Multi-target positioning, motion trend prediction and trajectory tracking method and system

Abstract

The invention belongs to the field of artificial intelligence, computer vision and computer graphics, and discloses a multi-target positioning, motion trend prediction and trajectory tracking method and system, and the method comprises the steps: obtaining a feature vector of a target; constructing a space-time diagram attention model; and inputting the feature vector into the time-space diagram attention model to realize target positioning, prediction of a motion trend and trajectory tracking. According to the method, a neural network method based on a graph structure is provided, the electromagnetic signal propagation process in reality is simulated through ray tracing, and therefore accurate multi-target position positioning, motion trend prediction and trajectory tracking are achieved. The complex relation between the targets is modeled through the graph neural network, the problem that the positioning precision is reduced due to mutual shielding and signal interference between the multiple targets is solved by learning the interaction between the targets, and the multi-target positioning precision is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence, computer vision, and computer graphics, and particularly relates to a method and system for multi-object localization, motion trend prediction, and trajectory tracking. Background Art

[0002] In today's rapidly developing technological field, multi-object localization, motion trend prediction, and trajectory tracking technologies are gradually becoming core technologies in numerous application scenarios. These applications not only cover high-tech fields such as intelligent transportation systems, autonomous vehicles, and drone monitoring and management, but also include aspects such as the security guarantee of large public events, the automated management of logistics warehousing, and large-space VR immersive tours. With the acceleration of urbanization and the increasing demand for intelligent life, the need for technologies that can accurately and efficiently process multi-object dynamic information is growing day by day.

[0003] In many actual scenarios, such as busy city streets, crowded shopping malls, or complex industrial environments, there are a large number of moving objects. Taking intelligent transportation as an example, vehicles, pedestrians, and other obstacles on the road constitute a complex and ever-changing environment. To ensure traffic safety and improve traffic efficiency, it is necessary to know the position of each moving object in real time and accurately predict its future motion trend. Similarly, in a logistics warehouse, an automated guided vehicle (AGV) needs to adjust its path planning according to changes in the surrounding environment to avoid collisions with other AGVs or fixed obstacles. In large public events, such as concerts and sports events, real-time monitoring and management of the crowd are one of the important measures to ensure the safety of participants.

[0004] However, traditional multi-object localization, motion trend prediction, and trajectory tracking methods face many challenges. First, due to the complexity and uncertainty of the environment, obtaining high-quality data itself is a major problem. Second, even if sufficient data is obtained, how to extract useful parts from this vast amount of information and make accurate predictions based on it is also a huge challenge. Moreover, existing algorithms often struggle to simultaneously achieve high accuracy and real-time performance, two key indicators, resulting in unsatisfactory results in practical applications. Summary of the Invention

[0005] To solve the problems existing in the prior art, the present invention provides a method and system for multi-target positioning, motion trend prediction, and trajectory tracking. This method proposes a neural network method based on a graph structure, which simulates the propagation process of electromagnetic signals in reality through ray tracing, thereby achieving accurate multi-target position positioning, motion trend prediction, and trajectory tracking. By using a graph neural network to model the complex relationships between targets and learning the interactions between targets, the problem of reduced positioning accuracy caused by mutual occlusion and signal interference between multi-targets is solved, and the accuracy of multi-target positioning is improved.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for multi-target positioning, motion trend prediction, and trajectory tracking, the method comprising:

[0008] Obtain the feature vector of the target;

[0009] Construct a spatio-temporal graph attention model;

[0010] Input the feature vector into the spatio-temporal graph attention model to achieve target positioning, prediction of motion trend, and trajectory tracking.

[0011] Preferably, the target includes: static target, dynamic target, or full parameters;

[0012] For the case of full parameters and only dynamic targets as inputs, the feature vector where is the initial position vector of node i, is the initial velocity vector of node i, is the initial acceleration vector of node i, is the initial electromagnetic signal feature vector of node i, is the initial time feature vector of node i, is the initial optical flow vector of node i, where the node is the target;

[0013] For the case of only static targets as inputs, the feature vector

[0014] Preferably, the method of inputting the feature vector into the spatio-temporal graph attention model to achieve target positioning, prediction of motion trend, and trajectory tracking includes:

[0015] Model multiple targets in the virtual scene as nodes in the graph, and each node represents a target, including the inherent attributes and dynamic attributes of the target;

[0016] The edges between nodes represent the relationships between targets, and the weights of the edges are dynamically adjusted based on the distance, relative position, motion trend, and CSI information between the targets;

[0017] Each layer of GAT consists of several attention heads. Each attention head calculates the local correlation between nodes. By stacking multiple layers of GAT, multi-level interaction relationships between targets are captured. In each layer, the features of the nodes are propagated to the corresponding neighbor nodes and weighted aggregation is performed through the attention mechanism, thereby updating the feature representation of the nodes;

[0018] Fuse the node features at different levels, capture deeper feature representations, output the position of each target at a single time point, as well as the position, speed, acceleration, and CSI signals in the time series, and achieve the prediction of the target motion trend and trajectory tracking.

