestamp of the second space-time trajectory point;

[0120] According to the speed characteristic vector corresponding to each time node Generate the velocity feature vector sequence corresponding to the space-time trajectory point sequence ,in, is the number of space-time trajectory points in the space-time trajectory point sequence.

[0121] S103 , determining an acceleration feature vector of the target object according to the first space-time trajectory point, the second space-time trajectory point, and the third space-time trajectory point in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence.

[0122] The acceleration characteristic vector of the target object is determined according to the first space-time trajectory point, the second space-time trajectory point, and the third space-time trajectory point in the space-time trajectory point sequence to form an acceleration characteristic vector sequence corresponding to the space-time trajectory point sequence, where the third space-time trajectory point is the space-time trajectory point next to the second space-time trajectory point.

[0123] Specifically, the acceleration eigenvector of the target object is calculated based on three consecutive spatiotemporal trajectory points in the sequence. The acceleration eigenvector can be obtained by calculating the rate of change of velocity (i.e., the velocity difference divided by the time difference). This calculation is performed for each consecutive set of three points, traversing the sequence (ignoring the first two points as there are not enough points before them to calculate the acceleration), to generate a sequence of acceleration eigenvectors.

[0124] In one possible implementation, the acceleration characteristic vector is determined based on the first, second, and third space-time trajectory points in the space-time trajectory point sequence using Formula 2: , Formula 2 is:

[0125] in, is the position coordinate of the third space-time trajectory point, is the timestamp of the third space-time trajectory point;

[0126] According to the acceleration characteristic vector corresponding to each time node Generate the acceleration feature vector sequence corresponding to the space-time trajectory point sequence .

[0127] S104: Determine a predicted position range of the target object at the next time node of the current time node according to the velocity feature vector sequence, the acceleration feature vector sequence, and a preset spatiotemporal trajectory prediction model.

[0128] In this step, the predicted position range of the target object at the next time node of the current time node is determined based on the velocity feature vector sequence, the acceleration feature vector sequence and the preset space-time trajectory prediction model. The space-time trajectory point corresponding to the current time node is the last space-time trajectory point in the space-time trajectory point sequence. The preset space-time trajectory prediction model is a prediction model trained based on a deep space-time neural network model. The convolutional neural network included in the deep space-time neural network model is used to extract the position change characteristics between each space-time trajectory point in the space-time trajectory point sequence.

[0129] Specifically, the velocity and acceleration feature vector sequences are fed into a pre-defined deep spatiotemporal neural network model for prediction. This model, trained using a training set of labeled spatiotemporal trajectories, captures the positional variations between trajectory points and outputs the predicted position range of the target object at future time points.

[0130] In one possible implementation, the deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, and the convolutional neural network layer is connected to the recurrent neural network layer. Specifically, the velocity feature vector sequence and the acceleration feature vector sequence are input to the convolutional neural network layer, so that the convolutional neural network uses Formula 3 to extract the behavioral feature values ​​of the spatiotemporal trajectory point sequence and output a feature map corresponding to the spatiotemporal trajectory point sequence. Formula 3 is:

[0131] in, The feature map is at position The behavioral characteristic value of is the preset activation function, is the size of the convolution kernel of the convolutional neural network, and is an odd number, is the expansion offset of the convolution kernel The weight of The position of the input data after expansion The input data includes the velocity feature vector sequence and the acceleration feature vector sequence. is the preset bias item;

[0132] The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

[0133] It is worth noting that the above formula 3 effectively extracts local features from the input data through convolution operations. Since the input data contains a sequence of velocity and acceleration feature vectors of the target object, the convolution operation can capture pattern changes in these sequences, such as sudden changes in velocity or acceleration, thereby generating a feature map that can characterize the motion characteristics of the target object. Among them, the convolution kernel performs a sliding window operation on the input data, and extracts spatial features at different levels by taking the dot product of the convolution kernel weights and the local area of ​​the input data. As the number of convolution layers increases, the network can learn increasingly abstract and complex feature representations, which helps the model better understand the motion patterns of the target object. In addition, the convolution kernel in formula 3 is shared across all positions of the input data, which greatly reduces the number of model parameters, reduces the risk of overfitting, and improves computational efficiency. By sharing parameters, the model can apply the same feature detector to the entire input data, thereby identifying similar features at different positions.

[0134] Furthermore, the convolution operation is translationally invariant, meaning that the kernel can detect a feature regardless of where it appears in the input data. This property makes the model insensitive to the specific location of the target object in the spatiotemporal trajectory, improving the model's generalization capabilities.

