Intelligent Management Method for Training Target Machines Based on Deep Learning

Abstract

This invention provides a deep learning-based intelligent management method for training target drones, belonging to the field of intelligent management technology for training target drones. The method includes: Step 1, constructing and training a deep learning neural network model for processing multi-source data from the training range; Step 2, using the deep learning neural network model to extract and fuse features from real-time video images and sensor data acquired from the training range, converting them into a set of spatial coordinates; determining feature points based on the coordinate set, and establishing a principal direction vector and a secondary direction vector based on the feature points, calculating the azimuth relationship between the two direction vectors, and constructing a spatial feature region based on the azimuth relationship. This invention, through model construction and training, data processing and recognition, parameter calibration and optimization, and closed-loop feedback, achieves adaptive adjustment of target drone operating parameters and improves the management process, enhancing the accuracy of target drone motion control and shooting event recognition, and ensuring the orderly operation of the training range.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for training target machines, and in particular to an intelligent management method for training target machines based on deep learning. Background Technology

[0002] When encountering strong crosswinds in outdoor training fields, the shortcomings of some traditional target drone management solutions may become apparent. One common issue is that traditional models often struggle to efficiently compensate for visual recognition errors caused by unexpected deviations in the physical attitude of the target drone.

[0003] Specifically, when training is conducted in outdoor training grounds such as beaches, sudden strong sea winds may cause the target drone's support rod to sway continuously or form a slight tilt angle with its base. This change in posture may cause some deformation or displacement in the target images captured by the cameras deployed on site. However, traditional deep neural network models may not have sufficient learning and coverage of such nonlinear feature changes caused by mechanical deformation during the early training process. This may lead to a need to further improve the model's adaptability to such scenarios. The potential impact of this situation is that when the model processes images and extracts feature points such as the bullseye, it may still rely heavily on the conventional prior assumption that the target drone is in an absolutely vertical state. This could lead to a potential systematic bias in the calculation of the spatial coordinates of the impact point or the spot.

[0004] For example, the system may sometimes misjudge a shot that actually hits the bullseye as a miss, or send a stand-up command based on conventional spatial coordinates to a target drone that has tilted. This may not only fail to correct the abnormal posture of the target drone, but may even exacerbate the deviation to some extent in certain situations, which may ultimately lead to a decrease in the accuracy of training data recording or cause some interference with the normal training process. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a deep learning-based intelligent management method for training target machines, so as to continuously improve the performance of the intelligent management system for training target machines and enhance training effectiveness and efficiency.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a deep learning-based intelligent management method for training target machines, the method comprising:

[0008] Step 1: Build and train a deep learning neural network model to process multi-source data from the training target range;

[0009] Step 2: Use a deep learning neural network model to extract and fuse features from video images and sensor data collected in real time at the training range, and convert them into a set of spatial coordinates; determine feature points based on the coordinate set, and establish a main direction vector and a secondary direction vector based on the feature points, calculate the orientation relationship between the two direction vectors, and construct a spatial feature region based on the orientation relationship.

[0010] Step 3: Set five spatial correction points within the spatial feature region; form a correction grid based on the spatial distribution of the spatial correction points; and obtain attitude calibration parameters based on the spatial transformation characteristics of the correction grid.

[0011] Step 4: Use attitude calibration parameters to perform coordinate transformation correction on the spatial coordinate set to obtain the correction result; Based on the correction result, identify the shooting event, bullet impact point or spot position to obtain the identification result;

[0012] Step 5: Determine the target drone control command based on the recognition result, adjust the target drone's tilting, swinging, concealment or movement mode, monitor the execution effect of the control command, and obtain feedback data;

[0013] Step 6: Based on the feedback data, optimize the operating parameters using a deep learning neural network model to obtain the optimized operating parameters; feed the optimized operating parameters back into the deep learning neural network model to achieve closed-loop adaptive optimization of the intelligent management process of the training target machine.

[0014] In a second aspect, a computing device includes:

[0015] One or more processors;

[0016] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0017] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0018] The above-described solution of the present invention has at least the following beneficial effects:

[0019] By constructing and training a deep learning neural network model, multi-source data from the training range can be effectively processed. It is no longer limited to a single data type or simple data processing methods, but can extract and fuse features from real-time acquired video images and sensor data, converting them into a set of spatial coordinates, thus improving the comprehensiveness and accuracy of data processing. Based on feature points, principal and secondary direction vectors are established, and azimuth relationships are calculated. This makes the understanding of the target drone's spatial position and attitude more precise. Compared to vague or inaccurate judgments of target drone position and attitude, this method can more accurately determine the target drone's state in space, avoiding errors caused by misjudgments. This addresses a series of problems caused by errors; by setting spatial correction points to form a correction grid and obtaining attitude calibration parameters, the spatial coordinate set can be corrected, improving the accuracy of identifying shooting events, bullet impact points, or light spot positions, and reducing misjudgments and omissions; the target drone's tilting, swaying, concealment, or movement modes are adjusted according to the recognition results, and problems in the control process are promptly identified and corrected by monitoring feedback data, making the target drone's operation more in line with training needs; the operating parameters of the deep learning neural network model are optimized based on feedback data, continuously improving the performance of the entire training target drone intelligent management system and enhancing training effectiveness and efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the intelligent management method for training target machines based on deep learning provided in an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating step 1 provided in an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, embodiments of the present invention propose an intelligent management method for training target machines based on deep learning, the method comprising the following steps:

[0024] Step 1: Build and train a deep learning neural network model to process multi-source data from the training target range;

[0025] Step 2: Use a deep learning neural network model to extract and fuse features from video images and sensor data collected in real time at the training range, and convert them into a set of spatial coordinates; determine feature points based on the coordinate set, and establish a main direction vector and a secondary direction vector based on the feature points, calculate the orientation relationship between the two direction vectors, and construct a spatial feature region based on the orientation relationship.

[0026] Step 3: Set five spatial correction points within the spatial feature region; form a correction grid based on the spatial distribution of the spatial correction points; and obtain attitude calibration parameters based on the spatial transformation characteristics of the correction grid.

[0027] Step 4: Use attitude calibration parameters to perform coordinate transformation correction on the spatial coordinate set to obtain the correction result; Based on the correction result, identify the shooting event, bullet impact point or spot position to obtain the identification result;

[0028] Step 5: Determine the target drone control command based on the recognition result, adjust the target drone's tilting, swinging, concealment or movement mode, monitor the execution effect of the control command, and obtain feedback data;

[0029] Step 6: Based on the feedback data, optimize the operating parameters using a deep learning neural network model to obtain the optimized operating parameters; feed the optimized operating parameters back into the deep learning neural network model to achieve closed-loop adaptive optimization of the intelligent management process of the training target machine.

[0030] In this embodiment of the invention, by constructing and training a deep learning neural network model, multi-source data from the training range can be effectively processed. It is no longer limited to a single data type or simple data processing method, but can perform feature extraction and fusion processing on real-time acquired video images and sensor data, converting them into a set of spatial coordinates, thus improving the comprehensiveness and accuracy of data processing. Based on feature points, principal and secondary direction vectors are established, and azimuth relationships are calculated. This makes the understanding of the target drone's spatial position and attitude more precise. Compared to the somewhat vague or inaccurate judgment of the target drone's position and attitude, this method can more accurately determine the target drone's state in space, avoiding... This system avoids a series of problems caused by misjudgment; by setting spatial correction points to form a correction grid and obtaining attitude calibration parameters, it can correct the spatial coordinate set, improving the accuracy of identifying shooting events, bullet impact points, or light spot positions, and reducing misjudgments and omissions; it adjusts the target drone's tilting, swaying, concealment, or movement modes based on the recognition results, and by monitoring feedback data, it can promptly identify and correct problems in the control process, making the target drone's operation more in line with training needs; and it optimizes the operating parameters of the deep learning neural network model based on feedback data, continuously improving the performance of the entire training target drone intelligent management system and enhancing training effectiveness and efficiency.

[0031] like Figure 2As shown, in another preferred embodiment of the present invention, step 1 includes:

[0032] In this embodiment of the invention, step 100 involves acquiring historical multi-source data from the training process at the training range. This historical multi-source data includes historical video image data and historical multimodal sensor data. Preprocessing the historical multi-source data yields processed historical multi-source data, specifically including: extracting historical training data from a storage server. The historical video image data is captured by fixed cameras installed at different locations within the range, including the static state of the target drone under various lighting conditions such as bright sunlight, low light on cloudy days, nighttime without supplemental lighting, and rain or snow; images of the bullet impact moment; and laser aiming trajectories. The historical multimodal sensor data includes tilt angle data and vibration frequency data recorded in real-time by attitude sensors installed on the target drone, and firing sound intensity data captured by sound sensors deployed around the range. The laser sensor integrated into the target surface acquires data such as the coordinates of the light spot center. When preprocessing these historical multi-source data, the historical video image data is processed first. That is, the image sharpness index is calculated frame by frame. This involves calculating the sum of grayscale differences between adjacent pixels in each frame. The higher the sum of the differences, the higher the sharpness. Frames with a sharpness index below a preset threshold of 500 (caused by camera shake) and black screen frames with all grayscale values ​​of 0 (caused by equipment failure) are filtered and deleted. For the remaining valid frames, the resolution of all frames is uniformly adjusted to 1920×1080 pixels using an image scaling algorithm. Then, each frame of color image is converted to grayscale by calculating the average value of the red, green, and blue channels of each pixel. This average value is used to replace the original pixel's three-channel values, converting the color image into a single-channel grayscale image.

