An adaptive regulation method and system suitable for three-dimensional radar dynamic visualization

Abstract

An adaptive regulation method and system suitable for three-dimensional radar dynamic visualization, the adaptive regulation method comprising real-time acquisition of basic three-dimensional radar data information and running performance data information; using a space-time forgetting mechanism to dynamically predict threats to the basic three-dimensional radar data information, obtaining an ADST-TPM dynamic threat index parameter corresponding to each target; generating an importance weight graph covering all target sets according to the basic three-dimensional radar data information and the ADST-TPM dynamic threat index parameter corresponding to each target; using the three-dimensional point cloud corresponding to the target and the importance weight graph to combine the running performance data information to simplify the topological structure of the three-dimensional grid model to generate a simplified grid; and according to the face sheet quantity reduction ratio of the simplified grid, adaptively regulating the performance index of the three-dimensional radar dynamic visualization running. The system comprises modules corresponding to the method steps.

Description

TECHNICAL FIELD

[0001] The application provides an adaptive regulation method and system suitable for three-dimensional radar dynamic visualization, and belongs to the technical field of three-dimensional radar visualization control. BACKGROUND

[0002] In the technical field of three-dimensional radar dynamic visualization, with the continuous progress of science and technology, the application demand for this technology is increasing, and it is widely used in many fields such as military monitoring, traffic control, weather forecasting, etc., and plays a key role in real-time acquisition and analysis of target information in complex environments.

[0003] In the prior art, three-dimensional radar dynamic visualization has many deficiencies in data processing and performance regulation. On the one hand, in the acquisition of basic three-dimensional radar data information and real-time acquisition of running performance data information, traditional methods are difficult to achieve comprehensive, accurate and efficient acquisition, and cannot meet the requirements of timeliness and integrity of data in complex and variable scenarios. For example, in some large-scale traffic scenarios or military monitoring scenarios, due to the missing or delay of the collected data, the positioning and tracking of the target deviate.

[0004] On the other hand, for the processing of basic three-dimensional radar data information, the prior art lacks an effective dynamic threat prediction mechanism. Most traditional methods do not fully consider the spatio-temporal characteristics of the data, and cannot accurately predict the dynamic threat indicator parameters of the target, making it difficult to make accurate judgments and countermeasures in advance when facing fast-moving or potentially threatening targets. For example, in weather monitoring, the threat level of a cloud layer that may develop into severe weather cannot be accurately predicted, affecting the accuracy and timeliness of weather warnings.

[0005] In addition, in generating the importance weight map of the associated target and point cloud cluster, the prior art cannot well combine the three-dimensional scene and the dynamic threat indicator parameters of the target, resulting in that the generated weight map cannot accurately reflect the actual importance and relevance of the target, and further affecting the subsequent processing of the three-dimensional grid model. In the topology simplification process of the three-dimensional grid model, the existing method often fails to fully combine the corresponding three-dimensional point cloud, importance weight map and running performance data information of the target, and the generated simplified grid cannot guarantee the visualization effect while realizing the effective optimization of the performance indicators of the three-dimensional radar dynamic visualization. Finally, in the adaptive regulation of the performance indicators of the three-dimensional radar dynamic visualization, due to the deficiencies of the above-mentioned links, the prior art cannot accurately and adaptively regulate the performance indicators according to the actual situation, resulting in problems such as poor visualization effect, slow processing speed and excessive resource consumption during the running of the system, which seriously affects the efficiency of the three-dimensional radar dynamic visualization technology in practical applications.

[0006] In summary, the existing three-dimensional radar dynamic visualization technology has obvious defects and deficiencies in data acquisition, processing, model simplification, and performance regulation, and an adaptive regulation method is urgently needed to solve these problems to improve the accuracy, efficiency, and stability of three-dimensional radar dynamic visualization technology in various fields of application. SUMMARY

[0007] The present application provides an adaptive regulation method and system suitable for three-dimensional radar dynamic visualization to solve the technical problems existing in the prior art, and the technical solutions adopted are as follows:

[0008] An adaptive regulation method suitable for three-dimensional radar dynamic visualization, the adaptive regulation method comprising:

[0009] Real-time acquisition of basic three-dimensional radar data information and running performance data information;

[0010] Using a space-time forgetting mechanism to perform ADST-TPM dynamic threat prediction on the basic three-dimensional radar data information, and obtaining an ADST-TPM dynamic threat index parameter corresponding to each target Target_ID;

[0011] Generating an importance weight map covering all target sets according to the three-dimensional scene and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID, wherein the importance weight map is used to associate a specific target Target_ID and the point cloud cluster corresponding to the target Target_ID;

[0012] Using the three-dimensional point cloud corresponding to the target Target_ID and the importance weight map in combination with the running performance data information to perform topological structure simplification on the three-dimensional grid model to generate a simplified grid;

[0013] According to the face number reduction ratio of the simplified grid, the performance index of three-dimensional radar dynamic visualization running is adaptively regulated.

[0014] Further, real-time acquisition of basic three-dimensional radar data information and running performance data information includes:

[0015] During the operation of the three-dimensional radar, the basic three-dimensional radar data information of the three-dimensional radar is real-time retrieved; wherein the basic three-dimensional radar data information includes original radar data, target tracking data, geographic fence data, multi-target relationship data, user interaction enhancement data, and system load data;

[0016] Real-time acquisition of running performance data information during the three-dimensional radar dynamic visualization process; wherein the running performance data information includes frame rate, GPU occupancy rate, ADST-TPM latency, and point cloud throughput.

[0017] Further, the ADST-TPM dynamic threat prediction is performed on the basic three-dimensional radar data information by using the space-time forgetting mechanism to obtain the ADST-TPM dynamic threat index parameters corresponding to each target, including:

[0018] The radar scanning period Δt is called from the original radar data contained in the basic three-dimensional radar data information, and the cache window length of the target data is set according to the radar scanning period Δt, and the target data is cached in the radar scanning period Δt;

[0019] The radial acceleration, stability index and peak frequency shift of the target Target_ID are called from the target tracking data contained in the basic three-dimensional radar data information, and the Doppler dynamics feature vector is generated by using the radial acceleration, stability index and peak frequency shift ; wherein, A represents the radial acceleration of the target Target_ID; stability represents the stability index of the target Target_ID; f peak represents the peak frequency shift of the target Target_ID;

[0020] The geographic fence list and target running data corresponding to the target Target_ID are called, and the space-time context feature vector corresponding to the target Target_ID is obtained according to the geographic fence list and target running data , wherein, W 01 and W 02 respectively represent the geographic threat correlation and target interaction threat corresponding to the target Target_ID;

[0021] The Doppler dynamics feature vector and the space-time context feature vector are weighted and fused to generate a fused feature vector;

[0022] The GPU utilization of the current system is monitored in real time, and a trained threat assessment model is selected according to the GPU utilization; wherein the threat assessment model includes a complete GRU network model, a lightweight time series convolution model and a lightweight time series convolution model;

[0023] The fused feature vector is input into the threat assessment model, and the initial threat probability corresponding to the target Target_ID is output, and the initial threat probability is forgotten by using the space-time forgetting mechanism to obtain the ADST-TPM dynamic threat index parameters corresponding to the target Target_ID.

[0024] Further, the geographic fence list corresponding to the target Target ID and the target running data are called, and the spatio-temporal context feature vector corresponding to the target Target ID is obtained according to the geographic fence list and the target running data, including:

[0025] The geographic fence list is traversed to obtain the minimum distance between the position of the target Target ID and the geographic fence;

[0026] The priority value corresponding to the fence in the geographic fence list is called;

[0027] The minimum distance between the position of the target Target ID and the geographic fence is combined with the priority value corresponding to the fence in the geographic fence list to obtain the geographic threat correlation degree corresponding to the target Target ID;

[0028] The neighborhood target within 500m of the target Target ID is searched, and the neighborhood target with an initial threat probability greater than 0.7 is selected as a high-threat neighborhood target;

[0029] The target interaction threat corresponding to the target Target ID is obtained by the speed vector of the target Target ID and the speed vector of the high-threat neighborhood target.

[0030] Further, the GPU utilization of the current system is monitored in real time, and a trained threat assessment model is selected according to the GPU utilization, including:

[0031] When the GPU utilization is less than 0.7, a complete GRU network model is selected as the threat assessment model;

[0032] When the GPU utilization is greater than or equal to 0.7 and less than 0.83, a lightweight time series convolution model is selected as the threat assessment model;

[0033] When the GPU utilization is greater than or equal to 0.83, a rule engine model is selected as the threat assessment model;

[0034] The pre-defined rules are as follows:

[0035] First, check the current GPU load, and when the GPU load exceeds 90%, enable the rule engine degradation mode;

[0036] Define a rule engine function to create a function named rule_engine, which receives the fused feature vector of the target Target ID as input and outputs a floating-point value representing the threat degree;

[0037] The execution rules of the rule_engine function are as follows:

[0038] Rule one: if the radial acceleration of the target Target_ID target.a_radial is greater than 20.0, and the distance of the target Target_ID to the geo-fence target.d_geo is less than 5000.0, it is determined that the high-speed approaching no-fly zone is a high threat situation, and the initial threat probability 0.9 is returned;

[0039] Rule two: check the peak frequency shift attribute target.Δf_peak of the target Target_ID, if the peak frequency shift attribute target.Δf_peak is greater than 50, it is determined that the Doppler spectrum changes dramatically, and the initial threat probability 0.7 is returned;

[0040] Rule three: check the target interaction threat attribute target.neighbor_threat of the target, if the value is greater than 1.5, calculate the initial threat probability by the function min(0.8, 0.3+target.neighbor_threat*0.2), and return the calculation result.