[0135] Furthermore, the activation function (such as ReLU) in the above formula 3 can introduce nonlinear factors, enabling the model to learn the complex mapping relationship between input data and output features, thereby helping to capture the nonlinear motion characteristics of the target object.

[0136] Although Formula 3 itself is aimed at spatial feature extraction, in the overall model, these spatial features will be further integrated with temporal information (such as timestamp sequences) in subsequent recurrent neural network (RNN) layers to achieve comprehensive capture and modeling of spatiotemporal information.

[0137] In this embodiment, the spatiotemporal trajectory information of the target object is obtained, and then the velocity feature vector of the target object is determined according to the first spatiotemporal trajectory point and the second spatiotemporal trajectory point sequence in the spatiotemporal trajectory information to form a velocity feature vector sequence corresponding to the spatiotemporal trajectory point sequence, and the acceleration feature vector of the target object is determined according to the first spatiotemporal trajectory point, the second spatiotemporal trajectory point and the third spatiotemporal trajectory point in the spatiotemporal trajectory point sequence to form an acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence, thereby determining the predicted position range of the target object at the next time node of the current time node according to the velocity feature vector sequence, the acceleration feature vector sequence and the preset spatiotemporal trajectory prediction model, thereby achieving efficient extraction of position change features in the spatiotemporal trajectory point sequence, thereby predicting the position range of the target object at future time nodes.

[0138] Figure 2This is a flow chart of an artificial intelligence spatiotemporal data processing method based on a large model according to another exemplary embodiment of the present application. Figure 2 As shown, the artificial intelligence spatiotemporal data processing method based on a large model provided in this embodiment includes:

[0139] S201: Acquire the spatiotemporal trajectory information of the target object.

[0140] In this step, the spatiotemporal trajectory information of the target object is obtained, wherein the spatiotemporal trajectory information includes a spatiotemporal trajectory point sequence, the spatiotemporal trajectory point sequence includes multiple spatiotemporal trajectory points of the target object, and each spatiotemporal trajectory point includes a position coordinate and a corresponding timestamp.

[0141] Specifically, in practical applications, we first need to collect the target object's spatiotemporal trajectory information. This information usually exists in the form of a dataset, containing the target object's location coordinates (such as longitude and latitude) at different time points and the corresponding timestamps. To process this data, we first need to define a data structure to store this information, such as using a list or array to store a sequence of spatiotemporal trajectory points, each of which contains location coordinates and a timestamp.

[0142] S202 : Determine a velocity feature vector of the target object according to the first space-time trajectory point and the second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence.

[0143] A velocity feature vector of the target object is determined according to a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, where the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is the space-time trajectory point next to the first space-time trajectory point.

[0144] Specifically, the velocity feature vector of the target object is calculated based on adjacent spatiotemporal trajectory points in the spatiotemporal trajectory point sequence. The velocity feature vector can be obtained by dividing the position difference between two points by the time difference. This calculation is performed for each pair of adjacent points through the entire spatiotemporal trajectory point sequence to generate a sequence of velocity feature vectors.

[0145] In a possible implementation, the velocity feature vector may be determined by using Formula 1 and the first and second space-time trajectory points in the space-time trajectory point sequence: , Formula 1 is:

[0146] in, is the position coordinate of the first space-time trajectory point, is the timestamp of the first space-time trajectory point, is the position coordinate of the second space-time trajectory point, is the timestamp of the second space-time trajectory point;

[0147] According to the speed characteristic vector corresponding to each time node Generate the velocity feature vector sequence corresponding to the space-time trajectory point sequence ,in, is the number of space-time trajectory points in the space-time trajectory point sequence.

[0148] S203 , determining the acceleration feature vector of the target object according to the first space-time trajectory point, the second space-time trajectory point, and the third space-time trajectory point in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence.

[0149] The acceleration characteristic vector of the target object is determined according to the first space-time trajectory point, the second space-time trajectory point, and the third space-time trajectory point in the space-time trajectory point sequence to form an acceleration characteristic vector sequence corresponding to the space-time trajectory point sequence, where the third space-time trajectory point is the space-time trajectory point next to the second space-time trajectory point.

[0150] Specifically, the acceleration eigenvector of the target object is calculated based on three consecutive spatiotemporal trajectory points in the sequence. The acceleration eigenvector can be obtained by calculating the rate of change of velocity (i.e., the velocity difference divided by the time difference). This calculation is performed for each consecutive set of three points, traversing the sequence (ignoring the first two points as there are not enough points before them to calculate the acceleration), to generate a sequence of acceleration eigenvectors.