[0019] Preferably, the calculation method of the edge weight includes:

[0020]

[0021] where λ s and λ t are weight coefficients used to balance the contributions of spatial, temporal, and CSI weights, is the spatial weight, is the temporal weight, is the CSI weight.

[0022] Preferably, the method by which the features of the nodes are propagated includes:

[0023]

[0024] where represents the node feature of node i at the k-th layer, represents the updated node feature of node i, where is the position vector of node i at the k-th layer, is the velocity vector of node i at the k-th layer, is the acceleration vector of node i at the k-th layer, is the electromagnetic signal feature vector of node i at the k-th layer, is the time feature vector of node i at the k-th layer, is the optical flow vector of node i at the k-th layer, ReLU(·) represents the ReLU activation function, represents the set of neighbor nodes of node i, α i,j is the attention coefficient, W (k) is the weight matrix of the k-th layer.

[0025] Preferably, the method of using the fused feature vector as the final feature of the node for subsequent position prediction and other tasks includes: using a fully connected layer to map the node features to position coordinates:

[0026]

[0027] where W pos is the weight matrix, b pos is the bias vector, L is the number of layers of the model, represents the node feature of node i at the L-th layer.

[0028] Preferably, the loss function of the spatio-temporal graph attention model is:

[0029]

[0030] where represents the position prediction error loss, represents the time series prediction error loss, represents the CSI loss, represents the regularization loss, α pos 、β vel 、γ csi and δ reg are weight coefficients used to balance the contributions of different loss functions.

[0031] The present invention also provides a system for multi-objective positioning, motion trend prediction and trajectory tracking, which is used to implement any one of the above methods. The system includes: an acquisition module, a construction module and a prediction module;

[0032] The acquisition module is used to acquire the feature vector of the target;

[0033] The construction module is used to construct a spatio-temporal graph attention model;

[0034] The prediction module is used to input the feature vector into the spatio-temporal graph attention model to realize target positioning, motion trend prediction and trajectory tracking.

[0035] The present invention also provides an electronic device, including a memory, a processor and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements any one of the above methods.

[0036] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed, it implements any one of the above methods.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] Multi-target precise positioning. One is multi-target, which can be dynamic targets; if they are static targets, the accuracy will be higher; the other is the accuracy of positioning, and it can also predict the movement trends and trajectory tracking of dynamic targets, etc.

[0039] 1. Spatiotemporal feature fusion: The spatiotemporal graph can simultaneously model the spatial position and time series information of the target. By fusing the spatial features of the target such as position, velocity, and acceleration with the time series features (such as optical flow field, rate of change of velocity, etc.), the model can more comprehensively understand the motion state and trajectory of the target.

[0040] 2. Dynamic adaptability: The spatiotemporal graph can dynamically adapt to the movement changes of the target. When the velocity, direction, or position of the target changes, the model can adjust the positioning and tracking of the target in real time by updating the node features and edge weights.

[0041] 3. Rich channel information: The CSI signal contains rich channel state information, such as signal strength, phase, multipath effect, etc. Compared with the traditional Received Signal Strength Indication (RSSI), CSI can provide more detailed channel characteristics, thus more accurately reflecting the interaction between the target and the channel.

[0042] 4. Signal feature fusion: By taking the amplitude and phase information of the CSI signal as part of the node features, the model can better utilize these signal features for target positioning and tracking.

[0043] 5. Automatic learning of attention weights: The graph attention network can automatically learn the attention weights between nodes, thus better capturing the interactions and dependencies between targets. Compared with traditional graph neural networks, GAT can more flexibly adjust the information propagation between nodes through the attention mechanism, thus better adapting to complex target motion scenarios.

[0044] 6. Scalability and robustness: This method has good scalability and robustness and can handle different numbers and types of targets. By sharing weights and parameters, the model can adapt to graph structures of different scales, thus having a wider applicability in practical applications. In addition, GAT has a certain robustness to noise and outliers and can maintain stable performance in complex environments.

[0045] 7. High-precision positioning: By fusing spatiotemporal features and CSI signals, the model can more accurately estimate the position and motion state of the target, thus achieving high-precision multi-target positioning and tracking. Compared with traditional methods, this method can better utilize the interactions between targets and channel characteristics, thus improving the accuracy and robustness of positioning.