[0041] Rule four: default initial threat probability If none of the above rules are met, it means that the target threat level is low, and the default initial threat probability 0.3 is returned.

[0042] Further, according to the three-dimensional scene and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID, an importance weight map covering all target sets is generated, including:

[0043] The three-dimensional scene is divided into a regular voxel grid with a size of 20m×20m×20m, and each voxel grid is scanned to capture the targets Target_ID appearing in each voxel grid;

[0044] The decision-making combined weight corresponding to the target Target_ID appearing in each voxel grid and the initial weight corresponding to the voxel grid are called;

[0045] The maximum value of the decision-making combined weight corresponding to the target Target_ID appearing in each voxel grid and the initial weight corresponding to the voxel grid is taken as the calibrated weight corresponding to each voxel grid;

[0046] The calibrated weight is subjected to anisotropic diffusion processing to obtain an anisotropic diffusion processed weight value;

[0047] The anisotropic diffusion processed weight value is normalized to obtain the maximum weight value corresponding to each voxel grid;

[0048] The maximum weight value corresponding to the voxel grid is encoded into a GPU directly renderable texture format to generate an importance weight map of all target sets.

[0049] Further, the three-dimensional point cloud corresponding to the target Target_ID and the importance weight map are combined with the running performance data information to simplify the topological structure of the three-dimensional grid model to generate a simplified grid, including:

[0050] Converting the three-dimensional point cloud coordinates into texture coordinates to generate three-dimensional point cloud texture coordinates;

[0051] Comparing the three-dimensional point cloud texture coordinates and the importance weight map to obtain the weight value corresponding to each point cloud;

[0052] Comparing the weight value corresponding to each point cloud with the first weight threshold and the second weight threshold, and determining high-weight targets, medium-weight targets, and low-weight targets according to the comparison results; wherein the first weight threshold is set to 0.6, and the second weight threshold is set to 0.3;

[0053] Performing edge collapse simplification processing on the high-weight targets to obtain the geometric details corresponding to the high-weight region;

[0054] For the medium-weight targets, only the medium-weight targets with a weight value exceeding 0.45 are subjected to edge collapse simplification to obtain the key features corresponding to the medium-weight region;

[0055] Performing vertex merging on the low-weight targets to obtain the basic feature corresponding to the low-weight region;

[0056] Integrating the geometric details corresponding to the high-weight region, the key features corresponding to the medium-weight region, and the basic features corresponding to the low-weight region contained in all voxel grids to generate a simplified grid.

[0057] Further, the face sheet quantity reduction ratio corresponding to the simplified grid is used to adaptively regulate and control the performance indicators of the three-dimensional radar dynamic visualization running, including:

[0058] Determining the face sheet quantity of the current simplified grid according to the simplified grid, and retrieving the face sheet quantity of the grid before simplification;

[0059] Obtaining the face sheet quantity reduction ratio corresponding to the current simplified grid according to the face sheet quantity of the current simplified grid and the face sheet quantity of the grid before simplification;

[0060] Retrieving the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID appearing in the current screen;

[0061] Using the face sheet quantity reduction ratio corresponding to the current simplified grid and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID to adaptively regulate and control the frame rate and display bandwidth occupancy rate of the current three-dimensional radar dynamic visualization running.

[0062] Further, the frame rate and display bandwidth occupancy rate of the current three-dimensional radar dynamic visualization operation are adaptively regulated by using the current simplified grid corresponding to the patch quantity reduction ratio and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID, including:

[0063] The ADST-TPM dynamic threat index parameter corresponding to each target Target_ID and the maximum weight value corresponding to the voxel grid where each target Target_ID is located are called.

[0064] The threat degree coefficient is obtained by weighted average processing of the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID and the maximum weight value corresponding to the voxel grid where each target Target_ID is located.

[0065] The patch quantity reduction ratio corresponding to the current simplified grid is called.

[0066] The frame rate and display bandwidth occupancy rate after regulation are obtained by using the patch quantity reduction ratio corresponding to the current simplified grid and the threat degree coefficient.

[0067] An adaptive regulation system suitable for three-dimensional radar dynamic visualization, the adaptive regulation system comprises:

[0068] A data acquisition module is configured to acquire basic three-dimensional radar data information and operation performance data information in real time.

[0069] A dynamic threat index parameter acquisition module is configured to use a space-time forgetting mechanism to perform ADST-TPM dynamic threat prediction on the basic three-dimensional radar data information, and obtain an ADST-TPM dynamic threat index parameter corresponding to each target Target_ID.

[0070] An importance weight map acquisition module is configured to generate an importance weight map covering all target sets according to a three-dimensional scene and an ADST-TPM dynamic threat index parameter corresponding to each target Target_ID, wherein the importance weight map is used to associate a specific target Target_ID and a point cloud cluster corresponding to the target Target_ID.

[0071] A simplified grid acquisition module is configured to use a three-dimensional point cloud corresponding to a target Target_ID and an importance weight map in combination with operation performance data information to simplify the topological structure of a three-dimensional grid model and generate a simplified grid.

[0072] An adaptive regulation module is configured to adaptively regulate the performance index of three-dimensional radar dynamic visualization operation according to the patch quantity reduction ratio corresponding to the simplified grid.

[0073] The present application has the following advantages:

[0074] The adaptive regulation method and system suitable for three-dimensional radar dynamic visualization proposed by the application collect two types of data in real time to ensure the timeliness and comprehensiveness of information, lay a solid foundation for accurate processing of subsequent links, and avoid processing deviation caused by data lag or loss. Through the ADST-TPM dynamic threat prediction based on the time-space forgetting mechanism, the dynamic threat characteristics of the target can be accurately captured, and the accuracy and dynamic adaptability of the target threat evaluation are effectively improved, so that the threat index parameter is more in line with the actual state change of the target. At the same time, the importance weight map generated based on the three-dimensional scene and the target threat index clearly associates the target with the point cloud cluster, and provides accurate importance basis for simplifying the topology structure of the three-dimensional grid model. Combined with the target three-dimensional point cloud, the importance weight map and the running performance data, the grid simplification can realize reasonable simplification of non-key areas while ensuring the visualization accuracy of key targets, and the simplified grid significantly reduces the complexity of the model under the premise of considering the visualization effect. Moreover, according to the running performance index based on the face number reduction ratio of the simplified grid, the adaptive regulation can dynamically balance the visualization effect and system running efficiency. This not only optimizes the resource occupation of the system and improves the speed of data processing and visualization, but also automatically adjusts according to the actual running performance, so that the three-dimensional radar dynamic visualization system can maintain stable and efficient running state in different scenes, and greatly improves the overall performance and application range of the system. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 The method described in the application corresponds to a flowchart;

[0076] Figure 2 The system described in the application corresponds to a system block diagram. DETAILED DESCRIPTION

[0077] The preferred embodiments of the application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0078] The adaptive regulation method suitable for three-dimensional radar dynamic visualization proposed by the embodiment of the application comprises: Figure 1 As shown in the figure, the adaptive regulation method comprises:

[0079] Real-time acquisition of basic three-dimensional radar data information and running performance data information;

[0080] ADST-TPM dynamic threat prediction of the basic three-dimensional radar data information is performed by using a time-space forgetting mechanism, and ADST-TPM dynamic threat index parameters corresponding to each target Target_ID are obtained;

[0081] According to the three-dimensional scene and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID, an importance weight graph covering all target sets is generated, wherein the importance weight graph is used to associate a specific target Target_ID and a point cloud cluster corresponding to the target Target_ID;

[0082] The three-dimensional grid model (3DMesh) is topologically simplified to generate a simplified grid by using the three-dimensional point cloud corresponding to the target Target_ID and the importance weight graph in combination with the running performance data information.

[0083] According to the face number reduction ratio corresponding to the simplified grid, the performance index of the three-dimensional radar dynamic visualization running is adaptively adjusted.

[0084] Among them, the real-time acquisition of basic three-dimensional radar data information and running performance data information includes:

[0085] In the process of three-dimensional radar running, the basic three-dimensional radar data information of the three-dimensional radar is real-time retrieved; wherein the basic three-dimensional radar data information includes original radar data, target tracking data, geographic fence data, multi-target relationship data, user interaction enhancement data and system load data;

[0086] The running performance data information in the process of three-dimensional radar dynamic visualization is real-time collected; wherein the running performance data information includes frame rate, GPU occupancy rate, ADST-TPM delay and point cloud throughput.