[0151] In one possible implementation, the acceleration characteristic vector is determined based on the first, second, and third space-time trajectory points in the space-time trajectory point sequence using Formula 2: , Formula 2 is:

[0152] in, is the position coordinate of the third space-time trajectory point, is the timestamp of the third space-time trajectory point;

[0153] According to the acceleration characteristic vector corresponding to each time node Generate the acceleration feature vector sequence corresponding to the space-time trajectory point sequence .

[0154] S204: Input the velocity feature vector sequence and the acceleration feature vector sequence into the convolutional neural network layer.

[0155] The deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, and the convolutional neural network layer is connected to the recurrent neural network layer. Specifically, the velocity feature vector sequence and the acceleration feature vector sequence are input to the convolutional neural network layer, so that the convolutional neural network uses Formula 3 to extract the behavioral feature values ​​of the spatiotemporal trajectory point sequence and output the feature map corresponding to the spatiotemporal trajectory point sequence. Formula 3 is:

[0156] in, The feature map is at position The behavioral characteristic value of is the preset activation function, is the size of the convolution kernel of the convolutional neural network, and is an odd number, is the expansion offset of the convolution kernel The weight of The position of the input data after expansion The input data includes the velocity feature vector sequence and the acceleration feature vector sequence. is the preset bias item;

[0157] The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

[0158] Furthermore, before extracting the behavioral feature values ​​of the spatiotemporal trajectory point sequence to output the feature map corresponding to the spatiotemporal trajectory point sequence, the method further includes:

[0159] Step a: When training a convolutional neural network, it is the expansion offset of the convolution kernel. Randomly assign an initial weight value ;

[0160] Step b: inputting a preset spatiotemporal trajectory training set into a convolutional neural network, performing forward conduction through the convolutional neural network to output a training predicted position range, wherein the preset spatiotemporal trajectory training set includes a labeled spatiotemporal trajectory information set and a labeled actual spatiotemporal trajectory point corresponding to each labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set, and the labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set includes a labeled velocity feature vector sequence and a labeled acceleration feature vector sequence corresponding to the labeled spatiotemporal trajectory point sequence;

[0161] Step c: determining a position prediction error based on the training prediction position range and the actual space-time trajectory point, where the position prediction error is used to represent the distance between the coordinates of the center point of the training prediction position range and the position coordinates of the actual space-time trajectory point;

[0162] Step d: Calculate the extended offset of the convolution kernel based on the position prediction error Current weight value The feature gradient of , and use formula 4 to update the next iteration weight value of the convolution kernel , Formula 4 is:

[0163] in, is the preset learning rate, is the current iteration number, The position prediction error determined by the gradient descent method with respect to the current weight value characteristic gradient of

[0164] Repeat steps b to d until the determined position prediction error is less than the preset position prediction error threshold or the current number of iterations reaches the maximum number of iterations, then determine the latest weight value as the extended offset of the convolution kernel Weight .

[0165] It's worth noting that Formula 4 calculates the gradient of the position prediction error with respect to the current weight and updates the weight value using gradient descent, allowing the convolution kernel to gradually learn the optimal weight configuration. This weight optimization process helps reduce prediction error and improve the model's prediction accuracy.

[0166] By repeatedly executing the weight update steps in Equation 4, the model gradually converges to the optimal solution during training. This process enables the model to gradually learn the inherent patterns and characteristics of the data, improving its generalization ability. By utilizing gradient information to guide weight updates, Equation 4 accelerates the model training process. Compared to random search or other heuristic methods, gradient descent is more efficient in finding the direction of weight optimization.

[0167] In summary, Formula 4 significantly improves the model's prediction accuracy and generalization capabilities by iteratively updating the convolution kernel weights in large-scale AI spatiotemporal data processing methods. Formula 4 is an indispensable part of the model training process and provides strong support for achieving high-precision and efficient spatiotemporal trajectory prediction.

[0168] Furthermore, before extracting the behavioral feature values ​​of the spatiotemporal trajectory point sequence to output the feature graph corresponding to the spatiotemporal trajectory point sequence, the following steps may be further included:

[0169] Using formula 5, the value of the input data at the expanded position is determined based on the velocity feature vector sequence and the acceleration feature vector sequence. , Formula 5 is:

[0170] in, For input data at location The original value of is the extension offset in the first direction, is the extension offset in the second direction, is the first weighting coefficient, is the second weighting coefficient, For location The neighborhood set within the preset range of For location Relative position The distribution weight of For input data at location The original value of .

