Target selection and motion trend prediction method, device and terminal equipment

Abstract

The application provides a target selection and motion trend prediction method, device and terminal equipment, which comprises the following steps: acquiring a dangerous characteristic value in real time, and constructing a data set; adaptively learning a weight value corresponding to the dangerous characteristic value, acquiring an optimal weight value, and constructing a relationship model A; calculating an arithmetic mean of an optimal dangerous coefficient, comparing the size, and determining a target to be tracked; constructing a relationship model B between the dangerous characteristic value increment change of the target to be tracked and the PTZ increment change of a video monitoring device, and predicting the motion trend of the target to be tracked. Through multi-target adaptive screening of a radar-video linkage system, an adaptive learning model is constructed, which can be copied to a similar geographical environment scene, reduces the trial and error cost of subsequent construction, improves the user experience of the system, quickly and accurately locks the target to be tracked in multiple moving targets, and synchronously rotates the linkage video monitoring device to the predicted position, thereby improving the capture rate and accuracy of the video monitoring device on the target.

Description

Technical Field

[0001] This invention belongs to the field of mobile target tracking technology, and in particular relates to a target selection and motion trend prediction method, device and terminal equipment. Background Technology

[0002] Radar, a transliteration of the English word "radar," is an abbreviation of "radio detection and ranging," meaning "radio detection and ranging," that is, using radio waves to detect targets and determine their spatial location. Therefore, radar is also known as "radio positioning." Optoelectronics refers to devices that utilize optoelectronics, optics, precision mechanics, and computer technology to acquire video images; it is also called video surveillance equipment.

[0003] The radar-visual linkage system consists of radar, optoelectronic components, and information processing devices. Specifically, the radar detects targets within its control range and reports the relevant information to the information processing device. The optoelectronic component uses its pan-tilt unit to rotate to the reported target's location to perform target search, matching, and tracking. This system is widely used in perimeter protection, border control, and special target protection.

[0004] In existing radar-electro-optical linkage equipment, after the radar detects multiple targets, it generally adopts a single characteristic such as the target's reflective area or movement speed for hazard measurement, which has some effectiveness. However, depending on the control area, the geographical environment varies greatly. For example, in flat areas, flight speed is of greater concern, while in hilly areas, the size of the target is more important. Currently, there is an urgent need for a site-specific method that selects targets that best match the terrain and topography of the deployment location to meet the actual needs of different scenarios.

[0005] Furthermore, after target selection, the system performs coordinate transformation on the radar-reported position of the target and then issues an electro-optical rotation command, without effectively considering the motion of the moving target. As a result, after the electro-optical equipment rotates to its position, the target often appears at the edge of the electro-optical field of view or even beyond it, leading to the failure of electro-optical target search, matching, and tracking. This affects the system's tracking efficiency, causes numerous false alarms, and requires significant manpower and resources for verification, greatly reducing the user experience.

[0006] Chinese patent literature discloses a "Target Motion Trend Judgment Method, Device, and Terminal Equipment," with application publication number CN 113030951A. This invention patent judges motion trends based on changes across multiple frames, effectively reducing the impact of jitter or random data on target motion trend judgment and improving the accuracy of target motion trend judgment. However, this invention is applicable to predicting the motion trend of a specific tracking target and cannot specifically filter out targets to be tracked from multiple targets in different application scenarios. Summary of the Invention

[0007] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a target selection and motion trend prediction method, device and terminal equipment to solve the problems of existing mobile target tracking technology being unable to quickly filter out the target to be tracked from multiple targets in different application scenarios and having low accuracy in motion trend prediction and low tracking efficiency.

[0008] To achieve the above and other related objectives, the present invention provides a method for target selection and motion trend prediction, comprising:

[0009] Real-time acquisition of hazard characteristic values ​​of multiple moving targets within the controlled area to construct a dataset;

[0010] Adaptive learning optimizes the weight values ​​corresponding to the hazard feature values, obtains the optimal weight values ​​corresponding to the hazard feature values ​​that best match the geographical environment of the control area in the current dataset, and constructs a relationship model A between the optimal hazard coefficient, the hazard feature values, and the optimal weight values ​​corresponding to the hazard feature values.

[0011] Based on relational model A, calculate the optimal risk coefficient of each moving target within a time period, calculate the arithmetic mean of the optimal risk coefficients of each moving target within that time period, compare the magnitude of the arithmetic mean, and determine the moving target with the largest arithmetic mean as the target to be tracked.