[0087] The working principle of the above technical solution is that: the technical solution realizes the adaptive adjustment of three-dimensional radar dynamic visualization through the cooperative operation of multiple links. First, the basic three-dimensional radar data information and the running performance data information are real-time collected to provide original data support for subsequent processing. Then, the ADST-TPM dynamic threat prediction is performed on the basic three-dimensional radar data information by using the space-time forgetting mechanism, focusing on each target Target_ID, and the corresponding ADST-TPM dynamic threat index parameter is calculated to quantify the threat degree of the target. Next, the importance weight graph covering all target sets is generated by combining the three-dimensional scene and the ADST-TPM dynamic threat index parameter of each target, which establishes the association between the specific target Target_ID and the point cloud cluster corresponding thereto, and clearly defines the importance level of different targets in the three-dimensional scene. On this basis, the three-dimensional grid model (3DMesh) is topologically simplified by using the three-dimensional point cloud corresponding to the target, the importance weight graph, and the running performance data information, and a simplified grid is obtained. Finally, the performance index of the three-dimensional radar dynamic visualization running is adaptively adjusted according to the face number reduction ratio corresponding to the simplified grid, forming a complete closed loop from data acquisition, processing, model optimization to performance adjustment.

[0088] The effect of the above technical solution is that: at the data processing level, real-time collection of two types of data ensures the timeliness and comprehensiveness of the information, laying a solid foundation for accurate processing of subsequent links, and avoiding processing deviation caused by data lag or loss. Through the ADST-TPM dynamic threat prediction based on the space-time forgetting mechanism, the dynamic threat characteristics of the target can be accurately captured, and the accuracy and dynamic adaptability of the target threat assessment can be effectively improved, so that the threat index parameters are more in line with the actual state change of the target.

[0089] In terms of model optimization, the importance weight map generated based on the three-dimensional scene and the target threat index clearly associates the target with the point cloud cluster, providing accurate importance basis for the topology simplification of the three-dimensional grid model. Combined with the target three-dimensional point cloud, the importance weight map and the running performance data, the grid simplification can ensure the visualization accuracy of the key target while reasonably simplifying the non-key area, and the generated simplified grid significantly reduces the complexity of the model while considering the visualization effect.

[0090] In terms of performance regulation, the running performance index is regulated adaptively according to the face number reduction ratio of the simplified grid, which can dynamically balance the visualization effect and system running efficiency. This not only optimizes the resource occupation of the system and improves the speed of data processing and visualization, but also automatically adjusts according to the actual running performance, ensuring that the three-dimensional radar dynamic visualization system can maintain stable and efficient running state in different scenes, greatly improving the overall performance and application range of the system.

[0091] In an embodiment of the present application, the ADST-TPM dynamic threat prediction is performed on the basis of three-dimensional radar data information using the space-time forgetting mechanism, and the ADST-TPM dynamic threat index parameters corresponding to each target are obtained, including:

[0092] The radar scanning period Δt is retrieved from the original radar data contained in the basis three-dimensional radar data information, and the cache window duration of the target data is set according to the radar scanning period Δt, and the target data is cached at the radar scanning period Δt; wherein the value range of the cache window duration is 10Δt-20Δt; the target data refers to the motion and signal related data of a single identified target (distinguished by Target_ID) in a certain time window in the radar monitoring scene, which specifically includes three types:

[0093] Position sequence (P[t]): the coordinates of the target in the three-dimensional space (x, y, z) at different times t, reflecting the motion trajectory.

[0094] Radial velocity sequence (v_radial[t]): calculated based on radar Doppler shift, reflecting the speed change of the target towards / away from the radar.

[0095] Doppler spectrum matrix (S[t,f]): the power value corresponding to different time t and different frequency f, used for analyzing the target motion characteristics (such as whether there is multi-component motion, whether there is interference, etc.).

[0096] The radial acceleration, stability index and peak frequency shift of the target Target_ID are retrieved from the target tracking data contained in the basic three-dimensional radar data information, and the radial acceleration, stability index and peak frequency shift are used to generate a Doppler dynamics feature vector ; wherein, A represents the radial acceleration of the target Target_ID; stability represents the stability index of the target Target_ID; f peak represents the peak frequency shift of the target Target_ID;

[0097] wherein, the stability index has a value range of 0-1, and is obtained by the following formula:

[0098] ;

[0099] wherein, stability ( t ) represents the stability index of the target Target_ID at time t; P norm ( f , t ) represents the normalized power spectrum at time t , frequency f , which is obtained by dividing the power spectrum value S ( f , t ) at this time by the sum of the power spectrum of all frequency bins; n represents the total number of frequency bins, i.e. the total number of discrete intervals of the spectrum; ε represents a minimum number, used to avoid P norm ( f , t ) = 0 when the logarithmic operation has no meaning;

[0100] The geographical fence list and target running data corresponding to the target Target_ID are retrieved, and the spatio-temporal context feature vector corresponding to the target Target_ID is obtained according to the geographical fence list and target running data , wherein, W 01 and W 02 respectively represent the geographical threat correlation and target interaction threat corresponding to the target Target_ID;

[0101] The Doppler dynamics feature vector and the space-time context feature vector are weighted and fused to generate a fused feature vector;

[0102] Real-time monitoring of the GPU utilization of the current system, according to the GPU utilization, select the trained threat assessment model; wherein, the threat assessment model includes a complete GRU network model, a light time convolution model and a light time convolution model;

[0103] The fused feature vector is input into the threat assessment model, and the initial threat probability corresponding to the target Target_ID is output, and the initial threat probability is forgotten by using the space-time forgetting mechanism, and the ADST-TPM dynamic threat index parameter corresponding to the target Target_ID is obtained.

[0104] The working principle of the above technical solution is: the technical solution is developed around obtaining the ADST-TPM dynamic threat index parameter corresponding to the target, and dynamic threat prediction is realized through multi-step data processing and analysis. First, the radar scanning period Δt is called from the original radar data of the basic three-dimensional radar data information, so as to set the cache window length of the 10Δt-20Δt value range, and the target data containing the position sequence, radial velocity sequence and Doppler spectrum matrix are cached according to Δt, and the basic data is reserved for subsequent processing.

[0105] Then, the radial acceleration, stability index and peak frequency shift of the target are extracted from the target tracking data, wherein the stability index is calculated by a formula based on the normalized power spectrum, and then a Doppler dynamics feature vector is generated. At the same time, the space-time context feature vector containing the geographical threat correlation and target interaction threat is obtained by combining the geographical fence list and running data of the target, and the two types of feature vectors are weighted and fused to obtain a fused feature vector.

[0106] After that, the system GPU utilization is monitored in real time, and the trained threat assessment model (complete GRU network model or light time convolution model) is selected according to the utilization rate, the fused feature vector is input into the selected model to obtain the initial threat probability, and finally the initial threat probability is forgotten by using the space-time forgetting mechanism, and the ADST-TPM dynamic threat index parameter corresponding to the target is finally obtained.

[0107] The effect of the above technical solution is: by setting a reasonable cache window length and caching target data according to a radar scanning period, the integrity and timeliness of target motion and signal related data are ensured, and a high-quality data foundation is provided for subsequent feature extraction and threat prediction. The quantitative calculation of the stability index accurately reflects the stability degree of the target signal, and enhances the accuracy of feature description. The weighted fusion of Doppler dynamics features and spatio-temporal context features comprehensively considers the target motion characteristics and external environmental correlation factors, so that the fused feature vector can more comprehensively represent the threat-related attributes of the target, and the comprehensiveness and reliability of threat assessment are improved. The model selection mechanism dynamically selects the threat assessment model according to the GPU utilization, realizes the adaptation of the calculation resources and the model complexity, ensures the assessment accuracy, effectively optimizes the system running efficiency, and avoids the problems of resource waste or insufficient performance. At the same time, the spatio-temporal forgetting mechanism processes the initial threat probability, can weaken the influence of outdated information, highlight the current key information, make the final ADST-TPM dynamic threat index parameter more consistent with the real-time state of the target, and improve the timeliness and accuracy of dynamic threat prediction, providing a reliable threat assessment basis for subsequent adaptive regulation of three-dimensional radar dynamic visualization.

[0108] In an embodiment of the present application, the geographic fence list corresponding to the target Target_ID and the target running data are called, and the spatio-temporal context feature vector corresponding to the target Target_ID is obtained according to the geographic fence list and the target running data, including:

[0109] The minimum distance between the position of the target Target_ID and the geographic fence is obtained by traversing the geographic fence list;

[0110] The priority value corresponding to the fence is called from the geographic fence list, wherein the priority value is a data value set by a person;

[0111] The minimum distance between the position of the target Target_ID and the geographic fence is obtained by traversing the geographic fence list;

[0112] And the geographic threat correlation degree corresponding to the target Target_ID is obtained by combining the minimum distance between the position of the target Target_ID and the geographic fence with the priority value corresponding to the fence called from the geographic fence list, wherein the geographic threat correlation degree reflects the threat degree generated by the spatial correlation between the target Target_ID and the geographic fence;

[0113] ;

[0114] Wherein, W 01 The geographic threat correlation degree corresponding to the target Target_ID is represented by Target_ID; i The sequence number corresponding to the fence is represented by Fence_ID.K Represents a collection of geofences; B i Indicates the i The priority value corresponding to each fence is a value used to indicate the importance or processing order of the fence in a specific system or scenario, and is set using methods known in the art, including but not limited to setting based on fence function, setting based on risk level, and setting based on historical data or user behavior; d i Indicates the target Target_ID to i Minimum distance between fences; L Indicates the geographical threat attenuation length, ranging from 800m to 1200m, preferably 1000m; specifically, the fence priority value B is introduced into the formula i , the threat contribution of fences can be measured differently according to actual scenarios (such as functions, risk levels, etc.) to avoid indiscriminate treatment; the minimum distance d i Combined with the exponential decay exp(−d i / L), accurately simulating the spatial pattern in which threats weaken as distance increases, allowing geographic threat correlation calculations to align with the real-world threat attenuation logic and improve the accuracy of spatial correlation threat quantification.