[0046] 8. Real-time and dynamic adaptability: The spatio-temporal graph and GAT can update node features and edge weights in real time, thus adapting to the dynamic changes of the target. This enables the model to maintain accurate tracking of the target in real-time applications and maintain good performance even in the case of complex target motion or environmental changes.

[0047] 9. Adaptability to complex scenarios: By fusing CSI signals and spatio-temporal features, the model can better adapt to complex scenarios and environments. For example, in the presence of occlusion, interference, or multipath effects, the model can accurately estimate the position and motion trajectory of the target by analyzing the interaction between the CSI signal and the target, thereby improving the reliability of positioning and tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] To more clearly illustrate the technical solutions of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 It is a schematic flowchart of a method for multi-target positioning, motion trend prediction, and trajectory tracking according to an embodiment of the present invention;

[0050] Figure 2 It is a schematic diagram of the overall architecture of multi-target positioning, motion trend prediction, and trajectory tracking according to an embodiment of the present invention;

[0051] Figure 3 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0053] 1010. Processor; 1020 Memory; 1030 Input / Output Interface; 1040 Communication Interface; 1050 Bus. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0055] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0056] Embodiment 1

[0057] AsFigure 1 , Figure 2 As shown in Figure 2 , an embodiment of the present invention provides a method for multi-target positioning, motion trend prediction, and trajectory tracking. It is a neural network method based on a graph structure. By ray tracing, the propagation process of electromagnetic signals in reality is simulated, so as to achieve accurate multi-target position positioning, motion trend prediction, and trajectory tracking. Using a graph neural network to model the complex relationships between targets, and by learning the interactions between targets, the problem that the positioning accuracy decreases due to mutual occlusion and signal interference between multi-targets is solved, and the accuracy of multi-target positioning is improved. The method includes:

[0058] Obtain the feature vectors of the targets;

[0059] Construct a spatio-temporal graph attention model;

[0060] Input the feature vectors into the spatio-temporal graph attention model to achieve target positioning, motion trend prediction, and trajectory tracking.

[0061] In this embodiment, the overall architecture consists of "input", "pre-trained target recognition model", "spatio-temporal graph attention model", and "output". In order to simulate the propagation path of electromagnetic signals by ray tracing in a virtual scene, this virtual scene needs to be a high-fidelity scene, meeting the shape structure and physical simulation characteristics, and the signal transmission process, including reflection, refraction, diffraction, etc., can be simulated by ray tracing.

[0062] Among them, the model input part:

[0063] (1) This network model supports inputting only "static targets", "dynamic targets", or full-parameter input, and can adapt to various positioning and tracking requirements.

[0064] (2) This model inputs the feature vectors of each target (node), including inherent attributes (such as size, shape, material, etc.) and dynamic attributes (such as position, velocity, acceleration, etc.). For the cases of full-parameter and only dynamic targets as input, where, is the initial position vector of node i, is the initial velocity vector of node i, is the initial acceleration vector of node i, is the initial electromagnetic signal feature vector of node i, is the initial time feature vector of node i, is the initial optical flow vector of node i. For the input of only "static targets",

[0065] (3) For "static targets", they can be directly input into the network according to the positions and electromagnetic signal feature attributes of the static targets in the virtual scene.

[0066] (4) Use the "pre-trained object recognition model" for "dynamic objects", and input the sequence diagram to identify parameters such as the initial position, speed, acceleration, and optical flow vector of the dynamic object.

[0067] Among them, the sequence diagram is the video containing the dynamic object. The "pre-trained object recognition model" can be an existing mature model, such as the object recognition models in the YOLO series, Faster R-CNN, etc., or the FlowNet for optical flow estimation, RAFT (Recurrent All-Pairs Field Transforms), etc. Here, it does not specifically refer to a certain model or the combination of two certain models, as long as it can identify the parameters of the dynamic object.

[0068] Among them, the model output: This model outputs the accurate position of each object at a single time point, as well as the position, speed, acceleration, and CSI signal (i.e., channel state information) in the time series, so as to realize the prediction of the motion trend and trajectory tracking. To meet the requirements of real-time and high-precision positioning, we output at each graph module. That is, for the case of N pairs of graph modules in the above figure, there are a total of 2N output results. When outputting, through a fully connected layer, the final output parameters, predicted position, and predicted CSI data are generated.

[0069] In this embodiment, the spatio-temporal graph attention model architecture includes:

[0070] 1. Graph structure:

[0071] (1) Model multiple objects in the virtual scene as nodes in the figure. Each node represents an object, including the inherent attributes (such as size, shape, material, etc.) and dynamic attributes (such as position, speed, acceleration, etc.) of the object.