[0171] It is worth noting that Formula 5 generates a richer and more detailed feature representation by considering the information within the neighborhood of the input data. This extension not only includes the information of the original data point, but also covers the data of its surrounding neighborhood, thereby improving the representation ability of the data.

[0172] Specifically, in spatiotemporal trajectory data, sudden changes in velocity or acceleration often contain important motion information. By expanding the input data, Equation 5 can more accurately capture the changes in these sudden changes and their surroundings, providing more precise basic data for subsequent feature extraction and prediction.

[0173] Furthermore, expanding the input data helps reduce the error propagation caused by inaccurate data representation. By calculating the weighted average or weighted sum within the neighborhood, the impact of a single data point on the overall prediction result can be reduced, improving the robustness of the prediction.

[0174] Furthermore, by incorporating neighborhood information, Equation 5 enables the model to maintain high prediction accuracy in the presence of noise or missing data. Even if some input data is inaccurate or missing, the model can compensate and correct it using information from other neighborhood data, thereby improving the model's generalization ability.

[0175] Because different target objects (such as vehicles and pedestrians) and different motion patterns (such as linear and curved motion) exhibit distinct characteristics along their spatiotemporal trajectories, Formula 5, by incorporating neighborhood information, enables the model to better adapt to these variations and improve prediction accuracy. Subsequent feature extraction layers (such as convolutional neural network layers) can extract more comprehensive and detailed feature information when processing input data expanded by Formula 5. This feature information helps the model better understand the target object's motion patterns and improve prediction accuracy.

[0176] S205: Input the feature map corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence into the recurrent neural network layer to determine the predicted position point.

[0177] In this step, the feature map corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence can be input into the recurrent neural network layer to determine the predicted position point, wherein the recurrent neural network layer is used to capture the dependency between the position information in the feature map and the time information in the spatiotemporal trajectory point sequence.

[0178] Specifically, the processed feature maps are obtained from the convolutional neural network layer. These feature maps contain the behavioral feature values ​​of each position point in the spatiotemporal trajectory sequence, which are key data for subsequent analysis. At the same time, the timestamp sequences corresponding to the spatiotemporal trajectory sequence are compiled. These timestamp sequences record the temporal information of each spatiotemporal trajectory point and are important for understanding the temporal patterns of the target object's motion.

[0179] The feature map and timestamp sequence are then fed simultaneously into the recurrent neural network layer. In this layer, the network uses a recurrent structure to gradually process the input sequence data, capturing the dependency between the positional information in the feature map and the temporal information in the timestamp sequence.

[0180] Next, the recurrent neural network layer, through its internal state transfer mechanism, captures how the position information in the feature map changes over time. This mechanism enables the network to understand how the positions of the target object at different time points are related to each other. By continuously using the output of the previous moment as the input (or partial input) of the current moment, the recurrent neural network layer can integrate historical information and make more accurate predictions about future positions. After processing by the recurrent neural network layer, the network ultimately outputs a predicted location point. This predicted location point is derived through internal learning and reasoning based on the input feature map and timestamp sequence, and represents the most likely location of the target object at the next time point.

[0181] S206: Determine the predicted position range of the next time node according to the predicted position point and the preset position extension range.

[0182] In this step, the feature map corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to determine the predicted position point, wherein the recurrent neural network layer is used to capture the dependency between the position information in the feature map and the time information in the spatiotemporal trajectory point sequence.

[0183] Specifically, Formula 6 may be used to determine the predicted position range of the next time node based on the predicted position point and the preset position extension range. Formula 6 is:

[0184] in, To predict the location range, To predict the location point, The preset position extension range is greater than the preset position prediction error threshold.

[0185] It's worth noting that in actual predictions, due to factors such as data noise and model errors, the prediction results often have a certain degree of uncertainty. Formula 6 accounts for this uncertainty by introducing a preset position expansion range. This results in the prediction result being a range rather than a single point, thereby improving the practicality and reliability of the prediction results.

[0186] Furthermore, a predicted location range can better meet the needs of practical applications than a single predicted location point. For example, in traffic management, knowing the area where a vehicle is likely to appear is more important than knowing its exact location, because regional information can help traffic managers allocate resources and plan traffic flows more efficiently.

[0187] Furthermore, by adjusting the size of the preset location extension range, it can flexibly adapt to the prediction needs of different scenarios. For scenarios requiring higher accuracy, the size can be appropriately reduced; for scenarios requiring a wider range, the size can be appropriately increased.