[0115] Search for neighboring targets within 500 meters of the target Target_ID and select those with an initial threat probability greater than 0.7 as high-threat neighboring targets;

[0116] The target interaction threat corresponding to the target Target_ID is obtained by comparing the velocity vector of the target Target_ID with the velocity vector of the high-threat neighboring targets. The target interaction threat reflects the threat level Target_ID generated by the interaction between the target and the surrounding high-threat neighboring targets.

[0117] Furthermore, the target interaction threat is obtained by the following formula:

[0118] ;

[0119] in, W 02 Indicates the target interaction threat corresponding to the target Target_ID; v t Represents the velocity vector of the target Target_ID; n Indicates the total number of high-threat neighborhood targets; v j represents the velocity vector of the jth high-threat neighborhood target; d j Indicates the distance between the jth high-threat neighborhood target and the target Target_ID;L n represents a preset interaction threat attenuation length, the value range is 120m-240m, preferably 200m; represents the vector norm L2 operation. Specifically, high-threat neighborhood targets with initial threat probability > 0.7 are screened, high-risk interaction objects are focused, and low-threat interference is reduced; the norm operation of the speed vector difference (v t −v j ) combined with the distance exponential attenuation exp(−d j / L n ), not only considers the relative motion trend (speed vector difference), but also reflects the influence of distance on interaction threat, accurately depicts the threat degree generated by the interaction of the target and the high-threat neighborhood target, and improves the quantitative accuracy of the interaction threat.

[0120] The working principle of the above technical solution is that the above technical solution of the embodiment realizes the calculation through the geographical threat correlation degree and the target interaction threat by obtaining the space-time context feature vector corresponding to the target Target_ID.

[0121] In terms of geographical threat correlation degree calculation, first, the minimum distance of the target to each geographical fence is obtained by traversing the geographical fence list; at the same time, the priority value set by each fence is called, and then the minimum distance, the priority value, and the preset geographical threat attenuation length (800m-1200m, preferably 1000m) are operated by using a specific formula to obtain the geographical threat correlation degree W01 reflecting the spatial correlation threat degree of the target and the geographical fence.

[0122] In terms of target interaction threat calculation, first, the neighborhood targets within 500m of the target are found, and high-threat neighborhood targets with an initial threat probability greater than 0.7 are screened; then, based on the speed vector of the target, the speed vector of each high-threat neighborhood target, and the distance between the target and each high-threat neighborhood target, combined with the preset interaction threat attenuation length (120m-240m, preferably 200m) and the vector norm operation, the target interaction threat W02 reflecting the interaction threat degree of the target and the surrounding high-threat neighborhood targets is calculated by the corresponding formula. The geographical threat correlation degree W01 and the target interaction threat W02 jointly constitute the space-time context feature vector of the target.

[0123] The effect of the above technical solution is that in the calculation of the geographical threat correlation degree, by introducing the priority value of the geographical fence, the differentiated influence of different geographical areas on threat assessment is highlighted, and by setting the minimum distance and the geographical threat attenuation length, the quantification of the geographical threat is more in line with the actual spatial relationship, which can accurately reflect the threat degree generated by the spatial correlation between the target and the geographical fence, and improves the accuracy of the geographical dimension threat feature.

[0124] In the target interaction threat calculation, the 500m range is limited, and the high threat neighborhood target with an initial threat probability greater than 0.7 is screened, the object that can really produce interaction threat is focused on, and the interference of irrelevant targets is avoided; meanwhile, the speed vector, distance and interaction threat attenuation length are taken into account, the influence of the motion relationship and spatial distance between targets on the interaction threat is comprehensively considered, the quantification of the target interaction threat is more reasonable, and the dynamic interaction risk between the target and the surrounding high threat object is effectively captured.

[0125] Meanwhile, the prior art makes a binary determination of the geographic fence threat, and the present scheme realizes semantic quantification of the geographic space threat by priority dynamic assignment (B i ) + distance attenuation modeling, converts semantic information such as function attribute and risk level of the geographic fence into a mathematically calculable threat weight, and realizes semantic quantification of the geographic space threat. For example, the'security check area' and 'runway area' of the airport fence can be distinguished by threat logic through priority, break through the traditional rough model that 'distance determines everything', and deeply bind the geographic threat and business semantics. The'vector dynamics' of the interaction threat captures only the 'distance proximity' or'speed speed' single dimension of the traditional target interaction threat, and the present embodiment accurately captures the'relative motion trend' through the norm operation of the speed vector difference (v t -v j ), the threat difference between 'head-on approach' and'same direction away' of two cars can be reflected by the length of the vector difference. This dynamic vector interaction modeling upgrades the threat assessment from'static distance correlation' to 'dynamic motion trend prediction', effectively improves the accuracy of the threat assessment and the response timeliness.

[0126] In one embodiment of the present application, the GPU utilization of the current system is monitored in real time, and a trained threat assessment model is selected according to the GPU utilization, including:

[0127] When the GPU utilization is less than 0.7, a complete GRU network model is selected as the threat assessment model, the model contains 3 layers of hidden units, is suitable for using the strong feature learning ability to reason when the GPU resources are relatively abundant, and is suitable for using the strong feature learning ability to reason when the GPU resources are relatively abundant;

[0128] When the GPU utilization is greater than or equal to 0.7 and less than 0.83, a light time convolution model is selected as the threat assessment model, the light time convolution model adopts a hollow convolution (the hollow convolution kernel is 3), has relatively low demand for GPU resources while ensuring a certain reasoning effect, and is suitable for the scene with medium GPU load;

[0129] When the GPU utilization is greater than or equal to 0.83, a rule engine model is selected as a threat assessment model, the rule engine model relies on a predefined rule, hardly consumes GPU computing power, and is used in a scenario of extreme shortage of GPU resources to ensure basic reasoning process;

[0130] The predefined rule is as follows:

[0131] First, the current GPU load is checked, and when the GPU load exceeds 90%, the rule engine degradation mode is enabled, and the purpose is to quickly calculate key features through a predefined rule when GPU resources are extremely scarce, and to ensure basic function operation;

[0132] The rule engine function is defined to create a function named rule_engine, which receives the fused feature vector of the target Target_ID as input and outputs a floating-point value representing the threat level.

[0133] The execution rule of the rule_engine function is as follows:

[0134] Rule one: if the radial acceleration of the target Target_ID target.a_radial is greater than 20.0 (meaning that the target is accelerating in the radial direction), and the distance of the target Target_ID to the geographic fence target.d_geo is less than 5000.0 (indicating proximity to a sensitive geographic area such as a no-fly zone), then it is determined as a high threat situation of high-speed approach to a no-fly zone, and the initial threat probability of 0.9 is returned.

[0135] Rule two: check the peak frequency shift attribute target.Δf_peak of the target Target_ID, if the peak frequency shift attribute target.Δf_peak is greater than 50 (indicating that the peak frequency of the Doppler spectrum has changed significantly, which may be target maneuvering or interference), it is determined as a threat scenario of dramatic change of Doppler spectrum, and the initial threat probability of 0.7 is returned.

[0136] Rule three: check the target interaction threat attribute target.neighbor_threat of the target, if the value is greater than 1.5 (indicating that there are more high-threat neighbor targets), the initial threat probability is calculated by the function min(0.8, 0.3+target.neighbor_threat*0.2), and the calculation result is returned.

[0137] Rule four: default initial threat probability is returned If all the above rule conditions are not met, it means that the target threat level is low, and the default initial threat probability of 0.3 is returned.

[0138] In summary, when the GPU load is too high, the rule_engine function is used to judge the target threat according to the three rules of "high-speed approach to no-fly zone, dramatic change of Doppler spectrum, and high-threat target cluster" in sequence. If the corresponding rule is met, the corresponding initial threat probability is returned. If none of the rules are met, the default low initial threat probability is returned.