[0012] The system collects the hazard characteristic values ​​of the target to be tracked and the PTZ motion parameters of the video surveillance equipment in real time. It constructs a relationship model B between the incremental change of the hazard characteristic value of the target to be tracked and the incremental change of the PTZ of the video surveillance equipment within a scanning cycle. In the next scanning cycle, the video surveillance equipment is PTZ-moved according to the relationship model B so that the target to be tracked appears in the center of the observable field of view, and the movement trend of the target to be tracked is predicted.

[0013] This invention utilizes a multi-target adaptive screening system with a radar-visual linkage system. Based on relationship model A and the calculation of the optimal risk coefficient, it quickly and accurately locks onto the target to be tracked among multiple moving targets. By establishing relationship model B, the video surveillance equipment is synchronously rotated to the predicted position to improve the target capture rate of the video surveillance equipment, thereby enhancing the overall accuracy of the radar-visual linkage system.

[0014] Preferably, the hazard characteristics of the target to be tracked include the target's speed, heading, and position status; the PTZ parameters of the video surveillance equipment include the video surveillance equipment's rotation speed, azimuth, and pitch status.

[0015] Preferably, the hazard characteristic values ​​of multiple moving targets within the controlled area are acquired in real time, including:

[0016] Data acquisition involves real-time collection of data such as batch number, position, distance, speed, heading, and RCS (radar cross-section) of each moving target using radar.

[0017] Data processing involves removing abnormal, duplicate, and invalid data.

[0018] Data reporting: Real-time reporting of the hazard characteristics of each moving target, such as distance, speed, heading, and RCS (radar cross-section).

[0019] As a preferred method, when the arithmetic mean of the optimal risk coefficients of multiple moving targets is the same, the variance of the optimal risk coefficient of each moving target within the time period is calculated, the magnitude of the variance is compared, and the moving target with the smallest variance is determined as the tracking target.

[0020] As a preferred method, adaptive learning optimizes the weight values ​​corresponding to dangerous features, including: deep learning network model selection, deep learning network model construction, dataset design, deep learning network model training, and deep learning network model optimization.

[0021] Adaptive learning optimizes the weight values ​​corresponding to hazardous feature values, using prior weight values ​​as the training set and as input to the selected deep learning network model. Under constraints, the model is trained on the trained data and compared with the validation set data to obtain the optimal solution for the current dataset. The weight ratios of different regions are continuously acquired as increments in the dataset, iteratively updating the model and continuously outputting solutions with better weight ratios. This application constructs a self-learning mathematical model for control areas, continuously supplementing the sample set and fine-tuning the model through supervised learning. This model can be replicated in similar geographical environments, reducing trial-and-error costs in subsequent construction while improving the user experience of the system.

[0022] Preferably, the deep learning network model selection includes designing a learning algorithm and model that can quickly and automatically acquire features from small sample datasets and quickly perform data regression, and can be used to learn the weight values ​​corresponding to the typical features of a moving target.

[0023] Preferably, the deep learning network model construction includes basic network modules, basic network branches, construction of the overall network structure (specifically involving 3 convolutional layers, 3 max pooling layers, 1 residual unit layer, 1 average pooling layer and 1 SVM regression layer), and corresponding loss functions and constraints.

[0024] Preferably, the design of the dataset includes dividing the known weight values ​​of other similar regions into training and validation sets proportionally, processing them through manual pre-labeling and removing extreme values, and allocating them in a 4:1 ratio for use as input for network model training and for validation of output.

[0025] Preferably, the training of the deep learning network model includes inputting the training set into the deep learning network model for training, obtaining a loss value through the loss function of the deep learning network model, and stopping training when the loss value reaches the expected value; then inputting the validation set into the trained deep learning network model, outputting the corresponding weight values, comparing them with the weight values ​​obtained in the original training set, and selecting the target result by multiplying them with the typical feature values ​​of danger; if the accuracy is higher than a threshold, the training of the deep learning network model is completed; further, the threshold is set to 0.95.

[0026] Preferably, the deep learning network model optimization includes continuously updating the training and validation sets while constantly collecting new datasets, repeating the above training steps, and continuously optimizing the weight ratio of typical dangerous feature values.