[0033] For historical multimodal sensor data, a reasonable data range is first determined based on the hardware parameters of each sensor. For example, the reasonable range of tilt angle for the attitude sensor is set to -15 degrees to 15 degrees based on the physical structure of the target drone support rod (with the vertical state of the target drone as 0 degrees, tilting to the left as negative, and tilting to the right as positive). Tilt angle data exceeding this range are marked as outliers and deleted. Then, using the system's unified timestamp as a reference, the data recorded by the attitude sensor, sound sensor, and laser sensor at the same timestamp are associated to achieve time synchronization of data from different sensors, ensuring that video image data and sensor data at the same moment can correspond one-to-one. Finally, all preprocessed video image data are converted to JPEG format, and sensor data are converted to CSV format to unify the data storage format, resulting in processed historical multi-source data.

[0034] Step 101: Label the processed historical multi-source data. The labeling information includes spatial three-dimensional coordinate information and shooting event category labels to obtain labeled historical multi-source data. Specifically, this includes: labeling the processed historical multi-source data using a data labeling tool (such as LabelImg); for spatial three-dimensional coordinate information, first, based on the three-dimensional coordinate system pre-set by the range (with the center point of the range ground as the origin, the horizontal direction eastward as the positive X-axis, the horizontal direction northward as the positive Y-axis, and the vertical direction upward as the positive Z-axis), locate the four targets of the target drone in the historical video images. Five key location points were identified: the corners (top left, top right, bottom left, bottom right) and the bullseye. Using an image ranging algorithm, combined with the actual dimensions of the target drone (e.g., the target height of 1.5 meters and its proportion in the image), the distances of each key location point relative to the origin of the coordinate system along the X, Y, and Z axes were calculated to determine the three-dimensional coordinates. For example, the bullseye was 5 meters from the origin along the X-axis, 3 meters along the Y-axis, and 1.2 meters along the Z-axis, resulting in three-dimensional coordinates of (5, 3, 1.2). Simultaneously, parameters related to these key location points were retrieved from sensor data. For example, the straight-line distance from the center of the laser spot detected by the laser sensor to the target center (calculating the Euclidean distance between the laser spot coordinates and the target center coordinates). If this distance exceeds a preset error threshold of 0.01 meters, the parameter calibration value of the laser sensor is adjusted based on the three-dimensional coordinates calculated by the coordinate system to ensure the accuracy of the three-dimensional coordinates of key positions. For shooting event category labels, annotations are made based on bullet hit results and data characteristics in the training records. For example, when the Euclidean distance between the bullet impact point coordinates and the target center coordinates in the video image is less than 0.05 meters, and the sound signature intensity captured by the sound sensor is high... At 80 dB, the data segment is labeled as hitting the bullseye; when the coordinates of the bullet impact point in the video image are outside the target area and the sound signature intensity is greater than 80 dB, it is labeled as missing the target; when the sound sensor continuously captures three or more sound signatures greater than 80 dB within one second and three or more corresponding bullet impact points appear in the video image, it is labeled as continuous firing; during the labeling process, the video image features and sensor data features need to be checked simultaneously to ensure that each data segment corresponds to only one shooting event category label, and that the label is consistent with the actual training situation, so as to obtain the labeled historical multi-source data.

[0035] Step 102: Based on the environmental characteristics of the training range and the features of historical multi-source data, construct a deep learning neural network model that integrates convolutional neural networks and recurrent neural networks. The convolutional neural network is used to extract the spatial features of video image data, and the recurrent neural network is used to process the time series features of multimodal sensor data. Specifically, it includes: combining the environmental characteristics of the training range, such as the fluctuation of light intensity and the slight swaying of the target machine due to wind, as well as the spatial features of historical video image data such as pixel distribution and edge contours, and the time-varying sequence features of multimodal sensor data, to design a fusion structure. The convolutional neural network part has 3 convolutional layers and 2 pooling layers.The first convolutional layer uses 32 3×3 pixel convolutional kernels with weights ranging from -0.05 to 0.05 and a bias of 0.01. The second convolutional layer uses 64 3×3 pixel convolutional kernels with weights ranging from -0.03 to 0.03 and a bias of 0.01. The third convolutional layer uses 128 5×5 pixel convolutional kernels with weights ranging from -0.02 to 0.02 and a bias of 0.01. The preprocessed 1920×1080 grayscale image pixel matrix is ​​input into the first convolutional layer. The convolutional kernels slide with a stride of 2 pixels. Each time the kernel slides, the product of the pixel value within the covered area and the corresponding kernel weight is calculated and accumulated. The bias value is then added to obtain the feature value at that location, which is used to extract the image edge features. The output of the first convolutional layer is connected to the first pooling layer. The pooling layer has a 2×2 pixel window, which slides with a stride of 2 pixels. Each time it slides, it takes the maximum value of the four feature values ​​in the window. For example, if the feature values ​​in the window are [15, 8, 22, 10], the maximum value is 22. The maximum value is used as the output to achieve feature compression. The second and third convolutional layers and the second pooling layer repeat the above operation to further extract deep spatial features such as image texture and target area. The recurrent neural network part adopts a long short-term memory structure with two hidden layers, each containing 64 memory units. The input gate weights range from [-0.04, 0.04].

[04] The weight range of the forget gate is [-0.04, 0.04], the weight range of the output gate is [-0.04, 0.04], and the bias value is 0.01 for both. The multimodal sensor data after time synchronization is input sequentially in 10-millisecond time steps. In the memory unit, the input gate calculates the weighted product of the current sensor data and the cell state at the previous moment to determine whether to update the cell state; the forget gate calculates the weighted product of the cell state at the previous moment to determine whether to retain the information; the output gate calculates the result of the activation function processing of the current cell state as the output, thereby retaining important information such as changes in shooting soundprints and fluctuations in the target tilt angle, forgetting random noise, and capturing different The time step data dependency is analyzed, such as the change in the target drone's tilt angle within 0.1 seconds after the soundprint of the shot appears. The 128-dimensional spatial feature vector output by the convolutional neural network and the 64-dimensional time series feature vector output by the recurrent neural network are concatenated in dimensional order to form a 192-dimensional fused feature vector. The fused feature vector is input into a fully connected layer with 128 neurons (weights range from -0.01 to 0.01, and the bias is 0.01). The input value of each neuron is calculated (the sum of the products of the fused feature vector elements and their corresponding weights plus the bias value). After processing by the activation function, a final feature representation that comprehensively reflects the spatial features of the video image and the time series features of the sensor is formed.

[0036] Step 103: Train the deep learning neural network model structure using labeled historical multi-source data. Adjust the deep learning neural network model structure parameters using the stochastic gradient descent optimization algorithm to obtain the trained deep learning neural network model. Specifically, the labeled historical multi-source data is divided into training, validation, and test sets in a 7:2:1 ratio. For example, when the total data volume is 10,000 records, the training set contains 7,000 records, the validation set contains 2,000 records, and the test set contains 1,000 records. The training set is used to update network weights and biases, while the validation set is used to adjust hyperparameters such as learning rate and batch size. The test set is used to evaluate the network's generalization performance. The training set data is input into a pre-constructed deep learning neural network structure in batches of 32. Video image data is input into a convolutional neural network (structure detailed in step 102), which extracts spatial features such as edges, textures, and target regions sequentially through three convolutional layers (configured with 32 3×3, 64 3×3, and 128 5×5 convolutional kernels). After compression by two pooling layers, a 128-dimensional spatial feature vector is output. Multimodal sensor data is input into a long short-term memory structure (structure detailed in step 102) according to time steps, and then processed through two hidden layers (each with 64 kernels). The memory unit processes time-series features and outputs a 64-dimensional time-series feature vector. The fusion layer concatenates the two types of feature vectors into a 192-dimensional fusion feature vector, which is input to a fully connected layer with 128 neurons (detailed in step 102). After processing by the activation function, the output layer outputs two branches. The regression branch contains three neurons (corresponding to the X, Y, and Z axes, respectively). Each neuron receives a 192-dimensional fusion feature vector, calculated by summing the products of each element of the feature vector and its corresponding weight, plus a bias value. The weight values ​​range from -0.01 to 0.01, and the bias value is 0.01. The output value is directly used as the prediction result of the three-dimensional coordinates of the bullseye; the classification branch contains 3 neurons (corresponding to the three types of shooting events respectively). Each neuron receives a 192-dimensional fused feature vector. The calculation method is to sum the products of each element of the feature vector and the weight of the neuron, with a value range of [-0.01, 0.01], plus a bias value of 0.01 to obtain a linear output. Then, it is processed by the softmax function. The linear output value of each neuron is used as the exponent of the exponential function. The ratio of the exponent value to the sum of the three exponent values ​​is calculated to obtain the probability distribution of each category, such as 0.9 for hitting the bullseye and 0 for missing the target.1. For continuous fire (0 shots): For 3D coordinates, calculate the difference between the predicted value and the labeled value along the X, Y, and Z axes, square them, and sum them to obtain the coordinate prediction loss; For the shooting event category, convert the labeled tags into one-hot encodings, such as hitting the bullseye corresponding to [1, 0, 0]. Use the cross-entropy function to calculate the difference between the predicted probability distribution and the one-hot encoding to obtain the category prediction loss; Add the two losses with a 1:1 weight to obtain the total loss value; When calculating the partial derivative of the total loss value with respect to each weight and bias, use the backpropagation algorithm starting from the output layer and calculating layer by layer according to the chain rule. Taking a certain weight w in the fully connected layer as an example, its partial derivative = the partial derivative of the loss with respect to the output of that neuron × the output value of the corresponding neuron in the previous layer; Multiply the partial derivative by the initial learning rate 0.001 to obtain the adjustment amount, and subtract the adjustment amount from the current parameter value to obtain the new value, such as when the current weight value is 0.5 and the partial derivative is 0.2. The adjustment amount is 0.2 × 0.001 = 0.0002, and the new value is 0.5 - 0.0002 = 0.4998. This process is repeated until all batches of the training set have completed one parameter update, which completes one training epoch. Every 5 epochs during training, the loss value is calculated using the validation set. If the validation set loss value does not decrease for three consecutive epochs (e.g., 0.05 in epoch 10, 0.052 in epoch 11, and 0.051 in epoch 12), training is stopped to prevent overfitting. The test set is then input into the trained network to calculate the matching degree. Specifically, the average error of the 3D coordinate prediction (the root mean square of the loss of all test data coordinate predictions) must be less than 0.02 meters, and the accuracy of the shooting event category prediction (the percentage of test data whose predictions match the labels) must be greater than 95%. This ensures the network's performance in processing untouched data, ultimately yielding the trained deep learning neural network model.