[0139] The working principle of the above technical solution is that the core of the technical solution is to dynamically select and adapt the trained threat assessment model according to the real-time GPU utilization of the system to balance the threat assessment performance and GPU resource consumption. The GPU utilization of the current system is monitored in real time, and the model is selected according to different thresholds: when the GPU utilization is less than 0.7, the complete GRU network model is selected, which contains 3 layers of hidden units, and relies on strong feature learning ability to perform reasoning when GPU resources are abundant; when the GPU utilization is greater than or equal to 0.7 and less than 0.83, a lightweight time series convolution model is selected, which uses a convolution kernel of 3, reduces the GPU resource requirement while ensuring a certain reasoning effect, and adapts to medium load scenarios; when the GPU utilization is greater than or equal to 0.83, a rule engine model is enabled, which relies on predefined rules and almost does not consume GPU power, and is used in resource-intensive scenarios to ensure basic reasoning. When the GPU load exceeds 90%, a degradation mode is enabled, and the fused feature vector is processed by the rule_engine function. The function judges in sequence according to the preset rules: rule one returns 0.9 for high-speed approach to no-fly zone; rule two returns 0.7 for dramatic change of Doppler spectrum; rule three calculates and returns the corresponding value according to the target interaction threat; if none of the rules are met, rule four returns the default value of 0.3, to quickly complete the threat assessment.

[0140] The effect of the above technical solution is that in terms of resource adaptation, by monitoring the GPU utilization in real time and dynamically switching the model, the threat assessment task is accurately matched with the GPU resources. When resources are abundant, the strong feature learning ability of the complete GRU network model is used to improve the assessment accuracy; when resources are medium, the lightweight time series convolution model controls resource consumption while maintaining good reasoning effect; when resources are extremely scarce, the rule engine model ensures basic functions with extremely low resource consumption, avoiding system crashes or assessment interruptions due to insufficient resources.

[0141] In terms of performance evaluation, the pre-defined rules of the rule engine model make quick judgments for key threat scenarios (such as high-speed approach to sensitive areas, spectrum change, and high-threat clusters), ensuring that high-threat targets can be accurately identified even in resource-limited situations, and the sequential execution logic of the rules makes the evaluation process efficient and orderly. At the same time, the hierarchical adaptation of different models not only takes full advantage of the high-performance model, but also ensures the robustness of the system through the degradation mechanism of the rule engine model, thereby improving the flexibility, efficiency and reliability of threat evaluation, and providing stable and adaptive resource state evaluation results for subsequent dynamic threat index parameter acquisition and three-dimensional radar visualization control.

[0142] In one embodiment of the present application, the initial threat probability is processed using a space-time forgetting mechanism to obtain the ADST-TPM dynamic threat index parameter corresponding to the target Target_ID, including:

[0143] Retrieving the initial threat probability corresponding to the target Target_ID and the initial threat probability corresponding timestamp determined so far;

[0144] Generating a threat probability queue using the initial threat probability corresponding to the target Target_ID and the initial threat probability corresponding timestamp;

[0145] Iterating through each initial threat probability and its timestamp contained in the threat probability queue, and calculating the time difference between the current time and the initial threat probability timestamp according to the current time and the initial threat probability timestamp;

[0146] Setting a time decay factor corresponding to each initial threat probability using the time difference between the current time and the initial threat probability timestamp;

[0147] Performing forgetting processing on all initial threat probabilities obtained so far using the time decay factor corresponding to each initial threat probability to obtain the ADST-TPM dynamic threat index parameter corresponding to the target Target_ID at the current time.

[0148] Wherein, the ADST-TPM dynamic threat index parameter corresponding to the target Target_ID at the current time is obtained by the following formula:

[0149] ;

[0150] Wherein, J ADST-TPM dynamic threat index parameter corresponding to the target Target_ID at the current time; M Total number of times of determining initial threat probability; P s The corresponding probability value of the initial threat probability determined for the first time; s times.R ( t , t c ) s Indicates the s The timestamp corresponding to the initial threat probability determined t c and the current moment t The corresponding time decay factor, and , lambda represents the time decay rate; r s Indicates the s The spatial accelerated forgetting factor corresponding to the initial threat probability determined this time is determined by:

[0151] ;

[0152] Among them, d s Indicates the s The actual distance between the target Target_ID in the initial threat profile determined this time and the fence.

[0153] The working principle of the above technical solution is as follows: This technical solution processes the initial threat probability through a spatiotemporal forgetting mechanism to obtain the ADST-TPM dynamic threat indicator parameter corresponding to the target Target_ID. The core of the technical solution is to weight and integrate the historical threat probabilities by combining time decay and spatial accelerated forgetting factors. First, all the initial threat probabilities of the target to date and their corresponding timestamps are retrieved to generate a threat probability queue. Then, each initial threat probability and timestamp in the queue is traversed, the time difference between the current moment and each timestamp is calculated, and the corresponding time decay factor is set based on this time difference. The time decay factor is determined by the time decay rate λ and the time difference. At the same time, the spatial accelerated forgetting factor is calculated based on the actual distance between the target and the fence when the sth initial threat probability is determined. Finally, each initial threat probability is weighted using the time decay factor and spatial accelerated forgetting factor. All initial threat probabilities are integrated through a formula to obtain the ADST-TPM dynamic threat indicator parameter J corresponding to the target at the current moment.

[0154] The effect of the technical scheme is as follows: in the time dimension, the historical initial threat probability is forgotten by the time attenuation factor, so that the influence of the threat information of an earlier time gradually weakens over time, the weight of recent threat data is highlighted, and it is ensured that the ADST-TPM dynamic threat index parameter can timely reflect the latest change of the target threat, and the time sensitivity and dynamic adaptability of the index are improved. In the spatial dimension, the spatial acceleration forgetting factor is introduced, and the threat probability is adjusted in combination with the actual distance between the target and the fence. When the target is far away from the fence, the influence of the historical threat information is accelerated and weakened, and when the target is close to the fence, the weight of the historical threat information is relatively reserved, so that the threat index parameter is more in line with the spatial position characteristics of the target, and the response capability to the spatial situation change of the target is enhanced. Meanwhile, the comprehensive application of the time-space forgetting mechanism enables the ADST-TPM dynamic threat index parameter to not only take into account the accumulation of historical threat information, but also focus on key threat factors under the current space-time state, thereby effectively improving the accuracy, real-time performance and pertinence of the threat index, and providing a more accurate and reliable threat evaluation basis for the subsequent adaptive regulation and control of the three-dimensional radar dynamic visualization. Meanwhile, the technical scheme of the embodiment solves the problem of how the historical threat decays over time, and supplements the problem of how the spatial distance affects the threat weight, which is different from the traditional threat processing method in a single dimension of time or space, realizes the dynamic decay modeling of the threat in the time-space double dimensions, and effectively improves the accuracy of the dynamic threat index parameter and the adaptability between the dynamic threat index parameter and the actual situation of the time-space jointly affected risk in the current real scene.

[0155] In an embodiment of the present application, an importance weight map covering all target sets is generated according to the three-dimensional scene and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID, which includes:

[0156] The three-dimensional scene is divided into a regular voxel grid with a size of 20m*20m*20m, and each voxel grid is scanned to capture the target Target_ID appearing in each voxel grid.

[0157] The decision combined weight corresponding to the target Target_ID appearing in each voxel grid and the initial weight corresponding to the voxel grid are called.

[0158] The maximum value of the decision combined weight corresponding to the target Target_ID appearing in each voxel grid and the initial weight corresponding to the voxel grid is taken as the calibrated weight corresponding to each voxel grid.

[0159] The calibration weight is anisotropically diffused to obtain a weight value after anisotropic diffusion processing; specifically, a three-dimensional convolution kernel kernel (in the example, a 3x3x3 structure) is created, and the numerical distribution is [[[0.05, 0.2, 0.05], [0.20, 0.0, 0.20], [0.05, 0.2, 0.05]]], the center weight is 0 (the original value is retained), and the surrounding weights are distributed according to a certain rule to control the diffusion direction and strength, achieving the effect of "anisotropy" (different diffusion degrees in different directions). Perform a convolution operation: use the convolve3d function, use the calibration weight as the input, use the kernel as the convolution kernel, and use the nearest (nearest neighbor) mode to perform three-dimensional convolution calculation to obtain the weight value after diffusion. This step is used to diffuse and adjust the voxel weight according to the weight of the surrounding voxels, simulating the weight transfer in space.

[0160] The weight value after anisotropic diffusion processing is normalized to obtain the maximum weight value corresponding to each voxel grid;

[0161] Encode the maximum weight value corresponding to the voxel grid into a GPU directly renderable texture format to generate an importance weight map of all target sets.

[0162] Each target Target_ID corresponds to a decision weight, which is obtained as follows:

[0163] Call the ADST-TPM dynamic threat indicator parameter corresponding to the current target Target_ID;

[0164] The ADST-TPM dynamic threat indicator parameter corresponding to the target Target_ID is nonlinearly amplified to obtain the ADST-TPM dynamic threat indicator parameter after nonlinear amplification;

[0165] Real-time monitoring of the distance between the position of the target Target_ID in the screen and the user's eye movement gaze coordinates;

[0166] Determine the attention degree of the current user to the target Target_ID through the distance between the position of the target Target_ID in the screen and the user's eye movement gaze coordinates;

[0167] The attention degree of the user to the target Target_ID is obtained by the following formula:

[0168] ;

[0169] Where G represents the attention degree of the user to the target Target_ID; d grepresents the distance between the position of the target Target_ID in the screen and the user's eye movement gaze coordinates; and sigma represents a Gaussian attenuation parameter, which is 50 pixels;

[0170] The decision-making weight of the target Target_ID is obtained by combining the user's attention to the target Target_ID with the ADST-TPM dynamic threat indicator parameter of the target Target_ID.

