A method for processing images of a test of parameters of a group of fragments under sequence laser illumination

Abstract

The application discloses a kind of based on sequence laser illumination under fragment group parameter test image processing method, by building two sequence multi-pulse lasers and the binocular intersection test system of cooperation of two high-speed cameras obtains target rear fragment parameter, high-speed camera is illuminated when multiple continuous fragment target instantaneous images of multiple continuous fragment target instantaneous images by laser pulse are shot, and histogram equalization is used to pretreatment;Based on the improved YOLOv4-tiny algorithm, the single visual angle image after pretreatment is identified to fragment group target;Kalman filter is used to track fragment trajectory;The same fragment target of two camera visual angles is matched based on SURF feature point detection algorithm;According to the two-dimensional coordinate of fragment pixel, three-dimensional coordinate is solved, and motion parameter is obtained.The application is analyzed to this to analyze the damage power of armor-piercing warhead, provides data support for establishing target weapon ammunition damage evaluation system.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an image processing method for testing fragment group parameters under sequential laser illumination. Background Technology

[0002] Back-to-target debris is the main medium causing secondary damage to the internal environment of armor, causing truly large-area and effective damage to core components and personnel inside the armor. Accurately understanding the dispersion characteristics of back-to-target debris is of great significance.

[0003] Currently, high-speed photography is widely used in various weapon parameter testing fields, such as measuring the speed and attitude of objects like fragments, bullets, and rockets. Due to the high speed of fragments, motion blur may occur in the images, and sufficient light is usually required during the shooting process. Lasers, with their good monochromaticity and short pulse time, combined with high-speed cameras, can produce clear images of high-speed moving fragment targets with minimal trailing blur. During shooting, the exposure is almost entirely determined by the laser, overcoming the strong self-illumination effects of transient processes such as warhead impact and explosion. Binocular stereo vision, an important branch of machine vision, performs measurement matching by simulating human eye observation and processing of objects. It is a fast, accurate, and easy-to-operate non-contact measurement method. Analyzing the captured images using binocular stereo vision technology can describe the fragment motion process from two-dimensional to three-dimensional space in the most intuitive way, calculating the three-dimensional coordinates and motion parameters of the fragments. Summary of the Invention

[0004] The purpose of this invention is to propose an image processing method for testing fragment group parameters under sequential laser illumination.

[0005] The technical solution to achieve the objective of this invention is: an image processing method for fragment group parameter testing under sequential laser illumination, comprising the following steps:

[0006] Step 1: Build a binocular intersection test system that combines two sequential multipulse lasers with two high-speed cameras, calibrate the two cameras, and simultaneously acquire fragmented sequence images from the perspectives of the two cameras and perform preprocessing.

[0007] Step 2: Add an ECA attention module after each CSP module. In addition to the two detection heads, add an extra detection head for result detection. Replace all activation functions of the backbone network with Mish activation functions. Based on this, construct an improved YOLOv4-tiny network model to detect fragment targets in fragment sequence images and obtain the target bounding box and center position coordinates.

[0008] Step 3: Model the state equation based on the debris scattering scene, and track the single-view debris target based on Kalman filtering;

[0009] Step 4: Based on the SURF algorithm, match the fragment feature points from the two camera perspectives and calculate the three-dimensional coordinates. Combine this with single-view target tracking to complete the motion trajectory recovery in the spatial coordinate system.

[0010] Step 5: Calculate the flight speed of the fragments within the field of view based on their three-dimensional spatial coordinates.

[0011] Further, in step 1, a binocular intersection testing system is built, consisting of two sequential multi-pulse lasers and two high-speed cameras. The two cameras are calibrated, and fragmented sequence images from both camera perspectives are acquired and preprocessed, including the following steps:

[0012] Step S11: Place two high-speed cameras of the same model outside the blast radius, adjust the field of view to converge in the test area and focus, place two sequential multi-pulse lasers and the high-speed cameras on the same side, adjust the laser optical path to align with the target and the location of the debris scattering, and place a soft light reflector behind the field of view.

[0013] Step S12: Before the experiment, camera calibration is performed using a checkerboard pattern. The checkerboard pattern is placed at various angles in the field of view. Two high-speed cameras simultaneously capture images of the same checkerboard pattern, and the parameters of the two cameras are solved using the checkerboard pattern images.