[0037] This embodiment, by acquiring and preprocessing various types of historical data, can eliminate the influence of invalid information and interference factors, making the data more in line with the model's processing needs. Detailed annotation of the processed data clarifies the spatial location and event type corresponding to the data, giving the model a clear learning objective during training and enabling it to accurately understand the actual meaning represented by different data, reducing learning bias caused by ambiguous data meanings. Constructing a structure that integrates two networks fully utilizes the spatial features of video images and the temporal features of sensor data, comprehensively capturing various types of information during training, avoiding the limitations of a single network in processing complex data, and improving the understanding of target drone status and shooting events. Using appropriate training methods and optimization algorithms to adjust model parameters allows the model to learn from historical data and grasp its patterns, improving the prediction accuracy for various situations in actual training scenarios and reducing recognition errors caused by unreasonable model parameters. Especially when the target drone experiences attitude shifts, it can better adapt and make accurate judgments.

[0038] In a preferred embodiment of the present invention, step 2 includes:

[0039] In this embodiment of the invention, step 200 involves inputting the real-time acquired video image data and multimodal sensor data into the trained deep learning neural network model for fusion processing to obtain a fused multidimensional feature representation. Specifically, this includes: acquiring video image data by using a fixed camera deployed at the firing range to capture real-time footage of the target drone and related shooting scenes. This data includes the target drone's appearance, real-time imaging of the bullet impact point, and laser aiming trajectory in the current environment; simultaneously, acquiring multimodal sensor data in real-time through attitude sensors installed on the target drone, sound sensors around the firing range, and laser sensors integrated into the target surface. The attitude sensors record the current tilt angle and vibration frequency of the target drone, the sound sensors capture the intensity of the acoustic signature during firing, and the laser sensors acquire the real-time position information of the laser spot center; processing the real-time acquired video image data and multimodal sensor data according to the preprocessing standards of the training phase, including grayscale conversion and unified segmentation of the video image data. The data is input after resolution (1920×1080 pixels). Multimodal sensor data is input after outlier removal and time synchronization (data is bound based on the system's millisecond-level timestamp). The trained deep learning neural network structure receives two types of data. The convolutional neural network part extracts features from the video image data layer by layer. It extracts spatial features such as edges, textures, and target regions through three convolutional layers (32 3×3, 64 3×3, and 128 5×5 convolutional kernels). After compression by the pooling layer, it outputs a 128-dimensional spatial feature vector. The long short-term memory structure part processes the multimodal sensor data with a time step of 10 milliseconds. It captures time series features through two hidden layers (64 memory units each) and outputs a 64-dimensional time series feature vector. The fusion layer concatenates the 128-dimensional spatial feature vector and the 64-dimensional time series feature vector in dimensional order to form a 192-dimensional fused feature vector. This vector is the fused multidimensional feature representation.

[0040] Step 201: Convert the multidimensional feature representation into a spatial coordinate set containing three-dimensional coordinate information. Perform density clustering analysis on the spatial coordinate set to identify several locations with the highest spatial distribution density as spatial distribution feature points. Based on the geometric center location and distribution pattern of the spatial distribution feature points, determine the principal direction vector and secondary direction vector using principal component analysis. Specifically, based on the pre-set three-dimensional coordinate system of the target range (with the center point of the target range ground as the origin, the positive X-axis direction being horizontally eastward, the positive Y-axis direction being horizontally northward, and the positive Z-axis direction being vertically upward), map each dimension of the 192-dimensional fused feature vector to a three-dimensional coordinate value in the coordinate system. Each dimensional feature corresponds to a three-dimensional coordinate (X, Y, Z value), and the three-dimensional coordinates corresponding to all dimensional features together constitute a spatial coordinate set. Density clustering analysis is performed on this spatial coordinate set, with a cluster radius of 0.05 meters and a minimum number of coordinate points of 5. For each coordinate point in the spatial coordinate set, the number of other coordinate points within a 0.05-meter radius centered on that point is counted. The coordinate points within the region with the highest number of coordinate points are selected, and the top 20 coordinate points within this region are taken as spatial distribution feature points. The geometric center position of the spatial distribution feature points is calculated by summing the X values ​​of all feature points and dividing by the total number of feature points. The center's X value is obtained by summing the Y values ​​of all feature points and dividing by the total number of feature points. The center's Z value is obtained by summing the Z values ​​of all feature points and dividing by the total number of feature points. These three values ​​together constitute the geometric center coordinates. Principal component analysis is used to determine the principal and secondary direction vectors. First, the deviations of the spatially distributed feature points relative to the geometric center are calculated: the X deviation for each feature point is the feature point's X value minus the center's X value; the Y deviation is the feature point's Y value minus the center's Y value; and the Z deviation is the feature point's Z value minus the center's Z value. Based on all deviations, a covariance matrix is ​​constructed. Specifically, if the number of spatially distributed feature points is n (here, n=20), the covariance matrix is ​​a 3×3 matrix. The row and column dimensions of this 3×3 covariance matrix correspond to the X-axis, Y-axis, and Z-axis deviation values, respectively. Specifically, the first row corresponds to the covariance of the X-axis deviation with respect to the X, Y, and Z-axis deviations; the first column corresponds to the covariance of the X-axis deviation with respect to the X, Y, and Z-axis deviations; the second row corresponds to the covariance of the Y-axis deviation with respect to the X, Y, and Z-axis deviations; the second column corresponds to the covariance of the Y-axis deviation with respect to the X, Y, and Z-axis deviations; the third row corresponds to the covariance of the Z-axis deviation with respect to the X, Y, and Z-axis deviations; and the third column corresponds to the covariance of the Z-axis deviation with respect to the X, Y, and Z-axis deviations. The deviation values ​​are arranged in an ordered manner within the matrix according to this row and column correspondence. The variance of the X-axis deviation value is represented by the sum of squares of the X-axis deviation values ​​of all feature points ÷ (n-1); The covariance between the X-axis deviation value and the Y-axis deviation value is represented by the sum of the products of the X-axis deviation values ​​and Y-axis deviation values ​​of all feature points ÷ (n-1); The covariance between the X-axis deviation value and the Z-axis deviation value is represented by the sum of the products of the X-axis deviation values ​​and the Z-axis deviation values ​​of all feature points ÷ (n-1); (The covariance matrix is ​​a symmetric matrix); The variance of the Y-axis deviation is represented by the sum of squares of the Y-axis deviations of all feature points ÷ (n-1); The covariance between the Y-axis deviation value and the Z-axis deviation value is represented by the sum of the products of the Y-axis deviation values ​​and the Z-axis deviation values ​​of all feature points ÷ (n-1); (The covariance matrix is ​​a symmetric matrix), representing the covariance between the Z-axis deviation and the X-axis deviation, calculated in the same way as... Consistent; (The covariance matrix is ​​a symmetric matrix), representing the covariance between the Z-axis deviation and the Y-axis deviation, calculated in the same way as... Consistent; The variance of the Z-axis deviation is represented by λ, calculated as the sum of squares of the Z-axis deviations of all characteristic points ÷ (n-1). When calculating the eigenvalues ​​and eigenvectors of the covariance matrix, λ represents the eigenvalues ​​(i.e., scalars satisfying the equation). First, solve the characteristic equation of the 3×3 covariance matrix. The formula for the characteristic equation is: Where C is the covariance matrix and I is the 3×3 identity matrix, the equation yields three eigenvalues ​​λ1, λ2, and λ3; then each eigenvalue is substituted into the equation. , where x is the eigenvector to be determined, and x≠0. Solve to obtain the eigenvector corresponding to each eigenvalue; take the eigenvector with the largest eigenvalue as the principal direction vector, and the eigenvector with the second largest eigenvalue as the secondary direction vector.

[0041] Step 202a: Based on the coordinate components of the principal and secondary direction vectors, calculate the dot product of the two vectors, and calculate the magnitudes of the principal and secondary direction vectors respectively; based on the dot product and the magnitudes of the two vectors, calculate the spatial angle between the two vectors using the inverse cosine function, specifically including: Let the coordinate components of the principal direction vector be (A1, A2, A3), and the coordinate components of the secondary direction vector be (B1, B2, B3). Calculate the dot product of the two vectors, that is, multiply the X component of the principal direction vector by the X component of the secondary direction vector, add the product of the Y component of the principal direction vector by the Y component of the secondary direction vector, and add the product of the Z component of the principal direction vector by the Z component of the secondary direction vector. Calculate the magnitudes of the two vectors respectively. The magnitude of the principal direction vector is... As a result, the magnitude of the secondary direction vector is The result is obtained by dividing the dot product by (the magnitude of the principal direction vector × the magnitude of the secondary direction vector). Substituting this ratio into the inverse cosine function, the result is the spatial angle between the two vectors.