[0171] The working principle of the above technical solution is as follows: The technical solution is developed around the generation of an importance weight map covering all target sets, and is realized through multi-step spatial division, weight calculation and processing. First, a three-dimensional scene is divided into a regular voxel grid according to 20m x 20m x 20m, and each grid is scanned to capture the target Target_ID therein.

[0172] In terms of weight calculation, the decision-making weight corresponding to the target in each voxel grid and the initial weight of the grid are first called, and the maximum value of the two is taken as the calibrated weight of the grid. The acquisition of the decision-making weight needs to go through a series of processing: the ADST-TPM dynamic threat indicator parameter of the target is called and nonlinearly amplified, and at the same time, the distance between the position of the target in the screen and the user's eye movement gaze coordinates is monitored in real time, the user's attention is calculated through the Gaussian attenuation formula, and then the decision-making weight is obtained by combining the amplified threat indicator parameter.

[0173] Subsequently, anisotropic diffusion processing is performed on the calibrated weight, a three-dimensional convolution kernel with a structure of 3 x 3 x 3 is used, a three-dimensional convolution operation is performed in nearest mode, and anisotropic diffusion adjustment of the weight is realized; then the weight value after diffusion is normalized to obtain the maximum weight value of each voxel grid, and finally it is encoded into a texture format that can be directly rendered by GPU to generate an importance weight map.

[0174] The effect of the above technical solution is that in the construction of spatial division and weight basis, the division of regular voxel grid makes the weight calculation of three-dimensional scene have a clear spatial unit, and the maximum value of the decision-making weight of the target and the initial weight of the grid is used to determine the calibrated weight, which highlights the importance brought by the threat of the target itself and the attention of the user, and also retains the inherent basic weight attribute of the grid, providing a reasonable starting point for subsequent weight processing.

[0175] The calculation of the decision-making weight comprehensively considers the dynamic threat indicator of the target (enhanced difference after nonlinear amplification) and the user's attention (accurately quantified based on eye movement data), so that the weight can reflect both the objective threat degree of the target and the subjective attention focus of the user, enhancing the comprehensiveness and pertinence of the weight.

[0176] The anisotropic diffusion processing realizes the spatial transmission of weights through a specific three-dimensional convolution kernel, so that the weights of adjacent voxels influence each other and the diffusion degrees in different directions are different, which is more consistent with the spatial characteristics of target correlation in a three-dimensional scene, and avoids the isolation of weights; the normalization processing unifies the weight scale, and ensures the comparability of weights of different voxels.

[0177] The finally generated GPU directly renderable texture format weight map can efficiently support the rendering process of three-dimensional radar dynamic visualization, so that the visualization effect can accurately reflect the importance distribution of the target set, improve the information transmission efficiency and accuracy of visualization, and provide a reliable importance basis for subsequent three-dimensional mesh model simplification and performance control.

[0178] In an embodiment of the present application, a three-dimensional mesh (3DMesh) is simplified to generate a simplified mesh by using the three-dimensional point cloud corresponding to the target Target_ID and the importance weight map in combination with the running performance data information, including:

[0179] The three-dimensional point cloud coordinates are converted into texture coordinates to generate three-dimensional point cloud texture coordinates;

[0180] The three-dimensional point cloud texture coordinates and the importance weight map are compared to obtain the weight value corresponding to each point cloud;

[0181] The weight value corresponding to each point cloud is compared with a first weight threshold and a second weight threshold, and high weight targets, medium weight targets and low weight targets are determined according to the comparison result; wherein the first weight threshold is set to 0.6, and the second weight threshold is set to 0.3; specifically: the point cloud with a weight value exceeding the first weight threshold is selected as a high weight target; the point cloud with a weight value not exceeding the first weight threshold but not lower than the second weight threshold is selected as a medium weight target; and the point cloud with a weight value lower than the second weight threshold is selected as a low weight target;

[0182] Edge folding simplification processing is performed on the high weight target to obtain geometric details corresponding to the high weight region; wherein the geometric details include but are not limited to the outline, key features and the like of the target;

[0183] For the medium weight target, edge folding simplification is only performed on the medium weight target with a weight value exceeding 0.45 to obtain key features corresponding to the medium weight region; wherein the key features include but are not limited to the approximate shape and motion trend of the target;

[0184] Vertex merging is performed on the low weight target to obtain basic features corresponding to the low weight region; wherein the basic features include but are not limited to the approximate position and existence of the target.

[0185] The geometric details corresponding to the high-weight region, the key features corresponding to the medium-weight region and the basic features corresponding to the low-weight region in all voxel grids are integrated to generate a simplified grid.

[0186] The working principle of the technical solution is that the technical solution generates a simplified grid by simplifying the topological structure of the three-dimensional grid model in multiple levels according to the three-dimensional point cloud, the importance weight map and the running performance data information of the target.

[0187] Firstly, the three-dimensional point cloud coordinates are converted into texture coordinates, and then compared with the importance weight map to obtain the weight value corresponding to each point cloud. Subsequently, according to the first weight threshold (0.6) and the second weight threshold (0.3), the point cloud is divided into a high-weight target (weight>0.6), a medium-weight target (0.3≤weight≤0.6) and a low-weight target (weight<0.3).

[0188] Different simplification strategies are adopted for different weight targets: the high-weight target is subjected to edge folding simplification processing to retain its geometric details; the medium-weight target is subjected to edge folding simplification only on the part with a weight exceeding 0.45 to retain key features; and the low-weight target is subjected to vertex merging to retain basic features. Finally, the features corresponding to different weight regions in all voxel grids are integrated to generate a simplified grid.

[0189] The effect of the technical solution is that the hierarchical simplification strategy based on the weight value enables the geometric details of the high-weight target to be retained, the key features of the medium-weight target to be maintained, and the basic features of the low-weight target to be retained, thereby realizing differentiated processing of grid simplification, ensuring the integrity of important target information, reasonably simplifying secondary targets, and avoiding loss of important information or excessive redundant information caused by indiscriminate simplification.

[0190] In terms of model optimization, the accurate application of different simplification methods such as edge folding and vertex merging effectively reduces the complexity of the three-dimensional grid model while taking into account the overall accuracy of the model. The retention of geometric details in the high-weight region ensures the visualization effect of the key target, the retention of key features in the medium-weight region maintains the basic shape and motion trend recognition of the target, and the retention of basic features in the low-weight region reduces the model data amount with minimal information loss.

[0191] The simplified grid generated by the technical solution can balance the model accuracy and complexity to meet the requirements of the running performance data information, providing an optimized model basis for subsequent performance regulation of three-dimensional radar dynamic visualization, improving the running efficiency of the visualization system, ensuring the effective transmission of key information, and enhancing the adaptability and practicality of the system in different scenarios.

[0192] An embodiment of the present application, according to the simplified grid corresponding to the number of patches adaptive regulation of performance indicators of three-dimensional radar dynamic visualization operation, comprising:

[0193] According to the simplified grid determines the number of patches of the current simplified grid, and call the number of patches before simplification;

[0194] According to the number of patches of the current simplified grid and the number of patches before simplification of the grid corresponding to the number of patches reduction ratio;

[0195] Call the ADST-TPM dynamic threat index parameters corresponding to each target Target_ID appeared in the current screen;

[0196] The number of patches corresponding to the current simplified grid and the ADST-TPM dynamic threat index parameters corresponding to each target Target_ID are used to adaptively regulate the frame rate and display bandwidth occupancy of the current three-dimensional radar dynamic visualization operation.

[0197] Wherein, the number of patches corresponding to the current simplified grid and the ADST-TPM dynamic threat index parameters corresponding to each target Target_ID are used to adaptively regulate the frame rate and display bandwidth occupancy of the current three-dimensional radar dynamic visualization operation, comprising:

[0198] Call the ADST-TPM dynamic threat index parameters corresponding to each target Target_ID and the maximum weight value corresponding to the voxel grid where each target Target_ID is located;

[0199] The ADST-TPM dynamic threat index parameters corresponding to each target Target_ID and the maximum weight value corresponding to the voxel grid where each target Target_ID is located are used to obtain the threat degree coefficient by weighted average processing;

[0200] Call the number of patches corresponding to the current simplified grid;

[0201] The number of patches corresponding to the current simplified grid and the threat degree coefficient are used to obtain the regulated frame rate and display bandwidth occupancy.

[0202] Wherein, the regulated frame rate and display bandwidth occupancy are obtained by the following formula:

[0203] ;

[0204] Wherein, FPS The regulated frame rate is represented by f; FPS old The frame rate before regulation is represented by f0; H The number of patches corresponding to the current simplified grid is represented by r;Y a threat degree coefficient is represented;

[0205]

[0206] wherein, B us a display bandwidth occupancy ratio after regulation is represented; B old a display bandwidth occupancy ratio before regulation is represented; H a patch quantity reduction ratio corresponding to the current simplified grid is represented; Y a threat degree coefficient is represented.

[0207] The working principle of the above technical solution is: the technical solution aims to adaptively regulate the performance indicators of three-dimensional radar dynamic visualization operation according to the patch quantity reduction ratio of the simplified grid, and the core is to realize dynamic adjustment of frame rate and display bandwidth occupancy ratio through quantitative analysis and formula calculation.