[0027] Preferably, a homogeneous representation is used to construct the dataset; the hazard feature values ​​include distance d, speed v, heading component h, and target reflectivity s, and the optimal weight values ​​corresponding to the hazard feature values ​​are η, respectively. opt μ opt δ opt ε opt The relational model A is:

[0028] w opt =[dvas]·[η opt μ opt δ opt ε opt ] T .

[0029] Preferably, the PTZ incremental change of the video surveillance equipment includes the pitch angle change value Δθ, the azimuth angle change value Δβ, and the optical lens zoom change value Δz; the relationship model B includes:

[0030]

[0031]

[0032]

[0033] Where Δs is the change in the distance s of the target to be tracked;

[0034] v x v y v z These represent the speed components of the target in the X, Y, and Z directions, respectively.

[0035] β is the relative azimuth angle with true north as 0°, and ΔT is the scanning period;

[0036] The zoom change value Δz of the optical lens is obtained by looking up the optical zoom relationship table based on Δs.

[0037] Δs 切向 The tangential distance of the target to be tracked relative to the observation point within one scan cycle;

[0038] The distance the target moves along the Z-axis within one scan cycle.

[0039] By predicting the movement trend of the target to be tracked through the radar-visual linkage system, and comprehensively considering the movement characteristics of the target, the system converts the information such as the orientation, pitch, and lens magnification of the video surveillance equipment based on the relational model B. The system moves "alongside" the target to be tracked, centered the target in the observable field of view of the video surveillance equipment, and greatly improves the efficiency and accuracy of the target tracking of the system.

[0040] The present invention also provides a target selection and motion trend prediction device, comprising:

[0041] The data acquisition unit is used to acquire the hazard characteristic values ​​of multiple moving targets within the controlled area in real time and construct a dataset;

[0042] The hazard feature assessment unit includes an adaptive learning optimization subunit, an acquisition subunit, and a first modeling subunit. The adaptive learning optimization subunit is used to adaptively learn and optimize the weight values ​​corresponding to the hazard feature values. The acquisition subunit is used to acquire the optimal weight values ​​corresponding to the hazard feature values ​​that best match the geographical environment of the control area in the current dataset. The first modeling subunit is used to construct a relationship model A between the optimal hazard coefficient and the hazard feature values ​​and the optimal weight values ​​corresponding to the hazard feature values.

[0043] The target selection unit includes a calculation subunit, a comparison subunit, and a determination subunit. The calculation subunit is used to calculate the optimal risk coefficient of each moving target within a time period according to the relationship model A, and to calculate the arithmetic mean of the optimal risk coefficients of each moving target within that time period. The comparison subunit is used to compare the magnitudes of the arithmetic means. The determination subunit is used to determine the moving target with the largest arithmetic mean as the target to be tracked.

[0044] The motion trend prediction unit includes an information acquisition subunit, a second modeling subunit, and a tracking subunit. The information acquisition subunit is used to collect the hazard characteristic values ​​of the target to be tracked and the PTZ parameters of the video surveillance equipment in real time. The second modeling subunit is used to construct a relationship model B between the incremental changes in the hazard characteristic values ​​of the target to be tracked and the incremental changes in the PTZ of the video surveillance equipment. The tracking subunit is used to perform PTZ movement of the video surveillance equipment in the next scanning cycle according to the relationship model B, so that the target to be tracked appears centered in the observable field of view, and predict the motion trend of the target to be tracked as the preferred one. The judgment subunit also includes an analysis module, which is used to calculate the variance of the optimal hazard coefficient of each moving target in the time period when the arithmetic mean of the optimal hazard coefficients of multiple moving targets is the same, compare the size of the variances, and determine the moving target with the smallest variance as the tracking target.

[0045] Preferably, the adaptive learning optimization subunit includes: a deep learning network model selection module, a deep learning network model construction module, a dataset design module, a deep learning network model training module, and a deep learning network model optimization module.

[0046] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0047] The present invention also provides a terminal device, comprising: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the terminal device to perform the method described above.