[0042] Step 202b: Calculate the normal vectors of the principal direction vector and the secondary direction vector using the cross product operation. The normal vectors are perpendicular to the plane formed by the principal and secondary direction vectors. Based on the spatial angle and the direction of the normal vector, determine the azimuth relationship between the principal and secondary direction vectors. Represent this azimuth relationship as azimuth description parameters containing the angle and the direction of the normal vector. Specifically, let the coordinate components of the principal direction vector be (A1, A2, A3) and the coordinate components of the secondary direction vector be (B1, B2, B3). Calculate the cross product of the two vectors to obtain the normal vector. The X component of the normal vector is... The Y component is The Z component is The three components together constitute the coordinates (C1, C2, C3) of the normal vector. The direction of the normal vector is determined based on its coordinates: if the Z component > 0, the direction is vertically upward; if the Z component < 0, the direction is vertically downward; if the Z component = 0, the X component is then determined: if the X component > 0, the direction is horizontally eastward; if the X component < 0, the direction is horizontally westward. Combining the spatial angle calculated in step 202a with the direction of the normal vector, the two are integrated into a directional description parameter. This parameter contains both the angle between the two vectors and the specific direction information of the normal vector.

[0043] Step 203: Based on the orientation relationship and the location distribution of spatial feature points, determine the boundary range and orientation of the spatial feature region, and construct the spatial feature region. Specifically, this includes: referring to the spatial angle and normal vector direction in the orientation description parameters, and combining the location distribution of spatial feature points to determine the boundary range of the spatial feature region; taking the geometric center of the spatial feature points as the region center, and setting the region radius according to the density of the feature points; if the feature points are concentrated within a 0.1-meter range around the geometric center, then the region radius is set to 0.15 meters to ensure that all spatial feature points are included within the region; determining the orientation of the spatial feature region based on the principal direction vector and secondary direction vector in the orientation description parameters, so that the major axis direction of the region is consistent with the principal direction vector, and the minor axis direction is consistent with the secondary direction vector; the length of the major axis is the region radius × 2, and the length of the minor axis is the region radius × 1.5; according to the determined region center, boundary range (radius 0.15 meters), and orientation (major axis along the principal direction vector, minor axis along the secondary direction vector), an elliptical spatial region is divided in the three-dimensional coordinate system, which is the constructed spatial feature region.

[0044] This embodiment integrates the spatial information of video images with the temporal series information of sensors by fusing two types of real-time acquired data into the trained structure. This avoids the limitations of single data types and solves the problem of incomplete target drone status judgment caused by single data dimensions. By converting multidimensional features into a set of spatial coordinates and performing density clustering, feature points reflecting the key state of the target drone can be accurately selected. Principal component analysis is then used to determine the primary and secondary direction vectors, making the description of the target drone's spatial attitude more targeted and reducing interference from irrelevant coordinate points. By calculating the spatial angle using dot product and modulus, and combining it with cross product to determine the direction of the normal vector, the orientation relationship between the primary and secondary direction vectors can be clearly defined, providing an accurate basis for the orientation of spatial feature regions and avoiding regional division errors caused by ambiguous orientation judgments. By constructing spatial feature regions based on orientation relationships and feature point distribution, the core activity range of the target drone can be accurately delineated, improving processing efficiency and accuracy, while also adapting to environmental characteristics such as slight swaying of the target drone due to wind in the target range.

[0045] In a preferred embodiment of the present invention, step 3 includes:

[0046] In this embodiment of the invention, step 300 involves mapping and setting five spatial correction points within the spatial feature region based on its boundary range and orientation. These five spatial correction points correspond to the four corner positions and one center point of the actual target plane. According to the three-dimensional coordinates and topological distribution of the five spatial correction points in space, they are interconnected through spatial mapping to construct a spatial correction grid covering the spatial feature region. Specifically, this includes: defining the structural dimensions of the actual target plane, for example, the actual target is a rectangular plane, 1.2 meters long and 0.8 meters wide, with the four corners being the upper left, upper right, and lower right corners, respectively. The top, bottom left, and bottom right positions are centered at the intersection of the two diagonals of a rectangle. Combining the boundary range of the spatial feature region (an elliptical region with a radius of 0.15 meters centered on the geometric center) and its orientation (major axis along the principal direction vector, minor axis along the secondary direction vector), the four corners and center point of the actual target plane are mapped onto the spatial feature region, resulting in the three-dimensional coordinates of five spatial correction points. The mapping of the top left correction point is based on the geometric center of the spatial feature region, moving 0.6 meters (half the actual target length) in the opposite direction of the principal direction vector and 0.4 meters (half the actual target width) in the positive direction of the secondary direction vector. The X, Y, and Z values ​​of this point; the mapping of the upper right correction point is determined by moving 0.6 meters along the positive direction of the principal direction vector and 0.4 meters along the positive direction of the secondary direction vector; the mapping of the lower left correction point is determined by moving 0.6 meters along the negative direction of the principal direction vector and 0.4 meters along the negative direction of the secondary direction vector; the mapping of the lower right correction point is determined by moving 0.6 meters along the positive direction of the principal direction vector and 0.4 meters along the negative direction of the secondary direction vector; the center correction point directly corresponds to the geometric center coordinates of the spatial feature region; after obtaining the three-dimensional coordinates of the five spatial correction points, according to the topological distribution relationship (four edges)... The corner correction points form a rectangle, and the center correction point is located at the center of the rectangle. Spatial mapping is performed by connecting the correction points: first, connect the top left and top right correction points, the top right and bottom right correction points, the bottom right and bottom left correction points, and the bottom left and top left correction points to form the four sides of the rectangle; then connect the top left and bottom right correction points, and the top right and bottom left correction points to form the two diagonals of the rectangle; finally, connect the center correction point to the four corner correction points to form four radial lines. Through these connection methods, the five correction points are interconnected to construct a spatial correction grid covering the entire spatial feature area. The lines of this grid correspond one-to-one with the outline, diagonals, and central radial lines of the actual target plane.

[0047] Step 301: Based on the spatial correspondence between the spatial correction mesh and the actual standard target mesh, the spatial transformation matrix from the spatial correction mesh to the standard target mesh is calculated to generate attitude calibration parameters for pose correction of the spatial coordinate set. The attitude calibration parameters include rotation matrices and translation vectors. Specifically, the three-dimensional coordinate system of the actual standard target mesh is determined. The actual standard target mesh takes the center point of the actual target plane as the origin. In the target plane, the horizontal direction to the right is the positive X-axis, the vertical direction upward is the positive Y-axis, and the direction perpendicular to the target plane outward is the positive Z-axis. The coordinates of each node in the mesh are determined according to the actual target size. For example, the coordinates of the four corner nodes are (-0.6, 0.4, 0), (0.6, 0.4, 0), and (-0.4, 0), respectively. 6, -0.4, 0), (0.6, -0.4, 0), with the center node coordinates being (0, 0, 0); select four sets of feature points corresponding to the spatial correction grid and the actual standard target grid, such as the four corner correction points of the upper left, upper right, lower right, and lower left, which correspond to the four corner nodes of the actual standard target grid respectively), and obtain the three-dimensional coordinates of each set of corresponding points; let the coordinates of the feature points in the spatial correction grid be (X1, Y1, Z1), (X2, Y2, Z2), (X3, Y3, Z3), (X4, Y4, Z4), and the coordinates of the corresponding feature points in the actual standard target grid be (X1', Y1', Z1'), (X2', Y2', Z2'), (X3', Y3', Z3'), (X4', Y4', Z4');

[0048] When calculating the spatial transformation matrix, the rotation matrix is ​​calculated first. First, four sets of feature points (e.g., the top left, top right, bottom right, and bottom left corner points) corresponding to the spatial correction mesh and the actual standard target mesh are selected as reference points for calculating the rotation relationship. The geometric centers of the two sets of feature points are calculated separately. For the four sets of feature points in the spatial correction mesh, their coordinates in the X-axis direction are added together and divided by 4 to obtain the X-axis coordinates of the center. Similarly, the coordinates in the Y-axis and Z-axis directions are added together and divided by 4 to obtain the Y-axis and Z-axis coordinates of the center. The calculation method for the center of the feature points in the actual standard target mesh is completely consistent with that of the spatial correction mesh. For each set of feature points, the X-axis coordinate of the feature point itself is subtracted from the X-axis coordinate of the corresponding mesh feature point center to obtain the deviation of the feature point in the X-axis direction. The deviations in the Y-axis and Z-axis directions are calculated using the same method. The deviations in the three directions together constitute the deviation vector of the feature point relative to its own mesh center. The deviation vectors of the four sets of feature points of the actual standard target mesh are calculated in the same way. The covariance matrix is ​​a 3x3 matrix used to reflect the relationship between the two sets of deviation vectors. When calculating the element of the first row and first column of the matrix, the X-axis deviation of the four sets of deviation vectors of the spatial correction mesh is multiplied by the X-axis deviation of the corresponding deviation vector of the actual standard target mesh, and then the four products are added together and divided by 4. The element of the first row and second column is the sum of the product of the X-axis deviation of the spatial correction mesh deviation vector and the Y-axis deviation of the actual standard target mesh deviation vector, and then divided by 4. The element of the first row and third column is the sum of the product of the X-axis deviation of the spatial correction mesh deviation vector and the Z-axis deviation of the actual standard target mesh deviation vector, and then divided by 4. The calculation logic of the element of the second row of the matrix is ​​similar to that of the first row, except that the reference dimension of the spatial correction mesh deviation vector is changed to the Y-axis. The element of the third row changes the reference dimension to the Z-axis, and finally the 3×3 covariance matrix M is constructed.