[0208] Firstly, the patch quantity of the current simplified grid is determined, and the patch quantity of the grid before simplification is called, and the ratio of the two is calculated to obtain the patch quantity reduction ratio. Then, the ADST-TPM dynamic threat index parameter of each target in the current screen and the maximum weight value of the voxel grid where the target is located are called, and the two parameters are weighted and averaged to generate a threat degree coefficient.

[0209] Subsequently, the patch quantity reduction ratio and the threat degree coefficient are substituted into a specific formula to calculate the frame rate and the display bandwidth occupancy ratio before regulation, and the frame rate (FPS) and the display bandwidth occupancy ratio (Bus) after regulation are obtained. Among them, the regulation formula of the frame rate is based on the frame rate before regulation, combined with the patch quantity reduction ratio and the threat degree coefficient for adjustment; the regulation formula of the display bandwidth occupancy ratio is also based on the value before regulation, and the patch quantity reduction ratio and the threat degree coefficient are used to realize dynamic adaptation.

[0210] The effect of the above technical solution is: in terms of the accuracy of performance regulation, the grid simplification degree is quantified by the patch quantity reduction ratio, combined with the threat degree coefficient (fusing target dynamic threat and voxel weight), so that the adjustment of frame rate and display bandwidth occupancy ratio can closely match the current grid complexity and target threat state. When the grid simplification degree is high (reduction ratio is large) and the threat degree is low, the system can moderately reduce the frame rate and bandwidth occupancy to avoid resource waste; when the threat degree is high, even if the grid is simplified, the coefficient can be adjusted to ensure performance support in critical scenes, realizing accurate matching of performance and demand.

[0211] ​In terms of system resource optimization, the quantitative parameter-based regulation mechanism avoids blind adjustment of performance indicators, enabling frame rate and bandwidth occupancy to dynamically balance according to actual scenarios. Under the premise of ensuring high-threat target visualization effect, resource consumption in low-threat scenarios is effectively reduced, improving the efficiency and stability of system operation. Meanwhile, the unified formulaic regulation logic ensures the consistency and predictability of performance adjustment, reduces the uncertainty of human intervention, and enhances the adaptive ability and overall operation efficiency of the three-dimensional radar dynamic visualization system in complex environments.

[0212] The embodiment of the present application proposes an adaptive regulation system suitable for three-dimensional radar dynamic visualization, as shown in Figure 2 The adaptive regulation system comprises:

[0213] A data acquisition module is configured to acquire basic three-dimensional radar data information and running performance data information in real time.

[0214] A dynamic threat index parameter acquisition module is configured to use a space-time forgetting mechanism to perform ADST-TPM dynamic threat prediction on the basic three-dimensional radar data information, and acquire ADST-TPM dynamic threat index parameters corresponding to each target Target_ID.

[0215] An importance weight map acquisition module is configured to generate an importance weight map covering all target sets according to the three-dimensional scene and the ADST-TPM dynamic threat index parameters corresponding to each target Target_ID, wherein the importance weight map is used to associate specific targets Target_ID and point cloud clusters corresponding to the targets Target_ID.

[0216] A simplified grid acquisition module is configured to use three-dimensional point clouds corresponding to the targets Target_ID and the importance weight map in combination with running performance data information to perform topological structure simplification on a three-dimensional mesh (3DMesh) to generate a simplified grid.

[0217] An adaptive regulation module is configured to regulate the performance indicators of three-dimensional radar dynamic visualization according to the face number reduction ratio of the simplified grid.

[0218] The working principle of the above technical solution is that the technical solution realizes adaptive regulation and control of three-dimensional radar dynamic visualization through the coordinated operation of multiple links. First, real-time acquisition of basic three-dimensional radar data information and running performance data information provides raw data support for subsequent processing. Then, the ADST-TPM dynamic threat prediction is performed on the basic three-dimensional radar data information using the space-time forgetting mechanism, focusing on each target Target_ID, and the corresponding ADST-TPM dynamic threat index parameter is calculated to quantify the threat level of the target. Next, the importance weight map covering all target sets is generated by combining the three-dimensional scene and the ADST-TPM dynamic threat index parameter of each target, which establishes the association between the specific target Target_ID and its corresponding point cloud cluster, and clearly defines the importance level of different targets in the three-dimensional scene. On this basis, the three-dimensional mesh model (3DMesh) is simplified by using the three-dimensional point cloud corresponding to the target, the importance weight map, and the running performance data information, and a simplified mesh is obtained. Finally, the performance index of three-dimensional radar dynamic visualization is adaptively regulated according to the face number reduction ratio of the simplified mesh, forming a complete closed loop from data acquisition, processing, model optimization to performance adjustment.

[0219] The effect of the above technical solution is that in the data processing layer, real-time acquisition of two types of data ensures the timeliness and comprehensiveness of the information, laying a solid foundation for accurate processing of subsequent links and avoiding processing deviation caused by data lag or loss. Through the ADST-TPM dynamic threat prediction by the space-time forgetting mechanism, the dynamic threat characteristics of the target can be accurately captured, effectively improving the accuracy and dynamic adaptability of the target threat evaluation, and making the threat index parameter more consistent with the actual state change of the target.

[0220] In the model optimization aspect, the importance weight map generated based on the three-dimensional scene and the target threat index clearly associates the target with the point cloud cluster, providing accurate importance basis for the topological structure simplification of the three-dimensional mesh model. Combined with the target three-dimensional point cloud, the importance weight map, and the running performance data for mesh simplification, the visualization accuracy of key targets can be ensured while the non-key areas are reasonably simplified, and the generated simplified mesh significantly reduces the complexity of the model while considering the visualization effect.

[0221] In the performance regulation, the running performance index is adaptively regulated according to the face number reduction ratio of the simplified mesh, which can dynamically balance the visualization effect and system running efficiency. This not only optimizes the resource occupation of the system, improves the speed of data processing and visualization, but also automatically adjusts according to the actual running performance, ensuring that the three-dimensional radar dynamic visualization system can maintain stable and efficient running state in different scenarios, greatly improving the overall performance and application range of the system.

Claims

1. An adaptive control method for three-dimensional radar dynamic visualization, characterized in that: The adaptive control method comprises: Real-time collection of basic 3D radar data and operational performance data; Use the spatiotemporal forgetting mechanism to perform ADST-TPM dynamic threat prediction on basic 3D radar data information and obtain the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID; Generate an importance weight map covering all target sets based on the three-dimensional scene and the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID, wherein the importance weight map is used to associate a specific target Target_ID with the three-dimensional point cloud corresponding to the target Target_ID; The 3D point cloud and importance weight map corresponding to the target Target_ID are combined with the operation performance data to simplify the topology of the 3D mesh model and generate a simplified mesh. Adaptively control the performance index of the current three-dimensional radar dynamic visualization operation according to the facet number reduction ratio corresponding to the simplified grid and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID; Among them, the spatiotemporal forgetting mechanism is used to perform ADST-TPM dynamic threat prediction on the basic 3D radar data information to obtain the ADST-TPM dynamic threat indicator parameters corresponding to each target, including: Retrieving a radar scanning period Δt from raw radar data included in the basic three-dimensional radar data information, setting a cache window duration for target data according to the radar scanning period Δt, and caching the target data for the radar scanning period Δt; The radial acceleration, stability index and peak frequency shift of the target Target_ID are retrieved from the target tracking data contained in the basic 3D radar data information, and the Doppler dynamic feature vector is generated using the radial acceleration, stability index and peak frequency shift. ;in, A Indicates the radial acceleration of the target Target_ID; stability Indicates the stability index of target Target_ID; f peak Indicates the peak frequency shift of the target Target_ID; Retrieve the geofence list and target operation data corresponding to the target Target_ID, and obtain the spatiotemporal context feature vector corresponding to the target Target_ID based on the geofence list and target operation data ,in, W 01 and W 02 They represent the geographical threat correlation and target interaction threat corresponding to the target Target_ID respectively; Performing weighted fusion on the Doppler dynamics feature vector and the spatiotemporal context feature vector to generate a fused feature vector; Monitor the GPU utilization of the current system in real time and select a trained threat assessment model based on the GPU utilization; wherein the threat assessment model includes a full GRU network model, a lightweight temporal convolution model, and a lightweight temporal convolution model; The fused feature vector is input into the threat assessment model, and the initial threat probability corresponding to the target Target_ID is output. The initial threat probability is forgotten using the spatiotemporal forgetting mechanism to obtain the ADST-TPM dynamic threat indicator parameter corresponding to the target Target_ID.

2. The adaptive control method for three-dimensional radar dynamic visualization according to claim 1, characterized in that: Real-time collection of basic 3D radar data and operational performance information, including: During the operation of the 3D radar, basic 3D radar data information of the 3D radar is retrieved in real time; wherein the basic 3D radar data information includes raw radar data, target tracking data, geo-fence data, multi-target relationship data, user interaction enhancement data and system load data; Real-time collection of operational performance data during 3D radar dynamic visualization; wherein the operational performance data includes frame rate, GPU occupancy, ADST-TPM latency, and point cloud throughput.