[0048] As described above, the complete invention title of this invention has the following beneficial effects: Through multi-target adaptive screening in the radar-visual linkage system, a self-learning mathematical model of the control area is constructed, continuously supplementing the sample set. In the supervised learning process, the model is continuously optimized, and can be replicated to scenarios with similar geographical environments, reducing the trial-and-error costs of subsequent construction while improving the user experience of the system; based on the calculation of relationship model A and the optimal risk coefficient, the target to be tracked among multiple moving targets can be quickly and accurately locked; by establishing relationship model B, the video surveillance equipment is synchronously rotated to the predicted position to improve the target capture rate of the video surveillance equipment, thereby improving the overall accuracy of the radar-visual linkage system. Attached Figure Description

[0049] Figure 1 The flowchart shown is for the target selection and movement trend prediction method.

[0050] Figure 2 Displayed as a real-time 3D spatial motion analysis diagram of the target to be tracked.

[0051] Figure 3 Displayed as a real-time planar (XY) motion analysis diagram of the target to be tracked.

[0052] Figure 4 Displayed as a real-time planar spatial (Z) motion analysis diagram of the target to be tracked.

[0053] Figure 5 This is a schematic diagram showing the tangential (azimuth) and pitch motion analysis of the target to be tracked after spatial synthesis.

[0054] Figure 6 The diagram shows a block diagram of a target selection and motion trend prediction device.

[0055] Figure 7 Displayed as Figure 6 Block diagram of medium-risk characteristic assessment unit 20.

[0056] Figure 8 Displayed as Figure 7 A block diagram of the adaptive learning optimization subunit 201.

[0057] Figure 9 Displayed as Figure 6 Block diagram of target selection unit 30.

[0058] Figure 10 Displayed as Figure 6 Block diagram of the motion trend prediction unit 40. Detailed Implementation

[0059] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0060] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0061] like Figure 1 As shown, this application embodiment provides a target selection and motion trend prediction method, including the following steps: S101: Real-time acquisition of the danger feature values ​​of multiple moving targets within the controlled area, and construction of a dataset; specifically:

[0062] The system uses radar to collect real-time hazardous characteristic values ​​of each moving target within the control range and reports the target's batch number, location, distance, speed, heading, and RCS (target cross-section).

[0063] The positional characteristics of each moving target within the monitoring range are collected in real time. The observation point is taken as the origin, and the straight-line distance between the origin and the target position is the target distance *d*. The line connecting the origin and the target position is the reference vector line, and the angle between the target heading and the vector line is α (0 ≤ α ≤ π). The heading component is defined.

[0064] The speed characteristics of moving targets within the monitoring range are collected in real time, and the collected speed is converted into a scalar speed v.

[0065] In this embodiment, the process of associating and converting the reported position, distance, speed, heading, and RCS (radar cross-section) of the moving target mainly includes:

[0066] The target distance d, speed v, heading component h, and RCS (target reflective surface area) of each moving target are acquired in real time and represented homogeneously as [dvhs].

[0067] Let the weights of the eigenvalues ​​corresponding to the target distance d, speed v, heading component h, and RCS (target radar cross-section) s of the moving target be η, μ, δ, and ε, respectively, and express them homogeneously as [η μ δ ε] T .

[0068] S102: Adaptive learning optimizes the weight values ​​corresponding to hazard feature values, obtains the optimal weight values ​​corresponding to the hazard feature values ​​that best match the geographical environment of the controlled area in the current dataset, and constructs a relationship model A between the optimal hazard coefficient, hazard feature values, and the optimal weight values ​​corresponding to the hazard feature values; specifically:

[0069] In this embodiment, the target distance d, speed v, heading component a, and RCS (target reflective surface area) s, and their corresponding weight values ​​η, μ, δ, and ε, will be adaptively learned using deep learning to adapt to changes in the geographical environment. The main process includes:

[0070] Choosing a network model, Convolutional Neural Networks (CNNs) can learn feature representations from a large number of samples, but CNN inference time is relatively long. Considering the dependence of machine learning methods on prior knowledge and the difficulty of obtaining on-site samples, we propose to use CNNs as the main body and introduce SVM regression functions to construct a learning algorithm and model for typical features of moving targets. Furthermore, the feature extraction of CNNs should be automatic through the network, which solves the problem that SVMs rely on experience to design and extract features, resulting in a complex and unsatisfactory process. Moreover, the features extracted by CNNs are invariant to translation, tilt, and rotation.

[0071] With limited samples, we use CNN as the main body and introduce SVM regression function to build a learning algorithm and model for the typical features of moving targets. By continuously collecting new samples to enrich the sample library, the system's adaptive training converges quickly, with high recognition rate and strong generalization.