[0049] Singular value decomposition of the covariance matrix M yields three matrices U, S, ... ( Let U be the transpose of V, where U and V are both 3×3 orthogonal matrices; calculate the rotation matrix R using U and V. ( Let U be the transpose of U. This ensures that the rotation of the spatial correction mesh deviation vector aligns with the direction of the actual standard target mesh deviation vector; that is, the direction of the feature points of the spatial correction mesh after rotation is consistent with the direction of the feature points of the actual standard target mesh. Subtract the coordinates of the feature points of the spatial correction mesh after rotation from the coordinates of the feature points of the actual standard target mesh to obtain the translation difference for each corresponding point. Take the average of the four translation differences to obtain the translation vector (Tx, Ty, Tz), where Tx is the average translation value in the X-axis direction, Ty is the average translation value in the Y-axis direction, and Tz is the average translation value in the Z-axis direction. The translation value is calculated as follows: Tx = (X1' - (the sum of the products of the first row elements of the rotation matrix multiplied by X1, Y1, and Z1 respectively) + X2' - (the sum of the products of the first row elements of the rotation matrix multiplied by X2, Y2, and Z2 respectively) + X3' - (the sum of the products of the first row elements of the rotation matrix multiplied by X3, Y3, and Z3 respectively) + X4' - (the sum of the products of the first row elements of the rotation matrix multiplied by X4, Y4, and Z4 respectively)) divided by 4. Ty and Tz are calculated in a similar way to Tx, only replacing the corresponding row elements of the rotation matrix. The resulting rotation matrix and translation vector together constitute the attitude calibration parameters.

[0050] This embodiment establishes a precise correlation between the virtual spatial region and the actual target by mapping five correction points corresponding to the actual target within the spatial feature region and constructing a spatial correction grid, thus avoiding correction deviations caused by misalignment between the spatial region and the actual target. By calculating the spatial transformation matrix to obtain the rotation matrix and translation vector, the pose difference between the spatial correction grid and the actual standard target grid can be accurately quantified. The generated attitude calibration parameters can specifically correct the deviation of the spatial coordinate set, solving the problem of recognition error caused by the difficulty in compensating for the physical attitude deviation of the target machine.

[0051] In a preferred embodiment of the present invention, step 4 includes:

[0052] In this embodiment of the invention, step 400 involves using attitude calibration parameters to perform spatial coordinate transformation on the spatial coordinate set to obtain corrected spatial coordinate data. Specifically, this includes: the attitude calibration parameters contain a rotation matrix (3×3 matrix) and a translation vector (Tx, Ty, Tz). The transformation process requires sequentially performing rotation and translation operations on each coordinate point in the spatial coordinate set. When performing the rotation operation, for a single coordinate point (X, Y, Z) in the spatial coordinate set, the intermediate value of the rotated X-axis is calculated, i.e., the X value of the coordinate point × the element of the first row and first column of the rotation matrix + the Y value of the coordinate point × the element of the first row and second column of the rotation matrix + the Z value of the coordinate point × the element of the first row and third column of the rotation matrix. The sum of these three products is the intermediate value of the rotated X-axis. The intermediate value of the rotated Y-axis is also calculated, i.e., the X value of the coordinate point × the element of the second row and first column of the rotation matrix + the Y value of the coordinate point × the element of the second row and second column of the rotation matrix. The intermediate value of the Y-axis after rotation is obtained by multiplying the Z-value of the coordinate point by the element of the rotation matrix, the Z-value of the coordinate point by the element of the rotation matrix, the element of the rotation matrix, the element of the rotation matrix, the element of the rotation matrix, the Z-value of the coordinate point by the element of the rotation matrix, the element of the rotation matrix, the element of the rotation matrix, the element of the rotation matrix, the Z-value of the coordinate point by the element of the rotation matrix, the Z-value of the coordinate point by the element of the rotation matrix, the Z-value of the coordinate point, the Z-value of the coordinate point, the Z-value of the coordinate point, and so on. The intermediate value of the X-axis after rotation is then multiplied by the translation vector Tx to obtain the corrected X-axis coordinate of the coordinate point. The intermediate value of the Y-axis after rotation is then multiplied by the translation vector Ty to obtain the corrected Y-axis coordinate of the coordinate point. The intermediate value of the Z-axis after rotation is then multiplied by the translation vector Tz to obtain the corrected Z-axis coordinate of the coordinate point. This process is repeated for all coordinate points in the spatial coordinate set. After all the points are processed, the corrected spatial coordinate data is formed by combining all the corrected coordinate points.

[0053] Step 401: Based on the corrected spatial coordinate data, analyze the changing characteristics of multimodal sensor data to identify the occurrence of shooting events, and detect the distribution of abnormal points in the spatial coordinate data. Determine the bullet impact coordinates or spot position based on the spatial location information of the abnormal points to obtain the identification result containing shooting event information and position information. Specifically, the multimodal sensor data includes the tilt angle and vibration frequency recorded by the target drone attitude sensor, the acoustic intensity captured by the target range sound sensor, and the real-time position data of the spot obtained by the target surface laser sensor. When analyzing the changing characteristics of sensor data to identify shooting events, first focus on the sound sensor data. The background acoustic value of the target range environment is stable between 30 and 40 decibels, such as the acoustic intensity corresponding to daily wind noise and equipment operating noise. If the acoustic intensity suddenly increases from this range, and the increased acoustic intensity reaches the preset shooting acoustic threshold of 80 decibels, and the duration of this high value state is between 50 and 200 milliseconds (which is consistent with the typical duration of a single shooting acoustic), then it is preliminarily determined that a shooting event may exist. Under normal conditions, the vibration frequency of the target drone is stable between 5 and 10 Hz. If the acoustic intensity increases... At the same moment, the target drone's vibration frequency exhibits an instantaneous peak of 100 to 200 Hz (matching the vibration characteristics caused by bullet impact or laser-simulated impact during firing); simultaneously, the target drone's tilt angle is normally within the range of -5 to 5 degrees. If there are no sudden changes in the tilt angle exceeding this range (excluding interference such as target drone tipping over or external force collisions), then the firing event is further confirmed, and the time of the firing event is recorded; after confirming the firing event, anomalies are detected in the corrected spatial coordinate data. Anomalies refer to points whose spatial positions differ significantly from most surrounding coordinate points and deviate from the specified range. Coordinate points whose deviation exceeds the preset deviation threshold of 0.03 meters are identified. During detection, each coordinate point in the corrected spatial coordinate data is traversed, and 10 adjacent coordinate points are selected for each coordinate point (selected from near to far in terms of spatial distance). The spatial distance between the coordinate point and each adjacent coordinate point is calculated (first calculate the square of the difference between the X-axis coordinate, the square of the difference between the Y-axis coordinate, and the square of the difference between the Z-axis coordinate, and then add the three squared values ​​and take the square root). If the spatial distance between the coordinate point and more than 7 adjacent coordinate points is greater than 0.03 meters, the coordinate point is marked as an anomaly.

[0054] The impact point or spot location is determined based on the spatial location of the anomaly. When determining whether it is a live-fire exercise and the impact point, the sound sensor data is checked first. If a sound signature above 80 decibels is captured, it matches the sound signature intensity characteristics of a live-fire exercise. Simultaneously, the attitude sensor data is checked. If a vibration peak of 100 to 200 Hz is recorded, it matches the vibration characteristics caused by a live-fire impact on the target, then it is determined to be a live-fire exercise. At this point, the coordinates of the anomaly are checked. If the X-axis coordinate of the anomaly is between -0.6 meters and 0.6 meters, the Y-axis coordinate is between -0.4 meters and 0.4 meters, and the Z-axis coordinate is between -0.02 meters and 0.02 meters, it falls within the preset range of the actual target plane. The three-dimensional coordinates of this anomaly are the impact point coordinates. The determination of laser simulation shooting and the spot location are then performed. When determining the location of the spot, first check the laser sensor data. If a spot signal is detected (the signal generated when the laser irradiates the target surface during laser simulation shooting), simultaneously check the attitude sensor data. If no vibration peak of 100 to 200 Hz is recorded (no live bullet impact, so there is no vibration in this frequency band), it is determined to be laser simulation shooting. At this time, calculate the spatial distance between the abnormal point and the real-time position of the spot recorded by the laser sensor. If the distance is less than 0.01 meters (the two positions are determined to match), the three-dimensional coordinates of the abnormal point are the spot position. Integrate the determined shooting event type (live bullet shooting or laser simulation shooting), the recorded shooting time, and the determined bullet impact point coordinates or spot position to form an identification result containing shooting event information and position information.

[0055] This embodiment transforms the spatial coordinate set through attitude calibration parameters, accurately correcting coordinate offsets caused by wind swaying, installation deviations, etc., ensuring precise matching between the corrected spatial coordinate data and the actual target position and attitude, thus resolving inaccurate target reporting caused by coordinate deviations. By combining the changing characteristics of multimodal sensor data to identify shooting events, it integrates multi-dimensional information such as sound signatures, vibrations, and light spots, avoiding misjudgments from single sensor data, such as mistaking ambient noise for shooting sounds based solely on sound signatures, thereby improving the accuracy of shooting event identification. Furthermore, by detecting and locating bullet impact points or light spots through anomaly detection, it can selectively identify key coordinate points related to shooting, eliminating irrelevant coordinate interference and ensuring the accuracy of position information, meeting the needs of rapid and accurate recording of shooting results in range training.