3. The adaptive control method for three-dimensional radar dynamic visualization according to claim 1, characterized in that: Retrieve the geofence list and target operation data corresponding to the target Target_ID, and obtain the spatiotemporal context feature vector corresponding to the target Target_ID based on the geofence list and target operation data, including: Traverse the geofence list and get the minimum distance between the target Target_ID location and the geofence; Retrieving a priority value corresponding to a fence from the geo-fence list; The minimum distance between the target Target_ID location and the geo-fence is used in combination with the priority value corresponding to the fence in the geo-fence list to obtain the geo-threat relevance corresponding to the target Target_ID; Search for neighboring targets within 500 meters of the target Target_ID and select those with an initial threat probability greater than 0.7 as high-threat neighboring targets; The target interaction threat corresponding to the target Target_ID is obtained through the velocity vector of the target Target_ID and the velocity vector of the high-threat neighboring target.

4. The adaptive control method for three-dimensional radar dynamic visualization according to claim 1, characterized in that: Monitor the GPU utilization of the current system in real time and select a trained threat assessment model based on the GPU utilization, including: When the GPU utilization is less than 0.7, the complete GRU network model is selected as the threat assessment model; When the GPU utilization is greater than or equal to 0.7 and less than 0.83, the lightweight temporal convolution model is selected as the threat assessment model; When the GPU utilization is greater than or equal to 0.83, the rule engine model is selected as the threat assessment model; The predefined rules corresponding to the rule engine model are as follows: First check the current GPU load. When the GPU load exceeds 90%, enable the rule engine degradation mode. Define the rule engine function and create a function named rule_engine. This function receives the fused feature vector of the target Target_ID as input and outputs a floating-point value representing the threat level. The execution rules of the rule_engine function are as follows: Rule 1: If the radial acceleration of target_ID (target.a_radial) is greater than 20.0, and the distance from target_ID to the geofence (target.d_geo) is less than 5000.0, then the target is considered to be approaching the no-fly zone at high speed, indicating a high threat situation. The initial threat probability is returned to 0.

9. Rule 2: Check the peak frequency shift attribute target.Δf_peak of the target Target_ID. If the peak frequency shift attribute target.Δf_peak is greater than 50, it is determined to be a threat scenario with a drastic change in the Doppler spectrum and the initial threat probability is returned to 0.

7. Rule 3: Check the target's target interaction threat attribute target.neighbor_threat. If the value is greater than 1.5, calculate the initial threat probability using the function min(0.8, 0.3 + target.neighbor_threat * 0.2) and return the calculated result. Rule 4: The default initial threat probability is returned. If all the above rule conditions are not met, it means that the target threat level is low, and the default initial threat probability of 0.3 is returned.

5. The adaptive control method for three-dimensional radar dynamic visualization according to claim 1, characterized in that: Generate an importance weight map covering all target sets based on the 3D scene and the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID, including: Divide the 3D scene into regular voxel grids of 20m×20m×20m in size, and scan each voxel grid to capture the target Target_ID that appears in each voxel grid; Retrieve the decision weight corresponding to the target Target_ID appearing in each voxel grid and the initial weight corresponding to the voxel grid; The maximum value of the decision weight corresponding to the target Target_ID appearing in each voxel grid and the initial weight corresponding to the voxel grid is used as the calibration weight corresponding to each voxel grid; Performing anisotropic diffusion processing on the calibration weights to obtain weight values ​​after the anisotropic diffusion processing; Normalize the weight values ​​after the anisotropic diffusion process to obtain the final weight value corresponding to each voxel grid; The final weight values ​​corresponding to the voxel grid are encoded into a texture format that can be directly rendered by a GPU to generate an importance weight map for all target sets.

6. The adaptive control method for three-dimensional radar dynamic visualization according to claim 1, characterized in that: The 3D point cloud and importance weight map corresponding to the target Target_ID are combined with the operational performance data to simplify the topology of the 3D mesh model and generate a simplified mesh, including: Convert 3D point cloud coordinates into texture coordinates to generate 3D point cloud texture coordinates; Compare the 3D point cloud texture coordinates with the importance weight map to obtain the weight value corresponding to each point cloud; Compare the weight value corresponding to each point cloud with a preset first weight threshold and a second weight threshold, and determine a high-weight target, a medium-weight target, and a low-weight target based on the comparison result; wherein the first weight threshold is set to 0.6 and the second weight threshold is set to 0.3; Perform edge folding and simplification processing on high-weight targets to obtain the geometric details corresponding to the high-weight areas; For medium-weight targets, only those with weight values ​​greater than 0.45 are subjected to edge folding simplification to obtain the key features corresponding to the medium-weight region; Vertex merging is performed on low-weight targets to obtain the basic features corresponding to the low-weight areas; the geometric details corresponding to the high-weight areas, the key features corresponding to the medium-weight areas, and the basic features corresponding to the low-weight areas contained in all voxel grids are integrated to generate a simplified mesh.

7. The adaptive control method for three-dimensional radar dynamic visualization according to claim 1, characterized in that: Adaptively adjust the performance indicators of the current 3D radar dynamic visualization operation based on the facet number reduction ratio corresponding to the simplified grid and the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID, including: Determine the number of faces of the current simplified mesh based on the simplified mesh, and retrieve the number of faces of the mesh before simplification; Obtain the face number reduction ratio corresponding to the current simplified mesh according to the face number of the current simplified mesh and the face number of the mesh before simplification; Retrieve the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID appearing on the current screen; The frame rate and display bandwidth occupancy of the current 3D radar dynamic visualization operation are adaptively controlled using the facet reduction ratio corresponding to the current simplified mesh and the ADST-TPM dynamic threat index parameters corresponding to each target Target_ID.

8. The adaptive control method for three-dimensional radar dynamic visualization according to claim 7, characterized in that: The frame rate and display bandwidth occupancy of the current 3D radar dynamic visualization are adaptively controlled using the facet reduction ratio corresponding to the current simplified mesh and the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID, including: Retrieve the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID and the final weight value corresponding to the voxel grid where each target Target_ID is located; The threat degree coefficient is obtained by performing weighted averaging processing using the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID and the final weight value corresponding to the corresponding voxel grid; Retrieve the face reduction ratio corresponding to the current simplified mesh; The adjusted frame rate and display bandwidth occupancy are obtained using the facet number reduction ratio and the threat degree coefficient corresponding to the current simplified mesh.

9. An adaptive control system for three-dimensional radar dynamic visualization, characterized in that: The adaptive control system comprises: Data acquisition module, used to collect basic 3D radar data and operating performance data in real time; The dynamic threat indicator parameter acquisition module is used to perform ADST-TPM dynamic threat prediction on basic 3D radar data information using the spatiotemporal forgetting mechanism, and obtain the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID; An importance weight map acquisition module is used to generate an importance weight map covering all target sets based on the three-dimensional scene and the ADST-TPM dynamic threat indicator parameters corresponding to each target Target_ID, wherein the importance weight map is used to associate a specific target Target_ID with the three-dimensional point cloud corresponding to the target Target_ID; The simplified mesh acquisition module is used to simplify the topology of the three-dimensional mesh model by using the three-dimensional point cloud and importance weight map corresponding to the target Target_ID in combination with the operation performance data information to generate a simplified mesh; An adaptive control module is used to adaptively control the performance index of the current three-dimensional radar dynamic visualization operation according to the facet number reduction ratio corresponding to the simplified grid and the ADST-TPM dynamic threat index parameter corresponding to each target Target_ID; Among them, the spatiotemporal forgetting mechanism is used to perform ADST-TPM dynamic threat prediction on the basic 3D radar data information to obtain the ADST-TPM dynamic threat indicator parameters corresponding to each target, including: Retrieving a radar scanning period Δt from raw radar data included in the basic three-dimensional radar data information, setting a cache window duration for target data according to the radar scanning period Δt, and caching the target data for the radar scanning period Δt; The radial acceleration, stability index and peak frequency shift of the target Target_ID are retrieved from the target tracking data contained in the basic 3D radar data information, and the Doppler dynamic feature vector is generated using the radial acceleration, stability index and peak frequency shift. ;in, A Indicates the radial acceleration of the target Target_ID; stability Indicates the stability index of target Target_ID; f peak Indicates the peak frequency shift of the target Target_ID; Retrieve the geofence list and target operation data corresponding to the target Target_ID, and obtain the spatiotemporal context feature vector corresponding to the target Target_ID based on the geofence list and target operation data ,in, W 01 and W 02 They represent the geographical threat correlation and target interaction threat corresponding to the target Target_ID respectively; Performing weighted fusion on the Doppler dynamics feature vector and the spatiotemporal context feature vector to generate a fused feature vector; Monitor the GPU utilization of the current system in real time and select a trained threat assessment model based on the GPU utilization; wherein the threat assessment model includes a full GRU network model, a lightweight temporal convolution model, and a lightweight temporal convolution model; The fused feature vector is input into the threat assessment model, and the initial threat probability corresponding to the target Target_ID is output. The initial threat probability is forgotten using the spatiotemporal forgetting mechanism to obtain the ADST-TPM dynamic threat indicator parameter corresponding to the target Target_ID.

Images