[0014] Step S13: The high-speed camera continuously captures instantaneous images of the debris target when it is illuminated by the laser to obtain a fragment sequence image;

[0015] Step S14: Histogram equalization is used to preprocess the fragmented image.

[0016] Further, in step 2, an ECA attention module is added after each CSP module. In addition to the two detection heads, an extra detection head is added for result detection. All activation functions in the backbone network are replaced with Mish activation functions. Based on this, an improved YOLOv4-tiny network model is constructed to detect fragment targets in fragmented image sequences, obtaining the target bounding box and center coordinates. This includes the following steps:

[0017] Step S21: Data annotation is performed on the preprocessed fragmented image, including target category and target bounding box, and the image is divided into training set and validation set in a 9:1 ratio.

[0018] Step S22, construct the improved YOLOv4-tiny network structure as follows:

[0019] 1) Backbone Feature Extraction Network Module

[0020] The backbone feature extraction network module consists of two cascaded CBM modules and three CSP modules. First, the input image passes through the two CBM modules sequentially, outputting feature map F1. Each CBM module consists of a convolutional layer, a batch normalization layer, and a Mish activation function layer. Then, feature map F1 is input into the three cascaded CSP modules to obtain feature map F2. Each CSP module has two branches. The first branch connects CBM1, CBM2, CBM3, and CBM4 modules sequentially, with the output of CBM2 also serving as the input to CBM4. The output of CBM4 is connected to the ECA attention mechanism module. The second branch processes the feature map only through the CBM1 module. The two branches converge and output through a max-pooling layer.

[0021] 2) Multi-scale feature fusion network module

[0022] F2 serves as the input to the multi-scale feature fusion network. It passes through modules CBL1 and CBL2 to output feature map P1. P1 passes through CBL3 and an upsampling layer to output feature map P2. P2 passes through CBL4 and CBL5 and an upsampling layer to output feature map P3. The CBL modules consist of a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer, in sequence.

[0023] 3) Prediction Output Module

[0024] The P1 feature map is processed by CBL6 and a convolutional layer to output the detection head Head1; the P2 feature map is processed by CBL4 and a convolutional layer to output the detection head Head2; and the P3 feature map is processed by CBL7 and a convolutional layer to output the detection head Head3.

[0025] Step S23: The network training uses CIOU as the loss function, optimizes the loss function using cosine annealing algorithm with momentum factor of 0.937, trains the network with a batch size of 4, uses a variable learning rate, with an initial learning rate of 0.01, trains the network 300 times, trains the YOLOv4-tiny network model, and obtains the optimal weights.

[0026] Step S24: Detect the fragment image to be tracked and matched, return the fragment target prediction result and output the center coordinates (x,y) of the bounding box.

[0027] Further, step 3, based on the state equation modeled in the debris scattering scene, tracks the single-view debris target using Kalman filtering, including the following steps:

[0028] Step S31: Model the debris dispersion process as a variable-speed motion affected only by air resistance, and create the Kalman filter state equation X for the debris target. t+1 =AX t +Bu tThe state vector of the fragment is represented as (x t ,y t ( ) represents the improved center coordinate position of the YOLOv4-tiny detection at time t. Let u be the velocity of the target at time t in the x and y directions. t Let the acceleration at time t be denoted by A. The state transition matrix A is defined as follows:

[0029]

[0030] The control matrix B is defined as follows:

[0031]

[0032] In the above formula, T is the time interval between two consecutive frames;

[0033] Step S32, divide several fragment targets (x t ,y t The value is initialized to the center position coordinates detected by the improved YOLOv4-tiny algorithm in the first frame. Set to 0 and initialize an ID number for each fragment;

[0034] Step S33: Using the Kalman filter algorithm, predict the center position of the fragment in the next frame as K1, K2...K i Let i represent the total number of predicted results, and let the center positions of the ground truth detection boxes in the next frame be D1, D2...D... j j represents the total number of detection results. The Mahalanobis distance between the i predicted positions and the j detected positions is calculated and used as the weight of the KM allocation algorithm to calculate the most relevant matching pair.

[0035] The association results are divided into three types: the predicted location and the detected location are successfully associated, the detected location is not associated, and the predicted location is not associated.