[0056] In a preferred embodiment of the present invention, step 5 includes:

[0057] In this embodiment of the invention, step 500 involves obtaining a target drone motion control command set for controlling the target drone's movement behavior based on the bullet impact point coordinates or spot position information in the identification results, combined with the timing of the shooting events. This target drone motion control command set is used to control the target drone to execute a tilting motion mode, a swinging motion mode, a concealment / revealing motion mode, or a movement trajectory mode. Specifically, this includes: first, extracting core information from the identification results; if the identification result is live-fire shooting, extracting the X-axis, Y-axis, and Z-axis coordinates of the bullet impact point; if the identification result is laser simulation shooting, extracting the X-axis, Y-axis, and Z-axis coordinates of the spot; simultaneously, extracting the occurrence time of each shooting event, and organizing the temporal relationship of consecutive shooting events in chronological order, including the time interval between two shots, the shooting... Information such as the dense shooting period; the target drone's impact area is determined based on the position of the impact point or light spot. The actual target plane is functionally divided into a central area and an edge area. The X-axis range of the central area is 0.2 meters to 0.2 meters, and the Y-axis range is 0.15 meters to 0.15 meters. The X-axis range of the edge area is 0.6 meters to 0.2 meters or 0.2 meters to 0.6 meters, and the Y-axis range is 0.4 meters to 0.15 meters or 0.15 meters to 0.4 meters. If the impact point or light spot falls in the central area, it is considered a core impact; if it falls in the edge area, it is considered an edge impact. If the impact point or light spot of three consecutive shots all fall in the same area, and the time interval between two adjacent shots is less than 2 seconds, it is considered a dense impact. The trigger condition for the knockdown action mode is that a single shot is considered a core impact. If a target is hit, or two consecutive shots are both determined to be edge hits with the time interval between the two shots being less than 1 second, a knockdown command is generated. This command includes the knockdown trigger time, knockdown speed, knockdown angle, and target repositioning delay time. The knockdown trigger time is 500 milliseconds after the initial shot; the knockdown speed is 0.5 meters per second; the knockdown angle is 90 degrees (from upright to horizontal); and the target repositioning delay time is 3 seconds (remaining horizontal for 3 seconds after knocking before returning to upright). The swing action mode is triggered when three consecutive shots are determined to be edge hits, with the three impact points or light spots distributed on the left and right sides of the target respectively. The X-axis range of the left area is 0.6 meters to 0.3 meters, and the X-axis range of the right area is 0.3 meters to 0.6 meters. The system generates a swing command, which includes the swing trigger time, swing amplitude, swing frequency, and swing duration. The swing trigger time is 300 milliseconds after the third shot. The swing amplitude is 15 degrees to the left and right, centered on the vertical axis of the target. The swing frequency is 1 time per second, and the swing duration is 5 seconds. The trigger condition for the concealment / reveal mode is that there is no subsequent shooting after a single shot for more than 5 seconds, or no new shooting after continuous shooting stops for more than 8 seconds. Once the condition is met, a concealment / reveal command is generated. The command includes the concealment / reveal trigger time, concealment speed, concealment position, and reveal delay time. The concealment / reveal trigger time is 2 seconds after the stop shooting. The concealment speed is 0.8 meters per second, and the concealment position is when the target is sunk into the target trench and its top is 0.5 meters below the ground.The target is 3 meters away, with a 10-second delay after it appears hidden. It remains in this state for 10 seconds before rising to an upright position. The trigger condition for the movement trajectory mode is a determination of dense impact and a fixed impact area. For example, if the impact points or light spots of four consecutive shots all fall within the same 10-centimeter range on the left edge, a movement command is generated upon meeting this condition. The command includes the movement trigger time, movement direction, movement speed, and movement distance. The movement trigger time is delayed by 1 second from the moment the fourth shot occurs. The movement direction is away from the impact area; if the impact area is on the left, it moves to the right. The movement speed is 0.3 meters per second, and the movement distance is 0.5 meters, maintaining the new position after movement. All the command parameters corresponding to the above-defined movement modes are integrated and ordered according to the order of their trigger times to form the target drone's motion control command set.

[0058] Step 501: During the execution of the target drone's motion control command set, the actual motion state of the target drone is monitored by real-time acquisition of multimodal sensor data, obtaining monitoring data including the target's spatial position, motion attitude, and motion response time. The monitoring data is compared and analyzed with the expected motion state parameters in the target drone's motion control command set to obtain feedback data reflecting the command execution effect. Specifically, this includes: acquiring real-time X-axis, Y-axis, and Z-axis position data of the target through a displacement sensor installed at the bottom of the target drone, with a sampling frequency of 10 times per second; acquiring real-time tilt angle and angular velocity of the target through an attitude sensor installed in the middle of the target drone, with a sampling frequency of 20 times per second; recording the command issuance time, the start time of the target drone's motion, and the completion time of the motion through a timer built into the target drone control module; calculating the monitoring data based on the acquired sensor data. The target's spatial position is calculated by taking the average of the X-axis, Y-axis, and Z-axis position data acquired by the displacement sensor every 100 milliseconds, and this average value is the actual spatial position of the target at the corresponding moment; the motion attitude is calculated by taking the attitude every 50 milliseconds. The average tilt angle collected by the sensors, combined with angular velocity data, determines the direction of motion. For example, if the tilt angle increases from 0 degrees to 90 degrees and the angular velocity is positive, it is determined to be a target inversion motion, ultimately obtaining the actual motion posture of the target. The action response time is calculated by subtracting the command issuance time from the start time of the target's action. For example, if the command to invert is issued at time T1 and the target begins to invert at time T2, then the action response time is T2 minus T1. The monitoring data is compared and analyzed with the expected parameters in the command set. During position comparison, the actual spatial position is calculated. The deviation from the expected position is calculated. For example, if the expected Z-axis position after the target is tilted is 0.5 meters, and the actual Z-axis position is 0.52 meters, the deviation is 0.02 meters. If the deviation is less than 0.05 meters, the position is considered to be in line with expectations. If the deviation is greater than 0.05 meters but less than 0.1 meters, the position is considered to be basically in line with expectations. If the deviation is greater than 0.1 meters, the position is considered to be unfavorable to expectations. During attitude comparison, the deviation between the actual tilt angle and the expected angle is calculated. For example, if the expected maximum tilt angle of the sway command is 15 degrees, and the actual maximum tilt angle is 14.5 degrees, the deviation is 0.For position, posture, and response time, the deviation is calculated as follows: if the deviation is less than 1 degree, the posture is considered to meet expectations; if the deviation is greater than 1 degree but less than 3 degrees, the posture is considered to basically meet expectations; if the deviation is greater than 3 degrees, the posture is considered to not meet expectations. For response time comparison, the deviation between the actual action response time and the expected command response time is calculated. For example, if the expected response time for a take-off / fall-off command is 500 milliseconds and the actual response time is 520 milliseconds, the deviation is 20 milliseconds. If the deviation is less than 50 milliseconds, the response time is considered to meet expectations; if the deviation is greater than 50 milliseconds but less than 100 milliseconds, the response time is considered to basically meet expectations; if the deviation is greater than 100 milliseconds, the response time is considered to not meet expectations. Based on the above comparison results, the number of items that meet expectations, basically meet expectations, and do not meet expectations in position, posture, and response time is counted. The compliance rate is calculated by dividing the number of items that meet expectations by the total number of items. Simultaneously, the specific deviation value of the items that do not meet expectations is recorded, forming feedback data reflecting the command execution effect. For example, if the position, posture, and response time of the take-off / fall-off action meet expectations, the compliance rate is 66.7%, and the response time deviation is 40 milliseconds.

[0059] This embodiment combines shooting results with timing to generate a target drone motion control command set, enabling the target drone's movements to match the shooting scenario, avoiding the mechanical nature of fixed motion patterns, and enhancing the realism and challenge of range training. By using multimodal sensors to monitor the target drone's motion status in real time and comparing it with expected parameters, it can promptly detect command execution deviations, such as insufficient target tilting angle or excessive response delay, ensuring the accuracy and stability of the target drone's movements. The generated feedback data can intuitively reflect the command execution effect, helping operators to grasp the target drone's operating status, reducing training interruptions caused by equipment failure or parameter deviations, and improving the continuity and efficiency of range training.

[0060] In a preferred embodiment of the present invention, step 6 includes:

[0061] Step 600: The feedback data is transmitted to the trained deep learning neural network model. Feature extraction and pattern recognition are performed on the feedback data to obtain recognition results. Specifically, this includes: organizing the feedback data, which covers instruction execution types such as falling, swaying, comparison results of each item (position meets expectations, basically meets expectations, does not meet expectations), compliance rate values, and specific deviation values ​​of non-compliant items, such as position deviation of 0.08 meters, posture deviation of 2 degrees, and response time deviation of 60 milliseconds. This data is converted into structured data that the neural network can read. Textual information, such as execution type, is converted into corresponding category codes (falling is coded as 1, swaying as 2, hidden / revealed as 3, and movement trajectory as 4). Numerical information, such as compliance rate and deviation value, retains the original values ​​and performs normalization, compressing the values ​​to the range of 0 to 1. The structured feedback data is then input into the trained deep learning neural network model. Feature extraction is performed first. The model input layer receives the structured data and performs operations through the convolutional units of the hidden layers. The weight values ​​of the hidden layer neurons range from [-0.02]. The weights within this range are multiplied by the corresponding input data, and all products are summed. A neuron bias of 0.01 is then added, and the result is processed by an activation function to obtain primary features. These primary features are then processed multiple times in subsequent hidden layers (where the neuron weights are also within the range of [-0.02, 0.02], with a bias of 0.01) to extract key features from the feedback data, such as positional deviation consistently greater than 0.05 meters, response time deviation concentrated between 50 and 100 milliseconds, and a fall / rise action accuracy rate below 80%. The model extracts key features and matches them with fault modes and parameter deviation modes learned during the training phase. For example, if the position deviation during the falling action is consistently greater than 0.05 meters and the falling speed parameter is normal, it matches the slight deformation mode of the target support structure; if the response time deviation of all actions is between 50 and 100 milliseconds, it matches the control signal transmission delay mode; if the attitude deviation during the swaying action is greater than 3 degrees and the angular velocity data fluctuates greatly, it matches the attitude sensor calibration offset mode. Finally, the matched mode is output as the recognition result.