[0072] (2) The edges between nodes represent the relationships between objects, and the weights of the edges can be dynamically adjusted based on factors such as the distance, relative position, motion trend, and CSI information between objects.

[0073] 2. Model structure:

[0074] (1) Multi-layer stacking: Each layer of GAT consists of several attention heads, and each attention head is responsible for calculating the local correlation between nodes. Through the stacking of multiple layers of GAT, the multi-level interaction relationships between objects can be captured. In each layer, the features of the nodes are propagated to their neighbor nodes and weighted aggregation is performed through the attention mechanism to update the feature representation of the nodes.

[0075] (2) Skip connection: Fuse the node features of different levels, which helps to alleviate the problem of gradient disappearance and can capture deeper feature representations at the same time.

[0076] Among them, edge weight calculation: In the spatio-temporal graph attention model, the calculation of edge weights needs to consider the spatial and temporal relationships between targets.

[0077] (1) Comprehensive weight: The spatial weight and the temporal weight are weighted and summed to obtain the final edge weight.

[0078]

[0079] Among them, λ s and λ t are weight coefficients used to balance the contributions of spatial, temporal, and CSI weights. is the spatial weight, is the temporal weight, is the CSI weight.

[0080] (2) Spatial weight Calculate the weight according to the Euclidean distance between targets. The closer the distance, the greater the weight. At the same time, consider the reduction effect of occlusion and interference relationships between targets on the weight.

[0081]

[0082] Among them, represents the spatial weight between target i and target j, p i and p j are the position vectors of target i and target j respectively, ‖p i - p j ‖ represents the Euclidean distance between node i and node j, σ s is the standard deviation of the spatial distance. Occlusion i,j is the occlusion degree of target j to target i, and its value range is from 0 to 1. 0 means complete non-occlusion, and 1 means complete occlusion. The occlusion and interference relationships can be determined by ray tracing technology. α s and β s are weight coefficients used to balance the effects of occlusion degree and distance.

[0083] (3) Temporal weight Calculate the weight according to the speed and acceleration of the targets. The more similar the speed and acceleration, the greater the weight. At the same time, consider the relative time delay effect between targets. The smaller the time delay, the greater the weight.

[0084]

[0085] Among them, is the temporal weight between target i and target j. v i and v j are the speed vectors of target i and target j respectively, ai and a j are the acceleration vectors of target i and target j respectively. Δt i,j is the time delay between target i and target j, and σ t is the standard deviation of the time delay. α t and β t are weight coefficients used to balance the effects of velocity, acceleration, and time delay.

[0086] There is a direct relationship between the phase difference of CSI and the time delay of signal propagation, which can be calculated through the wavelength and the radian of phase change, so as to convert the phase difference of CSI into the time delay Δt i,j . Since the phase difference Δθ (in radians) can be obtained through the wavelength λ L and the distance difference Δd i,j The relationship is expressed as Therefore According to the time delay Δt i,j and the distance difference Δd i,j The relationship where c L is the speed of light. According to the relationship between the wavelength, frequency, and speed of electromagnetic waves It can be obtained that where Δt i,j is the time delay (in seconds), Δθ is the phase difference (in radians), and f is the operating frequency of the electromagnetic wave (in hertz).

[0087] (4) CSI amplitude weight: Calculate the weight of the edge using the amplitude information of CSI. The magnitude of the signal strength can reflect the distance and occlusion between targets, thus affecting the weight of the edge.

[0088]

[0089] Among them, and are the CSI amplitude vectors of target i and target j. d max is the maximum distance threshold.

[0090] Among them, node update:

[0091] (1) Attention Coefficient: After the graph is constructed, the initial features of each target node include its inherent attributes (such as size, shape, material, etc.) and dynamic attributes (such as rough position, speed, acceleration, etc.). These features can be represented in vector form. After entering the GNN model, multiple layers of GAT start to propagate and update the node features. Taking the first layer of GAT as an example, when each attention head calculates the attention coefficient between node i and node j, it comprehensively considers the edge weights between them (including occlusion and interference relationships).

[0092] α i,j =LeakyReLU(a T ·[W·h i ‖W·h j )

[0093] Among them, h i and h j are the feature vectors of node i and node j, W is a learnable weight matrix, and a is an attention vector. Through this mechanism, GNN can automatically learn which neighbor nodes' features are more important to the current node. After multiple layers of propagation, the node features will fuse the information from neighbors at multiple levels and different ranges, thus more accurately representing the state of the target.