[0036] When the predicted location and the detected location are successfully associated, the same fragment target in two consecutive frames has the same ID number. The Kalman tracker is updated with the detected target state, and the prediction and association continue in the next frame.

[0037] If the detected locations are not associated, re-initialize these detected targets and continue prediction and association in the next frame;

[0038] If the predicted position is not associated, retain the filter parameters of the predicted position and continue the prediction for the next frame. Set a continuous frame variable n. If it is still not associated after n frames, delete this Kalman filter.

[0039] Once all frames have been processed, output the trajectory coordinates corresponding to each ID.

[0040] Further, in step 4, based on the SURF algorithm, fragment feature points from two camera perspectives are matched and their 3D coordinates are calculated. Combined with single-view target tracking, the motion trajectory in the spatial coordinate system is recovered, including the following steps:

[0041] Step S41: Read the left and right camera images after tracking processing, detect fragment feature points in the images based on the SURF algorithm, and match the same fragment target associated with feature points in the two camera views using the nearest neighbor method.

[0042] Step S42: The pixel coordinates of the left and right camera fragments need to be converted to three-dimensional spatial coordinates in the world coordinate system. Based on the principles of monocular camera imaging and coordinate transformation:

[0043]

[0044] In the formula, the pixel coordinates of the fragments are (u1, v1) and (u2, v2), respectively, and their corresponding world coordinates in space are P(x w ,y w ,z w ), A l and A r These are the intrinsic parameter matrices for the left and right cameras, respectively. R is the rotation matrix, t is the translation vector (obtained through calibration), and s1 and s2 are coefficients.

[0045] The three-dimensional coordinates P(x) of the fragment pixels were obtained using the least squares method. w ,y w ,z w The corresponding position in space is used to complete the motion trajectory recovery in the spatial coordinate system by combining the fragment ID obtained from target tracking.

[0046] Further, step 5, calculating the flight velocity of the debris within the field of view based on its three-dimensional spatial coordinates, includes the following steps:

[0047] With a fixed camera recording frequency, meaning a correlation between frame rate and time, the flight speed of debris within the field of view is determined by the pre-set frame rate of the high-speed camera and the actual spatial distance of the debris's flight.

[0048]

[0049] In the formula, v is the instantaneous velocity, i.e., the average velocity over a short interval, and s is the spatial distance of the debris within the flight time t.

[0050] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the fragment group parameter testing image processing method based on sequential laser illumination to perform fragment group parameter testing image processing based on sequential laser illumination.

[0051] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for processing fragment group parameters under sequential laser illumination is implemented to perform fragment group parameter testing image processing under sequential laser illumination.

[0052] Compared with the prior art, the significant advantages of this invention are:

[0053] 1) Adding an ECA convolutional attention module after each CSP module in YOLOv4-tiny allows the network to reuse useful channel information, and fragment targets will receive more weight attention; adding a new detection head to the network output can adapt to the recognition of fragments of different sizes; replacing the activation function of the network backbone with the Mish function can effectively improve the network detection accuracy.

[0054] 2) Convolutional neural networks, as a high-precision and fast detection algorithm, combined with the stable and efficient Kalman tracking algorithm, can improve the accuracy of target association between frames. Attached Figure Description

[0055] Figure 1 This is a system flowchart of the present invention.

[0056] Figure 2 This is a layout diagram of the test system of the present invention.

[0057] Figure 3 The diagram shows the improved YOLOv4-tiny network structure for this invention.

[0058] Figure 4 This is a flowchart of the fragment trajectory tracking algorithm.

[0059] Figure 5 This is a diagram showing the detection and tracking results of five consecutive frames simulating a scene of flying debris. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] like Figure 1 As shown, a fragment group parameter testing image processing method based on sequential laser illumination includes the following steps:

[0062] Step 1: Build a binocular intersection testing system using two sequential multi-pulse lasers and two high-speed cameras. Calibrate the two cameras and simultaneously acquire fragmented image sequences from both camera perspectives, then preprocess them. Specific operations are as follows:

[0063] like Figure 2 As shown, two ACS-1 high-speed cameras with a resolution of 1280×896 were placed 5m away from the epicenter. The two cameras were at the same height, about 30cm apart, and at a certain angle. The focal length of the lenses was 100mm, and the common field of view was 800mm×600mm. Two lasers were placed about 50cm away from the high-speed cameras, on the same side as the high-speed cameras. The laser beam path was adjusted to be aligned with the debris scattering area. A beam expander was used to diffuse and focus the light spot, ensuring that the image was completely covered by the laser beam. A soft-light reflector was placed 2.5m behind the epicenter to avoid environmental interference and obtain better imaging results.