[0062] Step 601: Adaptively optimize and adjust the operating parameters based on the recognition results. These operating parameters include illumination environment compensation parameters, target drone motion dynamics parameters, and target reporting accuracy calibration parameters, resulting in optimized operating parameters. Specifically, this includes: determining the type of operating parameter to be optimized based on the recognition results; if the recognition results indicate that the illumination environment causes target image acquisition deviation and affects target reporting accuracy, then focus on optimizing the illumination environment compensation parameters; if the recognition results indicate that target drone motion lag or speed fluctuations cause attitude deviations exceeding the range, then focus on optimizing the target drone motion dynamics parameters; if the recognition results indicate that the bullet impact point positioning deviation is consistently greater than 0.03 meters, leading to inaccurate target reporting data. The focus is on optimizing the target accuracy calibration parameters, and then further optimizing and adjusting different parameter types. When optimizing the lighting environment compensation parameters, these parameters include: image brightness compensation value (the equipment allows a range of 50 to 2000 cd / m²), used to adjust image brightness (the larger the value, the brighter the image); contrast compensation coefficient (range 0 to 1), with larger values ​​resulting in more obvious differences in image brightness; and white balance adjustment value, used to correct color temperature (neutral color temperature corresponds to 6500K). The values ​​of 0.3, 0.1, 0.5, and 0.15 used in the calculation are parameter adjustment coefficients, fixed coefficients that balance the adjustment effect with the stability of the equipment. If the recognition result shows that the target image is overexposed under strong light (the image appears washed out), (If the edge of the bullet impact point is blurred), increase the image brightness compensation value: for example, if the current value is 200 cd / m², the standard ambient light intensity is 5000 lux, and the actual value exceeds the standard value by 1000 lux, calculate the new brightness compensation value as current value + excess amount × 0.3, so the new brightness compensation value = 200 + 1000 × 0.3 = 500 cd / m², which is within the safe range of 50 to 2000 cd / m², thus suppressing overexposure by increasing brightness; at the same time, decrease the contrast compensation coefficient, for example, if the current value is 0.8, calculate the new coefficient as current value - 0.1, so the new coefficient = 0.8 - 0.1 = 0.7, reducing the difference between light and dark to restore the details of the bullet impact point; if the recognition result shows that the target image is too dark in low light. (The image is too dark, and the impact point is difficult to distinguish). Increase the image brightness compensation value. For example, if the current value is 100 cd / m², and the difference between the standard light intensity and the actual light intensity is 3000 lux, calculate the new brightness compensation value as current value + difference × 0.5 = 100 + 3000 × 0.5 = 1600 cd / m², which is within a safe range. This will enhance the brightness and reveal details. At the same time, increase the contrast compensation coefficient. For example, if the current value is 0.4, calculate the new coefficient as current value + 0.15 = 0.4 + 0.15 = 0.55. This will increase the difference between light and dark to highlight the impact point. Also, set the white balance adjustment value to 6500K neutral color temperature to avoid color cast affecting the identification of the impact point.

[0063] When optimizing the target drone's motion dynamics parameters, these parameters include: motor drive voltage (rated value 15V, allowable range 12 to 16.5V, controlling motor speed; higher values ​​result in faster speed); motion damping coefficient (range 0 to 1, higher values ​​result in smoother motion); and transmission mechanism clearance compensation value (allowable range 0.05 to 0.5mm, used to compensate for mechanical clearance; the value matches the actual clearance). The values ​​0.2, 0.05, and 1.2 used in the calculation are parameter adjustment coefficients to ensure smooth motion and prevent equipment damage after adjustment. If the identification result shows that the target drone's starting and falling speed is lower than expected (expected 0.5m / s, actual 0.3m / s), the motor drive voltage is increased. For example, if the current value is 12V, the new voltage is calculated as: current value + (expected speed - actual speed) × 0.2 = 12 + (0.5 - 0.3) × 0.2 = 12.04V, which is within the safe range of 12 to 16.5V. The speed is then increased to accelerate the falling speed. If the identification result shows that the target drone is swaying and jamming (angular velocity...), the motor drive voltage is increased. (To mitigate fluctuations), increase the motion damping coefficient. For example, if the current value is 0.3, calculate the new coefficient as 0.3 + 0.05 = 0.35, enhancing motion smoothness. Simultaneously, adjust the gap compensation value. For example, if the actual gap is 0.2mm, the standard gap is 0.1mm, and the current compensation value is 0.1mm, calculate the new compensation value as 0.1 + (actual gap - standard gap) × 1.2, which is within the safe range of 0.05 to 0.5mm, offsetting mechanical gaps to reduce jamming. If the identification result shows a large deviation in the target machine's movement position (expected stopping position X = 1m, actual X = 0.9m), control the motor drive voltage fluctuation range within ±0.1V. If it is stable at 15±0.1V, it is within the safe range, avoiding voltage fluctuations that cause unstable speed. Adjust the motion damping coefficient to match the movement speed, such as setting it to 0.4 when the movement speed is 0.3m / s, ensuring accurate stopping position.

[0064] When optimizing the target reporting accuracy calibration parameters, these parameters include a coordinate mapping correction coefficient (range 0.9 to 1.1, used to correct coordinate deviation; a value close to 1), a bullet impact point judgment threshold (range 1 to 3 mm, used to distinguish between the bullet impact point and interference points; a larger value results in fewer misjudgments), and sensor data fusion weights (the weights of the laser and image sensors sum to 1). The 0.4 and 0.02 used in the calculation are parameter adjustment coefficients to ensure accurate target reporting after calibration. If the identification result shows a large X-axis positioning deviation of the bullet impact point (actually 5 mm, allowable 3 mm), adjust the X-axis correction coefficient. For example, if the current value is 0.98, calculate the new coefficient as: current value + (actual deviation - allowable deviation) × 0.4, new coefficient = 0.98 + (5 - 3) × 0.4 = 1.78. This exceeds the safe range of 0.9 to 1.1, so the upper limit of 1.1 should be used to correct the X-axis coordinates to reduce the deviation. If the identification result shows a large X-axis positioning deviation of the bullet impact point... High false positive rate (misidentifying dust as impact point) To improve the accuracy of the fusion data, increase the judgment threshold. For example, if the current value is 2mm, calculate the new threshold as 2 + 0.02 = 2.02mm, which is within the safe range of 1 to 3mm. If the recognition result shows a large deviation between the laser and image sensor data fusion (laser positioning X = 2m, image positioning X = 1.95m), increase the weight of the laser sensor, such as from 0.4 to 0.6, and decrease the weight of the image sensor, such as from 0.6 to 0.4. The larger the deviation, the larger the adjustment. Organize all adjusted parameters to form optimized operating parameters, ensuring that all parameters are within the safe operating range of the equipment. Examples include image brightness compensation values ​​of 50 to 2000 cd / m², transmission mechanism gap compensation values ​​of 0.05 to 0.5mm, and motor drive voltage of 12 to 16.5V.