[0072] The deep learning network structure consists of 3 convolutional layers, 3 max pooling layers, 1 residual unit layer, 1 average pooling layer, and 1 SVM regression layer.

[0073] The constructed model uses a larger receptive field for data processing. First, it performs convolution operations with a kernel size of 1×4 to extract the original features. Then, it further extracts features through convolution operations of 1×4 and 1×2 respectively.

[0074] After performing convolution, the features are filtered by 1×2 pooling. Before processing the residuals, the feature size is reduced to 1 dimension. Then, an average pooling layer is passed through the output layer, which reduces the features and parameters while maintaining rotation, translation, and scaling. Finally, the SVM layer performs regression based on the features extracted in the previous steps to obtain the optimal weight ratio for the current dataset.

[0075] The network model was trained using a dataset that consisted of 800 training sets and 200 validation sets, based on valid datasets collected from other nearby regions.

[0076] The experimental environment for training the network was: Windows 10 64-bit operating system, Intel i9-12900H CPU, RTX3060TI dedicated graphics card, 32GB DDR4 3200MHz memory, and PyTorch framework.

[0077] Training parameters: batchsize = 128, epoch = 150, stochastic gradient inertia = 0.9, dynamic learning rate = 0.9. Where R0 = 0.01, s = 50, floor() is floor division, i.e., integer division; the loss function is... Where w std The input is the weight value of the input training set, and the output is the [m*n] weight value of the output, where m is 1 and n is 4.

[0078] If Loss = NAN, then stop training, reinitialize the learning rate and training parameters, and repeat the training steps.

[0079] Let w be the risk coefficient corresponding to the typical characteristics of the target, where w = [dvhs]·[ημδε] T ;

[0080] set up Where w output w represents the risk coefficient corresponding to the weight values ​​obtained from the current training and validation sets. 验证集 The risk coefficient corresponding to the validation set. Then output the optimal weight value η for the current dataset. opt μ opt δ opt ε opt ;

[0081] The algorithm was implemented and deployed on an industrial control computer with an Intel i7 4770 processor, 8GB of RAM, and Windows 10 operating system, and connected to the Raide-Vision linkage system. Relationship model A is as follows:

[0082] The optimal risk factor is: w opt =[dvas]·[η opt μ opt δ opt ε opt ] T ;

[0083] S103: Based on relational model A, calculate the optimal risk coefficient of each moving target within a time period, calculate the arithmetic mean of the optimal risk coefficients of each moving target within that time period, compare the magnitudes of the arithmetic means, and determine the moving target with the largest arithmetic mean as the target to be tracked; specifically:

[0084] The risk coefficient w of all targets within the control area is calculated in real time, and the arithmetic mean and variance of the risk coefficients of each target over a certain period of time are calculated, i.e.:

[0085]

[0086]

[0087] Select all targets The maximum value is the target with the highest danger level. If there are multiple targets with the same risk level, select all of them. The target with the smallest variance among similar targets is designated as the highest-risk target to be tracked.

[0088] S104: Real-time acquisition of the hazard characteristic values ​​of the target to be tracked and the PTZ motion parameters of the video surveillance equipment; construction of a relationship model B between the incremental change of the hazard characteristic value of the target to be tracked and the incremental change of the PTZ of the video surveillance equipment within a scanning cycle; in the next scanning cycle, according to the relationship model B, the video surveillance equipment is PTZ-moved to ensure that the target to be tracked appears centered within the observable field of view, and the movement trend of the target to be tracked is predicted. Specifically:

[0089] For the selected target, the radar reports the velocity components v in each direction at the selected time. x v y v z With true north as 0°, the relative azimuth angle β, and the radar scanning period ΔT;

[0090] See Figure 2 By combining the speed components in the X and Y directions, the tangential velocity of the target relative to the observation point can be obtained.

[0091] See Figure 3 The tangential movement distance within the time interval ΔT during which the radar next detects the target.

[0092]

[0093] If the reported target's location and distance 's' is ', then the azimuth change value' is... See Figure 4 The distance Δs that the radar moves along the Z-axis during the next time interval ΔT when the target is detected by the radar. 俯仰 =v 俯仰 ·ΔT= z v·Δ;

[0094] If the reported target position and distance 's' is used, then the pitch angle change value is...