[0188] By taking into account prediction uncertainty, Formula 6 makes the entire prediction model more robust in the face of data noise, outliers, etc. Even if there is an error in the predicted location point, as long as the error is within the preset location expansion range, the prediction result is still valid.

[0189] Furthermore, the predicted location ranges serve as the model’s output, facilitating subsequent processing and analysis. For example, the predicted location ranges can be used to set monitoring areas and adjust traffic light timings, enabling more efficient traffic management and control.

[0190] Based on the above embodiment, after determining the predicted position range of the next time node according to the predicted position point and the preset position extension range, the method further includes:

[0191] At the next time node, the actual space-time trajectory point of the target object is obtained;

[0192] The actual position prediction error is determined based on the predicted position range and the actual space-time trajectory points. If the actual position prediction error is greater than the prediction error threshold, newly annotated space-time trajectory information is generated and added to the preset space-time trajectory training set to generate an updated space-time trajectory training set. The updated space-time trajectory training set is used for iterative training of the deep space-time neural network model. The newly annotated space-time trajectory information includes space-time trajectory information and the predicted position range.

[0193] It's worth noting that at each time point, the acquisition of actual spatiotemporal trajectory points provides immediate feedback to the model. By comparing predicted results with actual data, the system can promptly identify prediction errors and adjust model parameters accordingly. This real-time feedback mechanism enables the model to quickly adapt to new data and dynamically changing environments.

[0194] When the actual position prediction error exceeds a preset threshold, the system generates new annotated spatiotemporal trajectory information and adds it to the preset spatiotemporal trajectory training set. This step effectively expands the diversity of training data, allowing the model to learn and optimize over a wider range of data.

[0195] The updated spatiotemporal trajectory training set is used to iteratively train the deep spatiotemporal neural network model. By continuously introducing new training samples, the model is able to gradually reduce prediction error and improve prediction accuracy. This iterative training process ensures continuous improvement in model performance.

[0196] By continuously receiving new data and feedback in real-world scenarios, the model can learn more details and variations about the target object's motion patterns. This learning method enhances the model's generalization ability, enabling it to better handle unknown or rare situations.

[0197] Continuous model optimization and iterative training reduce model instability and improve the reliability of prediction results. Even in complex and changing environments, the model can maintain high prediction accuracy and stability.

[0198] This process provides an effective framework for the long-term operation and maintenance of the model. By regularly acquiring real data, calculating prediction errors, and updating the training set, the system ensures that the model is always in optimal condition, providing reliable prediction support for decision-making.

[0199] By adjusting the model based on actual errors, the system can better meet user needs. As user feedback continues to accumulate, the model will become increasingly closer to actual user usage, improving user experience and satisfaction.

[0200] Figure 3 This is a schematic diagram of the structure of an artificial intelligence spatiotemporal data processing device based on a large model according to an exemplary embodiment of the present application. Figure 3 As shown, the artificial intelligence spatiotemporal data processing device 300 based on a large model provided in this embodiment includes:

[0201] An acquisition module 310 is configured to acquire spatiotemporal trajectory information of a target object, wherein the spatiotemporal trajectory information includes a spatiotemporal trajectory point sequence, wherein the spatiotemporal trajectory point sequence includes multiple spatiotemporal trajectory points of the target object, and each spatiotemporal trajectory point includes a position coordinate and a corresponding timestamp;

[0202] a processing module 320 for determining a velocity feature vector of the target object based on a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, wherein the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is the space-time trajectory point next to the first space-time trajectory point;

[0203] The processing module 320 is further configured to determine an acceleration feature vector of the target object based on the first, second, and third space-time trajectory points in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence, wherein the third space-time trajectory point is a space-time trajectory point subsequent to the second space-time trajectory point;

[0204] The processing module 320 is also used to determine the predicted position range of the target object at the next time node of the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence and the preset space-time trajectory prediction model. The space-time trajectory point corresponding to the current time node is the last space-time trajectory point in the space-time trajectory point sequence. The preset space-time trajectory prediction model is a prediction model trained based on a deep space-time neural network model. The convolutional neural network included in the deep space-time neural network model is used to extract the position change characteristics between each space-time trajectory point in the space-time trajectory point sequence.

[0205] Optionally, the processing module 320 is specifically configured to:

[0206] The determining the velocity feature vector of the target object according to the first space-time trajectory point and the second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence includes:

[0207] determining the velocity feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, and the timestamp of the second space-time trajectory point in the space-time trajectory point sequence;

[0208] The speed feature vector sequence corresponding to the space-time trajectory point sequence is generated according to the speed feature vector corresponding to each time node.