[0094] (2) Position Encoding: Position encoding is a method of embedding the spatial position information of the target into the node features to enhance the model's perception ability of the target's spatial position.

[0095]

[0096] Among them, represents the position encoding vector of the i-th node in the input stage. x i , y i , z i are the coordinates of node i (that is, target i, because each target is modeled as a node in the graph) in three-dimensional space. is the scaling factor, used to adjust the frequency of encoding.

[0097] (3) Time Encoding: Time encoding is a method of embedding time information into the node features. It adds a time dimension to the feature vector of each node, enabling the model to distinguish the states of the target at different time points. Assign a timestamp t to each node, representing the state of the target at a specific time point, and use trigonometric functions to convert the timestamp into a multi-dimensional time feature vector.

[0098]

[0099] Among them, Denote the time feature vector of the $i$-th node in the input stage. $\xi$ is the dimension of the time feature vector, $t$ is the timestamp, is a scaling factor, such as used to convert the timestamp to an appropriate range.

[0100] (4) Optical flow vector $\mathbf{F}$ i : Optical flow method is a method for calculating the motion field of an object, which can capture the motion trend and trajectory information of the object.

[0101]

[0102] Among them, $\mathbf{F}$ i is the optical flow vector of object $i$, representing the motion trend of the object in the image sequence. $I$ is the image luminance, is the time derivative of the image luminance, is the spatial gradient of the image luminance, $\mathbf{F}$ i $= [f$ i,x , $f$ i,y $ T is the optical flow field of object $i$, $f$ i,x and $f$ i,y are the components of the optical flow vector at position $(x, y)$. The optical flow field $\mathbf{F}$ i can be solved by the least squares method.

[0103] (5) Feature propagation: Using the attention mechanism of GAT, according to the edge weight $w$ i,j , attention coefficient $\alpha$ i,j and node feature $\mathbf{h}$ i to perform feature propagation.

[0104]

[0105] Among them, represents the node feature of node $i$ at the $k$-th layer, represents the updated node feature of node $i$, where is the position vector of node $i$ at the $k$-th layer, is the velocity vector of node $i$ at the $k$-th layer, is the acceleration vector of node $i$ at the $k$-th layer, is the electromagnetic signal feature vector of node $i$ at the $k$-th layer, is the time feature vector of node $i$ at the $k$-th layer, is the optical flow vector of node $i$ at the $k$-th layer. ReLU(·) represents the ReLU activation function, represents the set of neighbor nodes of node $i$, $\alpha$ i,j is the attention coefficient, $\mathbf{W}$ (k) is the weight matrix of the $k$-th layer.

[0106] Among them, 1. The weight matrix W (k) is the learnable parameter of the k-th layer in the graph neural network, which is used for linear transformation of node features. It is a global parameter and is applied to the features of all nodes. During the feature propagation process, the weight matrix W (k) performs a linear transformation on the feature vector of the node to extract a higher-level feature representation. 2. The edge weight w i,j is the weight of the edge between node i and node j in the graph, indicating the degree of influence of node j on node i. It is a local parameter that reflects the relationship strength between two nodes. The edge weight w i,j is used to adjust the contribution of neighbor node features to the update of the current node features. The larger the edge weight, the greater the influence of the neighbor node features on the update of the current node features. 3. The weight matrix W (k) and the edge weight w i,j are combined through the attention coefficient α i,j The attention coefficient α i,j is based on the edge weight w i,j and is calculated to dynamically adjust the contribution of neighbor node features. The specific formula is α i,j = LeakyReLU(a T ·[W·h i ‖W·h j ) as described above.

[0107] (6) Location prediction: The fused feature vector is used as the final feature of the node for subsequent location prediction and other tasks. A fully connected layer can be used to map the node features to location coordinates:

[0108]

[0109] where W pos is the weight matrix, b pos is the bias vector, L is the number of layers of the model, represents the node feature of node i at the L-th layer.

[0110] In this embodiment, the model loss function:

[0111] (1) Joint loss function:

[0112]

[0113] where represents the location prediction error loss, represents the time series prediction error loss, represents the CSI loss, represents the regularization loss. α pos 、β vel 、γcsi and δ reg are weight coefficients used to balance the contributions of different loss functions.

[0114] (2) Location prediction error loss Used to measure the gap between the predicted location and the true location.

[0115]

[0116] Among them, is the location vector predicted by node i, p i is the true location vector of node i, N O is the number of samples, including all samples, both dynamic and static samples.

[0117] (3) Time series prediction error loss

[0118]

[0119] Among them, and are the predicted velocity and acceleration vectors respectively, v i and a i are the true velocity and acceleration vectors respectively. N D is the number of dynamic samples.