[0095] Reference Figure 5 The distance the target moves within the time interval ΔT between the next radar detection of the target is...

[0096] The optical side ratio of photoelectric transmission within the range of distance Δs is found in a table as Δz;

[0097] Based on the changes in azimuth angle Δβ, pitch angle Δθ, and optical lens zoom Δz, the video surveillance equipment performs PTZ motion, moving to the predicted position within the ΔT time interval, while simultaneously searching for the target.

[0098] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding embodiments.

[0099] Computer-readable storage media:

[0100] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0101] This invention also provides a terminal device, including: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal device performs any of the methods described above.

[0102] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0103] Corresponding to the aforementioned embodiments of the target selection and motion trend prediction method, this application also provides an embodiment of a target selection and motion trend prediction device.

[0104] like Figure 6 As shown, this application provides a target selection and motion trend prediction device, comprising:

[0105] Data acquisition unit 10 is used to acquire the hazard characteristic values ​​of multiple moving targets within the controlled area in real time and construct a dataset;

[0106] Hazard assessment unit 20, refer to Figure 7The system includes an adaptive learning optimization subunit 201, an acquisition subunit 202, and a first modeling subunit 203. The adaptive learning optimization subunit is used to adaptively learn and optimize the weight values ​​corresponding to the hazard feature values. The acquisition subunit is used to acquire the optimal weight values ​​corresponding to the hazard feature values ​​that best match the geographical environment of the controlled area in the current dataset. The first modeling subunit is used to construct a relationship model A between the optimal hazard coefficient, the hazard feature values, and the optimal weight values ​​corresponding to the hazard feature values. (Refer to...) Figure 8 The adaptive learning optimization subunit 201 includes: a deep learning network model selection module 2011, a deep learning network model construction module 2012, a dataset design module 2013, a deep learning network model training module 2014, and a deep learning network model optimization module 2015.

[0107] Target selection unit 30, refer to Figure 9 The system includes a calculation subunit 301, a comparison subunit 302, and a judgment subunit 303. The calculation subunit calculates the optimal risk coefficient of each moving target within a time period based on a relational model A, and calculates the arithmetic mean of the optimal risk coefficients of each moving target within that time period. The comparison subunit compares the magnitudes of the arithmetic means. The judgment subunit determines the moving target with the largest arithmetic mean as the target to be tracked. The judgment subunit 303 also includes an analysis module 3031, which calculates the variance of the optimal risk coefficients of each moving target within that time period when the arithmetic mean of the optimal risk coefficients of multiple moving targets is the same, compares the magnitudes of the variances, and determines the moving target with the smallest variance as the tracking target.

[0108] Movement Trend Prediction Unit 40, see Figure 10 The system includes an information acquisition subunit 401, a second modeling subunit 402, and a tracking subunit 403. The information acquisition subunit is used to collect the hazard characteristic values ​​of the target to be tracked and the PTZ parameters of the video surveillance equipment in real time. The second modeling subunit is used to construct a relationship model B between the incremental changes of the hazard characteristic values ​​of the target to be tracked and the incremental changes of the PTZ of the video surveillance equipment. The tracking subunit is used to perform PTZ motion on the video surveillance equipment in the next scanning cycle according to the relationship model B, so that the target to be tracked appears centered in the observable field of view, and to predict the movement trend of the target to be tracked.