[0209] Optionally, the processing module 320 is specifically configured to:

[0210] determining the acceleration feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, the timestamp of the second space-time trajectory point, the position coordinates of the third space-time trajectory point, and the timestamp of the third space-time trajectory point in the space-time trajectory point sequence;

[0211] The acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence is generated according to the acceleration feature vector corresponding to each time node.

[0212] Optionally, the deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, and the convolutional neural network layer is connected to the recurrent neural network layer; optionally, the processing module 320 is specifically used to:

[0213] Inputting the velocity feature vector sequence and the acceleration feature vector sequence into the convolutional neural network layer, so that the convolutional neural network extracts the behavioral feature values ​​of the spatiotemporal trajectory point sequence, and outputs a feature map corresponding to the spatiotemporal trajectory point sequence;

[0214] The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

[0215] Optionally, the processing module 320 is specifically configured to:

[0216] Step a: When training the convolutional neural network, randomly assigning an initial weight value to the expansion offset of the convolution kernel in the convolutional neural network;

[0217] Step b: inputting a preset spatiotemporal trajectory training set into the convolutional neural network and performing forward conduction through the convolutional neural network to output a training predicted position range, wherein the preset spatiotemporal trajectory training set includes a labeled spatiotemporal trajectory information set and a labeled actual spatiotemporal trajectory point corresponding to each labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set, and the labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set includes a labeled velocity feature vector sequence and a labeled acceleration feature vector sequence corresponding to the labeled spatiotemporal trajectory point sequence;

[0218] Step c: determining a position prediction error based on the training predicted position range and the actual space-time trajectory point, wherein the position prediction error is used to represent the distance between the coordinates of the center point of the training predicted position range and the position coordinates of the actual space-time trajectory point;

[0219] Step d, calculating the characteristic gradient of the current weight value of the extended offset of the convolution kernel according to the position prediction error, and updating the next iterative weight value of the convolution kernel;

[0220] Repeat steps b to d until the determined position prediction error is less than a preset position prediction error threshold or the current number of iterations reaches the maximum number of iterations, and then determine the latest weight value as the weight of the extended offset of the convolution kernel.

[0221] Optionally, the processing module 320 is specifically configured to:

[0222] The value of the input data at the expanded position is determined based on the velocity feature vector sequence and the acceleration feature vector sequence, wherein the value of the input data at the expanded position is used to determine the feature map, and the input data includes the velocity feature vector sequence and the acceleration feature vector sequence.

[0223] Optionally, the processing module 320 is specifically configured to:

[0224] Inputting the feature graph corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence into the recurrent neural network layer to determine the predicted position point, wherein the recurrent neural network layer is used to capture the dependency between the position information in the feature graph and the time information in the spatiotemporal trajectory point sequence;

[0225] The predicted position range of the next time node is determined according to the predicted position point and the preset position extension range.

[0226] Figure 4 FIG. 1 is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of the present application. Figure 4 As shown, this embodiment provides an electronic device 400 including: a processor 401 and a memory 402; wherein:

[0227] The memory 402 is used to store computer programs. The memory may also be a flash memory.

[0228] The processor 401 is configured to execute the execution instructions stored in the memory to implement each step in the above method. For details, please refer to the relevant description in the above method embodiment.

[0229] Optionally, the memory 402 may be independent or integrated with the processor 401 .

[0230] When the memory 402 is a device independent of the processor 401, the electronic device 400 may further include:

[0231] The bus 403 is used to connect the memory 402 and the processor 401 .

[0232] This embodiment further provides a readable storage medium, in which a computer program is stored. When at least one processor of an electronic device executes the computer program, the electronic device executes the methods provided in the various aforementioned embodiments.

[0233] This embodiment further provides a program product, which includes a computer program stored in a readable storage medium. At least one processor of an electronic device can read the computer program from the readable storage medium, and at least one processor can execute the computer program to cause the electronic device to implement the methods provided in the various embodiments described above.

[0234] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered merely as exemplary, and the true scope and spirit of the present application are indicated by the claims.