[0120] (4) CSI loss Add a CSI reconstruction loss term to the loss function to ensure that the model can accurately reconstruct CSI features. This can be achieved by minimizing the difference between the predicted CSI and the true measured CSI.

[0121]

[0122] Among them, represents the predicted CSI matrix, is the true measured CSI matrix. N csi is the number of CSI samples. Usually N csi ≤N D That is, not all dynamic objects in the scene are to be located. Some dynamic objects may be dynamic obstacles.

[0123] (5) Regularization loss To prevent the model from overfitting, a regularization term is introduced.

[0124]

[0125] Among them, W is the weight matrix of the model, λ reg is the regularization coefficient.

[0126] Further, after the neural network model predicts the "current predicted position" of each target, it is necessary to emit light rays into this virtual scene. Through ray tracing, the propagation process of electromagnetic signals is simulated. By calculating the value of CSI received at the "current predicted position" of the target, the CSI simulation data at the target position is predicted. Then, by taking the gradient of the measured CSI data in the actual environment and the simulated CSI data, the precise position of the target is predicted.

[0127] Among them, the CSI generated through ray tracing can accurately reflect the propagation characteristics of wireless signals in complex environments, providing a basis for high-precision positioning. CSI contains rich environmental information, such as path delay, angle information, etc., and can be used for environmental perception and scene modeling. During the ray tracing process, the system will record the parameters of each ray propagation path, including the path delay That is, the propagation path parameters of the ray from the transmitter to the receiver; the angle information (AOA or AoD), that is, the angle of arrival and the angle of departure of the ray; the amplitude and phase of the electromagnetic field, that is, the change in the electromagnetic field calculated based on the interaction between the ray and the object. That is

[0128] In theory, CSI can be calculated by integrating all possible propagation paths. However, since there is no closed solution for path integration, the Monte Carlo method is usually used for numerical estimation, that is:

[0129]

[0130] where N w is the number of sampled paths, and w i is the i-th sampled path.

[0131] In an actual wireless system, multiple antennas and orthogonal frequency division multiplexing (OFDM) technology are usually used. Therefore, CSI is a complex-valued matrix, representing the signal propagation characteristics on different antennas and different subcarriers. For the CSI between the n t th transmitting antenna and the n r th receiving antenna, it can be expressed as:

[0132]

[0133] where and are the radiation pattern functions of the receiving and transmitting antennas respectively, is the transmission matrix of path i, describing the change in the electromagnetic field after the ray interacts with the object. represents the phase change caused by the path delay. f is the electromagnetic frequency, and N w is the number of paths.

[0134] The CSI matrix N is an N t × N r × N s complex-valued matrix, where N t and N r are the numbers of transmit and receive antennas respectively, and N s is the number of subcarriers. Each element of the CSI matrix represents the signal propagation characteristics on a specific antenna pair and subcarrier:

[0135]

[0136] where is the frequency of the j-th subcarrier, f c is the center frequency, and Δf is the subcarrier spacing.

[0137] where the loss function provides a direction for the optimization process by measuring the difference between the simulated CSI and the actual CSI. The optimization algorithm (such as gradient descent) will adjust the target position according to the gradient information of the loss function to minimize the value of the loss function. By combining the amplitude and delay information of the CSI, the loss function can more comprehensively reflect the propagation characteristics of the wireless signal, thereby improving the positioning accuracy.

[0138] (1) CSI amplitude error: The amplitude of the CSI reflects the signal strength and is one of the important pieces of information in the positioning process. By minimizing the difference between the simulated CSI amplitude and the actual CSI amplitude, the positioning accuracy can be improved.

[0139]

[0140] where N tot is the total number of elements in the CSI matrix, the actually measured CSI amplitude, is the simulated CSI amplitude.

[0141] (2) CSI delay error: The delay (or phase) of the CSI reflects the signal propagation time and is another important piece of information in the positioning process. By minimizing the difference between the simulated CSI delay and the actual CSI delay, the positioning accuracy can be further improved.

[0142]

[0143] where is the actually measured CSI delay, is the simulated CSI delay.

[0144] (3) Regularization term: To prevent the optimization process from falling into local minima and improve the generalization ability of the model, a regularization term is introduced. The regularization term is usually a simple quadratic term used to constrain the change of the target position.

[0145] R(x,y) = γ2(x 2 +y 2 )

[0146] where x and y are the coordinates of the target position, and γ2 is the regularization coefficient used to balance the weights of the regularization term and other error terms.

[0147] (4) Composite loss function:

[0148]

[0149] where γ1 and γ2 are weight coefficients used to balance the importance of different error terms.