[0109] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for target selection and motion trend prediction, characterized in that, include: Real-time acquisition of hazard characteristic values ​​of multiple moving targets within the controlled area to construct a dataset; Adaptive learning optimizes the weight values ​​corresponding to the hazard feature values, obtains the optimal weight values ​​corresponding to the hazard feature values ​​that best match the geographical environment of the control area in the current dataset, and constructs a relationship model A between the optimal hazard coefficient, the hazard feature values, and the optimal weight values ​​corresponding to the hazard feature values. Based on relational model A, calculate the optimal risk coefficient of each moving target within a time period, calculate the arithmetic mean of the optimal risk coefficients of each moving target within that time period, compare the magnitude of the arithmetic mean, and determine the moving target with the largest arithmetic mean as the target to be tracked. The system collects the hazard characteristic values ​​of the target to be tracked and the PTZ parameters of the video surveillance equipment in real time, and constructs a relationship model B between the incremental change of the hazard characteristic value of the target to be tracked and the incremental change of the PTZ of the video surveillance equipment within a scanning cycle. In the next scanning cycle, the video surveillance equipment is moved in PTZ according to the relationship model B so that the target to be tracked appears in the center of the observable field of view, and the movement trend of the target to be tracked is predicted. The adaptive learning method for optimizing the weight values ​​corresponding to dangerous features includes: deep learning network model selection, deep learning network model construction, dataset design, deep learning network model training, and deep learning network model optimization; using prior weight values ​​as the training set and as input to the selected deep learning network model; training the model under constraints to obtain trained data, and comparing it with validation set data under comparison conditions to obtain the optimal solution for the current dataset; continuously obtaining the weight ratio of regions as the increment of the dataset, iteratively updating the model, and continuously outputting solutions with better weight ratios; The dataset is constructed using homogeneous representation; the danger feature values ​​include distance. speed , heading component and target reflective area The optimal weight values ​​corresponding to the dangerous feature values ​​are respectively , , , The relational model A is: ； The PTZ increment change of the video surveillance equipment includes the pitch angle change value. azimuth change value And the zoom change value of optical lenses The relational model B includes: in, This represents the change in distance *s* between the target and its location. , , These represent the speed components of the target in the X, Y, and Z directions, respectively. The relative azimuth is defined with true north as 0°. The scan cycle; The zoom change value of the optical lens according to The optical margin ratio is obtained by looking up a table. The tangential distance of the target to be tracked relative to the observation point within one scan cycle; For the target to be tracked within a scan cycle Axis movement distance.

2. The method according to claim 1, characterized in that, When the arithmetic mean of the optimal risk coefficients of multiple moving targets is the same, calculate the variance of the optimal risk coefficient of each moving target within that time period, compare the magnitude of the variances, and determine the moving target with the smallest variance as the tracking target.

3. A target selection and motion trend prediction device, used to perform the method as described in claim 1, characterized in that, include: The data acquisition unit is used to acquire the hazard characteristic values ​​of multiple moving targets within the controlled area in real time and construct a dataset; The hazard feature assessment unit includes an adaptive learning optimization subunit, an acquisition subunit, and a first modeling subunit. The adaptive learning optimization subunit is used to adaptively learn and optimize the weight values ​​corresponding to the hazard feature values. The acquisition subunit is used to acquire the optimal weight values ​​corresponding to the hazard feature values ​​that best match the geographical environment of the control area in the current dataset. The first modeling subunit is used to construct a relationship model A between the optimal hazard coefficient and the hazard feature values ​​and the optimal weight values ​​corresponding to the hazard feature values. The target selection unit includes a calculation subunit, a comparison subunit, and a decision subunit; the calculation subunit is used to calculate the optimal risk coefficient of each moving target within a time period according to the relation model A, and to calculate the arithmetic mean of the optimal risk coefficients of each moving target within the time period; the comparison subunit is used to compare the magnitude of the arithmetic mean. The determination subunit is used to determine the moving target with the largest arithmetic mean as the target to be tracked. The motion trend prediction unit includes an information acquisition subunit, a second modeling subunit, and a tracking subunit. The information acquisition subunit is used to collect the hazard characteristic values ​​of the target to be tracked and the PTZ parameters of the video surveillance equipment in real time. The second modeling subunit is used to construct a relationship model B between the incremental changes of the hazard characteristic values ​​of the target to be tracked and the incremental changes of the PTZ of the video surveillance equipment. The tracking subunit is used to perform PTZ motion on the video surveillance equipment in the next scanning cycle according to the relationship model B, so that the target to be tracked appears centered in the observable field of view, and predict the motion trend of the target to be tracked.

4. The apparatus according to claim 3, characterized in that, The determination subunit also includes an analysis module, which is used to calculate the variance of the optimal risk coefficient of each moving target within the time period when the arithmetic mean of the optimal risk coefficients of multiple moving targets is the same, compare the magnitude of the variances, and determine the moving target with the smallest variance as the tracking target.

5. The apparatus according to claim 3, characterized in that, The adaptive learning optimization subunit includes: a deep learning network model selection module, a deep learning network model construction module, a dataset design module, a deep learning network model training module, and a deep learning network model optimization module.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 2.

7. A terminal device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the terminal device to perform the method as described in any one of claims 1 to 2.

Images