[0235] It should be understood that the present application is not limited to the exact structure described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. An artificial intelligence spatiotemporal data processing method based on a large model, characterized in that: include: Acquire spatiotemporal trajectory information of a target object, the spatiotemporal trajectory information including a spatiotemporal trajectory point sequence, the spatiotemporal trajectory point sequence including a plurality of spatiotemporal trajectory points of the target object, each spatiotemporal trajectory point including a position coordinate and a corresponding timestamp; determining a velocity feature vector of the target object based on a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, wherein the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is a space-time trajectory point next to the first space-time trajectory point; determining an acceleration feature vector of the target object based on the first, second, and third space-time trajectory points in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence, wherein the third space-time trajectory point is a space-time trajectory point subsequent to the second space-time trajectory point; Determining a predicted position range of the target object at a time node next to the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence, and a preset spatiotemporal trajectory prediction model, wherein the spatiotemporal trajectory point corresponding to the current time node is the last spatiotemporal trajectory point in the spatiotemporal trajectory point sequence, and the preset spatiotemporal trajectory prediction model is a prediction model trained based on a deep spatiotemporal neural network model, wherein the convolutional neural network included in the deep spatiotemporal neural network model is used to extract position change features between each spatiotemporal trajectory point in the spatiotemporal trajectory point sequence; The deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, and the convolutional neural network layer is connected to the recurrent neural network layer. Correspondingly, determining the predicted position range of the target object at the next time node of the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence, and the preset spatiotemporal trajectory prediction model includes: The velocity feature vector sequence and the acceleration feature vector sequence are input into the convolutional neural network layer, so that the convolutional neural network extracts the behavioral feature values ​​of the spatiotemporal trajectory point sequence using Formula 3 to output a feature map corresponding to the spatiotemporal trajectory point sequence. Formula 3 is: in, The feature map is at position The behavioral characteristic value of is the preset activation function, is the size of the convolution kernel of the convolutional neural network, and is an odd number, is the expansion offset of the convolution kernel The weight of The position of the input data after expansion The input data includes the velocity feature vector sequence and the acceleration feature vector sequence. The bias term is preset, the convolution kernel is shared at all positions of the input data to reduce the number of parameters of the deep spatiotemporal neural network model, and the same feature detector is applied to the entire input data to identify similar features at different positions. The convolution operation of the convolution kernel is used to capture velocity mutations in the velocity feature vector sequence and acceleration mutations in the acceleration feature vector sequence; The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

2. The artificial intelligence spatiotemporal data processing method based on a large model according to claim 1 is characterized in that: The determining the velocity feature vector of the target object according to the first space-time trajectory point and the second space-time trajectory point in the space-time trajectory point sequence to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence includes: determining the velocity feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, and the timestamp of the second space-time trajectory point in the space-time trajectory point sequence; The speed feature vector sequence corresponding to the space-time trajectory point sequence is generated according to the speed feature vector corresponding to each time node.

3. The artificial intelligence spatiotemporal data processing method based on a large model according to claim 2 is characterized in that: The determining of the acceleration feature vector of the target object according to the first space-time trajectory point, the second space-time trajectory point, and the third space-time trajectory point in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence includes: determining the acceleration feature vector according to the position coordinates of the first space-time trajectory point, the timestamp of the first space-time trajectory point, the position coordinates of the second space-time trajectory point, the timestamp of the second space-time trajectory point, the position coordinates of the third space-time trajectory point, and the timestamp of the third space-time trajectory point in the space-time trajectory point sequence; The acceleration feature vector sequence corresponding to the spatiotemporal trajectory point sequence is generated according to the acceleration feature vector corresponding to each time node.

4. The artificial intelligence spatiotemporal data processing method based on a large model according to claim 1 is characterized in that: Before extracting the behavior feature value of the spatiotemporal trajectory point sequence to output the feature map corresponding to the spatiotemporal trajectory point sequence, the method further includes: Step a: When training the convolutional neural network, randomly assigning an initial weight value to the expansion offset of the convolution kernel in the convolutional neural network; Step b: inputting a preset spatiotemporal trajectory training set into the convolutional neural network and performing forward conduction through the convolutional neural network to output a training predicted position range, wherein the preset spatiotemporal trajectory training set includes a labeled spatiotemporal trajectory information set and a labeled actual spatiotemporal trajectory point corresponding to each labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set, and the labeled spatiotemporal trajectory information in the labeled spatiotemporal trajectory information set includes a labeled velocity feature vector sequence and a labeled acceleration feature vector sequence corresponding to the labeled spatiotemporal trajectory point sequence; Step c: determining a position prediction error based on the training predicted position range and the actual space-time trajectory point, wherein the position prediction error is used to represent the distance between the coordinates of the center point of the training predicted position range and the position coordinates of the actual space-time trajectory point; Step d, calculating the characteristic gradient of the current weight value of the extended offset of the convolution kernel according to the position prediction error, and updating the next iterative weight value of the convolution kernel; Repeat steps b to d until the determined position prediction error is less than a preset position prediction error threshold or the current number of iterations reaches the maximum number of iterations, and then determine the latest weight value as the weight of the extended offset of the convolution kernel.