[0150] During the optimization of the target position, the loss landscape of the loss function is usually complex, with problems such as sparse gradients and local minima. These problems can cause the optimization process to fall into local optimal solutions and make it difficult to find the global optimal solution. To solve this problem, this system introduces a coarse-to-fine Gaussian smoothing technique to improve the convergence of the optimization process by smoothing the loss function. Gaussian smoothing is a commonly used smoothing technique that provides smoother gradient information for the optimization process by convolving a Gaussian kernel on the loss function. The role of the Gaussian kernel is to perform a weighted average of the values of the loss function around the target position, thereby reducing local fluctuations and helping the optimization algorithm to escape local minima more easily. Gaussian smoothing helps the optimization algorithm to escape local minima more easily by reducing the local fluctuations of the loss function, thereby improving the convergence speed and accuracy.

[0151] (1) Definition of Gaussian kernel: The Gaussian kernel is a two-dimensional Gaussian function,

[0152]

[0153] where u = (x,y) is the target position, and σ is the standard deviation of the Gaussian kernel, which controls the degree of smoothing.

[0154] (2) Smoothed loss function: The smoothed loss function is obtained by convolving the original loss function with the Gaussian kernel to get

[0155]

[0156] where P iis the initial position of the target, P u is the smoothed target position.

[0157] (3) Monte Carlo sampling: Since directly calculating the convolution is computationally expensive, the Monte Carlo method is used to estimate the smoothed loss function

[0158]

[0159] where N p is the number of sampling points, and u j is the position of the sampling point.

[0160] (4) Introducing bias: To further improve the optimization efficiency, a bias term is introduced to assign higher weights to the sampling points with smaller loss values

[0161]

[0162] where α is the bias coefficient that controls the strength of the bias. The bias term does not participate in the gradient calculation but affects the weights of the sampling points. After introducing the bias,

[0163]

[0164] where represents the bias term for the j-th sampling.

[0165] (5) Adaptive variance adjustment: During the optimization process, the variance of the Gaussian kernel is adaptively adjusted. Initially, it is large to provide a global smoothing effect, which helps the optimization algorithm to perform global exploration; as the optimization progresses, it gradually decreases, enabling the optimization algorithm to optimize the target position more precisely to improve the optimization accuracy. Variance adjustment strategy:

[0166]

[0167] where σ0 is the initial variance, σ m is the minimum variance, T is the total number of iterations, t is the current iteration number, and β is the parameter that controls the variance adjustment speed. where B c is the bias value of the current sampling point, and the value range of β is [-2, 2].

[0168] Embodiment 2

[0169] The present invention also provides a system for multi-target positioning, motion trend prediction, and trajectory tracking. The system is used to implement the method described in any one of the first embodiments. The system includes an acquisition module, a construction module, and a prediction module;

[0170] The acquisition module is used to acquire the feature vector of the target;

[0171] The construction module is used to construct a spatio-temporal graph attention model;

[0172] The prediction module is used to input the feature vector into the spatio-temporal graph attention model to achieve target positioning, prediction of motion trends, and trajectory tracking.

[0173] Embodiment III

[0174] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method for restoring the three-dimensional structure of a high-quality urban renewal landscape building described in any one of the above embodiments.

[0175] Figure 3 FIG. shows a more specific schematic diagram of the hardware structure of the electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. Among them, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other inside the device through the bus 1050.

[0176] The processor 1010 may be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0177] The memory 1020 may be implemented in the form of a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0178] The input / output interface 1030 is used to connect to the input / output module to achieve information input and output. The input / output module can be configured as a component in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Among them, the input device can include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device can include a display, a speaker, a vibrator, an indicator light, etc.

[0179] The communication interface 1040 is used to connect to the communication module (not shown in the figure) to achieve communication interaction between this device and other devices. Among them, the communication module can achieve communication through a wired method (such as USB (Universal Serial Bus), network cable, etc.) or through a wireless method (such as a mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0180] The bus 1050 includes a path for transmitting information between various components of the device (such as the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0181] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device may also only include the components necessary to implement the solution of the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0182] The system of the above embodiment is used to implement the three-dimensional structure restoration method of a corresponding high-quality urban renewal landscape building in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0183] Embodiment 4

[0184] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure also provides a non-transitory computer-readable storage medium, and the non-transitory computer-readable storage medium stores computer instructions for causing the computer to execute a three-dimensional structure restoration method of a high-quality urban renewal landscape building as described in any of the foregoing embodiments.

[0185] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage, or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0186] The computer instructions stored in the storage medium of the above embodiment are used to cause the computer to execute a method for restoring the three-dimensional structure of a high-quality urban renewal landscape building as described in any one of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0187] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present disclosure (including the claims) is limited to these examples; under the concept of the present disclosure, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the embodiments of the present disclosure as described above, and they are not provided in detail for the sake of brevity.