[0065] Step 602 involves feeding back the optimized operating parameters to the trained deep learning neural network model and the target machine motion control command generation process, achieving closed-loop adaptive optimization of the intelligent management process of the training target machine. Specifically, this includes: using the optimized lighting environment compensation parameters, target machine motion dynamics parameters, and target reporting accuracy calibration parameters as new sample data to supplement the model's training sample library; simultaneously using the changes in feedback data before and after parameter optimization, such as increasing the start-up / fall-down action compliance rate from 70% to 90% and decreasing the position deviation from 8mm to 3mm, as labels for incremental training of the model; and updating the weight values ​​of the model's hidden layers using the new sample data during training. For example, if the original weight is 0.01, the update method is to add a gradient descent adjustment of 0.002 to the weight value. The new weights are set to 0.01 + 0.002 = 0.012. This allows the model to more accurately identify parameter deviation patterns. The optimized operating parameters are then fed back into the target drone motion control command generation process. When generating new target drone motion control commands, the optimized parameters are used as constraints and references for command generation. Specifically, the voltage-speed mapping relationship of the target drone motor is pre-calibrated. For example, every 0.1V increase in the motor drive voltage corresponds to a 0.02m / s increase in the stable motion speed of the target drone. It is known that the stable speed of the motor is 0.4m / s at 12V and 0.5m / s at 12.5V. When the optimized motor drive voltage is 12.04V, the difference between the voltage and 12V is first calculated. Value: 12.04V - 12V = 0.04V; then calculate the speed increment according to the voltage-speed mapping relationship, i.e., 0.04V ÷ 0.1V × 0.02m / s = 0.008m / s. Therefore, the basic stable speed corresponding to 12.04V = the speed corresponding to 12V (0.4m / s) + speed increment (0.008m / s) = 0.408m / s; considering the speed fluctuation of ±0.002m / s during motor operation, and to reserve a safety margin, the safe speed range is set to the basic stable speed ±0.002m / s and rounded up, finally yielding a safe speed range of 0.41 to 0.55m / s. The upper limit of 0.55m / s corresponds to the speed corresponding to the motor's rated voltage of 12.5V, avoiding voltage fluctuations. Overload caused the speed to exceed the limit. Therefore, the target-falling speed was adjusted to 0.5 m / s, which is close to the expected speed of 0.5 m / s and is within the safe operating range of the motor. The optimized target reporting accuracy judgment threshold is 2.02 mm. Based on this, the allowable range of position deviation in the command is set to no more than 2 mm to ensure that the actual deviation is within the acceptable judgment range. The optimized brightness compensation value is 500 cd / ㎡. This value has been verified to ensure that the image sensor can capture a clear image under the current lighting conditions, providing reliable data support for subsequent shooting event recognition and bullet impact point positioning. When generating movement commands, the movement distance calculation method is adjusted from target distance to target distance × 1 based on the X-axis coordinate mapping correction coefficient of 1.06.06. By compensating for mechanical transmission errors through coefficients, the actual position of the target drone after movement is ensured to match the commanded target position. Through the above feedback, optimized parameters are applied in real time to the model and command generation stages. All subsequent shooting event recognition, command generation, status monitoring, and parameter optimization processes are based on the optimized parameters, forming a complete closed loop of data acquisition, command generation, status monitoring, parameter optimization, and feedback application, achieving closed-loop adaptive optimization for intelligent management of the training target drone.

[0066] This embodiment inputs feedback data into the trained model for feature extraction and pattern recognition, accurately pinpointing the root causes of parameter deviations during target machine operation, such as issues related to lighting, motors, and calibration. This avoids blindly adjusting parameters and improves the accuracy of problem diagnosis. Based on the recognition results, adaptive optimization of operating parameters can address different types of operational problems, such as recognition deviations caused by lighting or motion deviations caused by motors. This ensures that the target machine's performance indicators remain stable within a reasonable range, reducing equipment malfunctions or reporting errors caused by improper parameters. Through closed-loop feedback, the optimized parameters are applied to the model and command generation process, continuously improving the model's recognition capabilities and command accuracy. This creates a self-improving cycle in the intelligent management of the target machine, adapting to changes in the range environment and performance degradation after long-term operation, ensuring the long-term stable and accurate operation of the training target machine.

[0067] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0068] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0069] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent management of training targets based on deep learning, characterized in that, The method comprises: Step 1, constructing and training a deep learning neural network model for processing training range multi-source data; Step 2, using the deep learning neural network model to perform feature extraction and fusion processing on video images and sensor data collected in real time at the training range, and converting the same into a spatial coordinate set; determining feature points based on the coordinate set, and establishing a primary direction vector and a secondary direction vector according to the feature points, calculating the orientation relationship between the two direction vectors, and constructing a spatial feature region based on the orientation relationship; Step 3, setting five spatial correction points in the spatial feature region; forming a correction grid according to the spatial distribution of the spatial correction points, and obtaining attitude calibration parameters based on the spatial transformation characteristics of the correction grid; Step 4, performing coordinate transformation correction on the spatial coordinate set using the attitude calibration parameters to obtain a correction result; identifying a shooting event, a point of impact or a light spot position according to the correction result to obtain an identification result; Step 5, determining a target machine control instruction according to the identification result, adjusting the up-down, swing, hide-and-show or movement mode of the target machine, and monitoring the control instruction execution effect to obtain feedback data; Step 6, optimizing the running parameters using the deep learning neural network model according to the feedback data to obtain optimized running parameters; feeding back the optimized running parameters to the deep learning neural network model to realize closed-loop adaptive optimization of the intelligent management process of the training target machine. 2.The deep learning-based intelligent management method of training targets according to claim 1, wherein, The step 1 comprises: Obtaining historical multi-source data in a historical training process in the training range, wherein the historical multi-source data comprises historical video image data and historical multi-modal sensor data; and performing preprocessing on the historical multi-source data to obtain processed historical multi-source data; Labeling the processed historical multi-source data, wherein the labeling information comprises spatial three-dimensional coordinate information and shooting event category labels, and obtaining labeled historical multi-source data; According to the environmental characteristics of the training range and the features of the historical multi-source data, a deep learning neural network model integrating a convolutional neural network and a recurrent neural network is constructed, wherein the convolutional neural network is used to extract spatial features of video image data, and the recurrent neural network is used to process time sequence features of multi-modal sensor data; The deep learning neural network model structure is trained using the labeled historical multi-source data, and the deep learning neural network model structure parameters are adjusted through a stochastic gradient descent optimization algorithm to obtain a trained deep learning neural network model. 3.The deep learning-based intelligent management method of training targets according to claim 2, characterized in that, The step 2 comprises: Inputting the real-time collected video image data and multi-modal sensor data into the trained deep learning neural network model for fusion processing to obtain a multi-dimensional feature representation after fusion; Converting the multi-dimensional feature representation into a spatial coordinate set containing three-dimensional coordinate information, performing density clustering analysis on the spatial coordinate set, and identifying a plurality of position points with the highest spatial distribution density as spatial distribution feature points; determining a primary direction vector and a secondary direction vector through principal component analysis according to the geometric center position and distribution form of the spatial distribution feature points; Calculating the spatial included angle between the primary direction vector and the secondary direction vector through a vector operation method, and determining the orientation relationship between the primary direction vector and the secondary direction vector based on the spatial included angle; According to the position distribution of the orientation relationship and the spatial distribution feature points, the boundary range and the orientation of the spatial feature region are determined, and the spatial feature region is constructed. 4.The deep learning-based intelligent management method of training targets according to claim 3, wherein, The spatial included angle between the main direction vector and the secondary direction vector is calculated by using the vector operation method, and the orientation relationship between the main direction vector and the secondary direction vector is determined based on the spatial included angle, including: The dot product of the two vectors is calculated based on the coordinate components of the main direction vector and the secondary direction vector, and the lengths of the main direction vector and the secondary direction vector are calculated respectively; the spatial included angle between the two vectors is calculated by the inverse cosine function according to the dot product and the lengths of the two vectors; The normal vector of the main direction vector and the secondary direction vector is calculated by the vector cross product operation, and the normal vector is perpendicular to the plane formed by the main direction vector and the secondary direction vector; the orientation relationship between the main direction vector and the secondary direction vector is determined based on the spatial included angle and the direction of the normal vector, and the orientation relationship is represented as the orientation description parameter containing the included angle size and the normal vector direction information. 5.The deep learning-based intelligent management method of training targets according to claim 4, characterized in that, The step 3 includes: Based on the boundary range and the orientation of the spatial feature region, five spatial correction points are mapped and set in the spatial feature region, and the five spatial correction points correspond to the positions of four corners and a center point of the actual target plane respectively; the spatial correction grid is constructed by connecting the correction points to each other through the spatial mapping connection mode according to the three-dimensional coordinates and the topological distribution relationship of the five spatial correction points in space; Based on the spatial corresponding relationship between the spatial correction grid and the actual standard target grid, the pose calibration parameter for correcting the spatial coordinate set is generated by calculating the spatial transformation matrix from the spatial correction grid to the standard target grid, and the pose calibration parameter includes a rotation matrix and a translation vector. 6.The deep learning-based intelligent management method of training targets according to claim 5, wherein, The step 4 includes: The spatial coordinate transformation processing is performed on the spatial coordinate set by using the pose calibration parameter, and the corrected spatial coordinate data is obtained; Based on the corrected spatial coordinate data, the change characteristics of the multi-modal sensor data are analyzed to identify the occurrence of the shooting event, and the abnormal point distribution in the spatial coordinate data is detected, the impact point coordinates or the light spot positions are determined according to the spatial position information of the abnormal points, and the identification result containing the shooting event information and the position information is obtained. 7.The deep learning-based intelligent management method of training targets according to claim 6, characterized in that, The step 5 includes: According to the impact point coordinates or the light spot position information in the identification result, and in combination with the time sequence of the occurrence of the shooting event, the target machine action control instruction set for controlling the movement behavior of the target machine is obtained, and the target machine action control instruction set is used to control the target machine to execute the rise and fall action mode, the swing action mode, the hidden and visible action mode or the moving track mode; In the process that the target machine executes the target machine action control instruction set, the actual movement state of the target machine is monitored through the real-time collected multi-modal sensor data, and the monitoring data containing the target space position, the movement posture and the action response time are obtained; the monitoring data and the expected movement state parameters in the target machine action control instruction set are compared and analyzed, and the feedback data reflecting the instruction execution effect is obtained. 8.The method of claim 7, wherein, The step 6 includes: The feedback data is transmitted to the trained deep learning neural network model, the feature extraction and pattern recognition of the feedback data are performed, and the identification result is obtained. Based on the recognition result, adaptive optimization adjustment is performed on operation parameters, including illumination environment compensation parameters, target motion dynamics parameters and target calibration parameters, to obtain optimized operation parameters; The optimized operation parameters are fed back to the trained deep learning neural network model and the target action control instruction generation process, to realize closed-loop adaptive optimization of the intelligent management process of the training target.

9. A computing device, comprising: Comprise: One or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method as claimed in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program which, when executed by a processor, implements the method as claimed in any one of claims 1 to 8.

Images