5. The artificial intelligence spatiotemporal data processing method based on a large model according to claim 4 is characterized in that: Before extracting the behavior feature value of the spatiotemporal trajectory point sequence to output the feature map corresponding to the spatiotemporal trajectory point sequence, the method further includes: The value of the input data at the expanded position is determined based on the velocity feature vector sequence and the acceleration feature vector sequence, wherein the value of the input data at the expanded position is used to determine the feature map, and the input data includes the velocity feature vector sequence and the acceleration feature vector sequence.

6. The artificial intelligence spatiotemporal data processing method based on a large model according to claim 5 is characterized in that: Inputting the feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence into the recurrent neural network layer to output the predicted position range includes: Inputting the feature graph corresponding to the spatiotemporal trajectory point sequence and the timestamp sequence corresponding to the spatiotemporal trajectory point sequence into the recurrent neural network layer to determine the predicted position point, wherein the recurrent neural network layer is used to capture the dependency between the position information in the feature graph and the time information in the spatiotemporal trajectory point sequence; The predicted position range of the next time node is determined according to the predicted position point and the preset position extension range.

7. An artificial intelligence spatiotemporal data processing device based on a large model, characterized in that: include: an acquisition module, configured to acquire spatiotemporal trajectory information of a target object, wherein the spatiotemporal trajectory information includes a spatiotemporal trajectory point sequence, wherein the spatiotemporal trajectory point sequence includes a plurality of spatiotemporal trajectory points of the target object, and each spatiotemporal trajectory point includes a position coordinate and a corresponding timestamp; a processing module, configured to determine a velocity feature vector of the target object based on a first space-time trajectory point and a second space-time trajectory point in the space-time trajectory point sequence, so as to form a velocity feature vector sequence corresponding to the space-time trajectory point sequence, wherein the first space-time trajectory point and the second space-time trajectory point are any two adjacent space-time trajectory points in the space-time trajectory point sequence, and the second space-time trajectory point is a space-time trajectory point next to the first space-time trajectory point; The processing module is further configured to determine an acceleration feature vector of the target object based on the first space-time trajectory point, the second space-time trajectory point, and a third space-time trajectory point in the space-time trajectory point sequence to form an acceleration feature vector sequence corresponding to the space-time trajectory point sequence, wherein the third space-time trajectory point is a space-time trajectory point subsequent to the second space-time trajectory point; The processing module is further used to determine the predicted position range of the target object at the next time node of the current time node based on the velocity feature vector sequence, the acceleration feature vector sequence, and a preset spatiotemporal trajectory prediction model, the spatiotemporal trajectory point corresponding to the current time node is the last spatiotemporal trajectory point in the spatiotemporal trajectory point sequence, the preset spatiotemporal trajectory prediction model is a prediction model trained based on a deep spatiotemporal neural network model, and the convolutional neural network included in the deep spatiotemporal neural network model is used to extract position change features between each spatiotemporal trajectory point in the spatiotemporal trajectory point sequence; The deep spatiotemporal neural network model includes at least one convolutional neural network layer and at least one recurrent neural network layer, wherein the convolutional neural network layer is connected to the recurrent neural network layer; correspondingly, the processing module is specifically configured to: The velocity feature vector sequence and the acceleration feature vector sequence are input into the convolutional neural network layer, so that the convolutional neural network extracts the behavioral feature values ​​of the spatiotemporal trajectory point sequence using Formula 3 to output a feature map corresponding to the spatiotemporal trajectory point sequence. Formula 3 is: in, The feature map is at position The behavioral characteristic value of is the preset activation function, is the size of the convolution kernel of the convolutional neural network, and is an odd number, is the expansion offset of the convolution kernel The weight of The position of the input data after expansion The input data includes the velocity feature vector sequence and the acceleration feature vector sequence. The bias term is preset, the convolution kernel is shared at all positions of the input data to reduce the number of parameters of the deep spatiotemporal neural network model, and the same feature detector is applied to the entire input data to identify similar features at different positions. The convolution operation of the convolution kernel is used to capture velocity mutations in the velocity feature vector sequence and acceleration mutations in the acceleration feature vector sequence; The feature map corresponding to the spatiotemporal trajectory point sequence and the spatiotemporal trajectory point sequence are input into the recurrent neural network layer to output the predicted position range.

8. An electronic device, characterized in that: include: processor; as well as, a memory for storing executable instructions of the processor; The processor is configured to perform the method according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 6 when executed by a processor.

Images