[0188] In addition, for the sake of simplicity of description and discussion, and in order not to make the embodiments of the present disclosure difficult to understand, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. In addition, the device may be shown in block diagram form to avoid making the embodiments of the present disclosure difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present disclosure are to be implemented (i.e., these details should be fully within the understanding of those skilled in the art). In the case where specific details (such as circuits) are set forth to describe the exemplary embodiments of the present disclosure, it will be apparent to those skilled in the art that the embodiments of the present disclosure can be implemented without these specific details or with variations of these specific details. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0189] Although the present disclosure has been described in connection with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art in light of the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0190] Thus, the units of the examples described in the embodiments of the present application can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0191] The embodiments described above are only descriptions of the preferred modes of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solution of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for multi-target positioning, motion trend prediction, and trajectory tracking, characterized in that The method includes: Obtain the feature vector of the target; Construct a spatio-temporal graph attention model; Input the feature vector into the spatio-temporal graph attention model to achieve target localization, prediction of motion trends, and trajectory tracking.

2. The method according to claim 1, wherein The target includes: static target, dynamic target, or full parameters; For the case of full parameters and only dynamic targets as inputs, the feature vector wherein is the initial position vector of node i, is the initial velocity vector of node i, is the initial acceleration vector of node i, is the initial electromagnetic signal feature vector of node i, is the initial time feature vector of node i, is the initial optical flow vector of node i, wherein the node is the target; For the case where only static targets are used as inputs, the feature vectors 3. The method according to claim 2, wherein The method of inputting the feature vector into the spatio-temporal graph attention model to achieve target localization, prediction of motion trends, and trajectory tracking includes: Model multiple targets in the virtual scene as nodes in the graph, where each node represents a target and includes the inherent attributes and dynamic attributes of the target; The edges between nodes represent the relationships between targets, and the weights of the edges are dynamically adjusted based on the distance, relative position, motion trends, and CSI information between targets; Each layer of GAT consists of several attention heads. Each attention head calculates the local correlation between nodes. By stacking multiple layers of GAT, multi-level interaction relationships between targets are captured. In each layer, the features of the nodes are propagated to the corresponding neighbor nodes and weighted aggregation is performed through the attention mechanism, thereby updating the feature representation of the nodes; Fuse the node features at different levels to capture deeper feature representations, output the position of each target at a single time point, as well as the position, speed, acceleration, and CSI signals in the time series, to achieve the prediction of target motion trends and trajectory tracking.

4. The method according to claim 3, wherein The calculation method of the weights of the edges includes: Among them, λ s and λ t are weight coefficients used to balance the contributions of spatial, temporal, and CSI weights, is the spatial weight, is the temporal weight, is the CSI weight.

5. The method according to claim 4, characterized in that, The method by which the features of the nodes are propagated includes: Among them, represents the node feature of node i at the k-th layer, represents the updated node feature of node i, where is the position vector of node i at the k-th layer, is the velocity vector of node i at the k-th layer, is the acceleration vector of node i at the k-th layer, is the electromagnetic signal feature vector of node i at the k-th layer, is the time feature vector of node i at the k-th layer, is the optical flow vector of node i at the k-th layer, and ReLU(·) represents the ReLU activation function, represents the set of neighbor nodes of node i, and α i,j is the attention coefficient, and W (k) is the weight matrix of the k-th layer.

6. The method according to claim 5, wherein The method of using the fused feature vector as the final feature of the node for subsequent position prediction and other tasks includes: using a fully connected layer to map the node features to position coordinates: Among them, W pos is the weight matrix, b pos is the bias vector, L is the number of layers of the model, represents the node feature of node i at the L-th layer.

7. The method according to claim 6, wherein The loss function of the spatio-temporal graph attention model is: Among them, represents the position prediction error loss, represents the time series prediction error loss, represents the CSI loss, represents the regularization loss, α pos , β vel , γ csi and δ reg are weight coefficients used to balance the contributions of different loss functions.

8. A system for multi-target positioning, motion trend prediction, and trajectory tracking, the system being used to implement the method described in any one of claims 1-7, characterized in that, The system includes: an acquisition module, a construction module, and a prediction module; The acquisition module is used to obtain the feature vector of the target; The construction module is used to construct a spatio-temporal graph attention model; The prediction module is used to input the feature vector into the spatio-temporal graph attention model to achieve target localization, prediction of motion trends, and trajectory tracking.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed, it implements the method according to any one of claims 1 to 7.

Images