[0064] Before the experiment, a checkerboard pattern was used for calibration. Two high-speed cameras simultaneously photographed the same checkerboard pattern in the area where the debris was scattered. The checkerboard pattern was placed at various angles in the field of view. The intrinsic and extrinsic parameters of the two cameras were solved according to Zhang Zhengyou's calibration method.

[0065] Under the coordination of the laser high-speed photography synchronization controller, the high-speed camera captures multiple consecutive instantaneous images of the debris target when illuminated by the laser, obtaining a sequence of images of the debris flight.

[0066] Because laser speckle exists in the acquired images, the fragment targets are not clear enough, which is not conducive to subsequent target detection. Therefore, histogram equalization is used to enhance the fragment images, improve image quality and highlight fragment targets.

[0067] Histogram equalization uses the cumulative distribution function of the probability of each gray level in the fragmented image as the transformation function. Following this transformation relationship, an image with a uniform gray-level probability density distribution can be obtained. The cumulative distribution function can be expressed as:

[0068]

[0069] In the formula, 0≤r j ≤1 and k=0,1,…L-1. j T(r) represents the normalized gray level before transformation. k ) is the transformation function, s k For the transformed normalized gray levels, n j p is the number of pixels in the original fragment image that have the k-th gray level, n is the total number of pixels in the fragment image, and p r (r j ) represents the probability that the fragment image before transformation takes the k-th gray level value.

[0070] Step 2: Add an ECA attention module after each CSP module. In addition to the two detection heads, add one extra detection head for result detection. Replace all activation functions in the backbone network with the Mish activation function. Based on this, construct an improved YOLOv4-tiny network model to detect fragment targets in fragmented image sequences and obtain the target bounding boxes and center coordinates. The specific operations are as follows:

[0071] The dataset was labeled using the Labellmg software, generating corresponding txt files. Each fragment's location coordinates were recorded in the generated txt files, including the center coordinates (x, y) of the bounding box, its width w, and height h. The training and validation sets for the detection model were divided in a 9:1 ratio.

[0072] The improved YOLOv4-tiny network structure is constructed as follows:

[0073] 1) Backbone Feature Extraction Network Module

[0074] The backbone feature extraction network module consists of two cascaded CBM modules and three CSP modules. First, the input image passes through the two CBM modules sequentially, outputting feature map F1. Each CBM module consists of a convolutional layer, a batch normalization layer, and a Mish activation function layer. Then, feature map F1 is input into the three cascaded CSP modules to obtain feature map F2. Each CSP module has two branches. The first branch connects CBM1, CBM2, CBM3, and CBM4 modules sequentially, with the output of CBM2 also serving as the input to CBM4. The output of CBM4 is then connected to the ECA attention mechanism module. The second branch processes the feature map only through the CBM1 module. The two branches converge and output through a max-pooling layer.

[0075] 2) Multi-scale feature fusion network module

[0076] F2 serves as the input to the multi-scale feature fusion network. After passing through modules CBL1 and CBL2, it outputs feature map P1. P1 then passes through CBL3 and an upsampling layer to output feature map P2. P2 then passes through CBL4 and CBL5 and an upsampling layer to output feature map P3. The CBL modules consist of convolutional layers, batch normalization layers, and LeakyReLU activation function layers, in sequence.

[0077] 3) Prediction Output Module

[0078] The P1 feature map is processed by CBL6 and a convolutional layer to output the detection head Head1. The P2 feature map is processed by CBL4 and a convolutional layer to output the detection head Head2. The P3 feature map is processed by CBL7 and a convolutional layer to output the detection head Head3.

[0079] The network training uses CIOU as the loss function, and the loss function is optimized using a cosine annealing algorithm with a momentum factor of 0.937. The batch size of the training network is 4, and a variable learning rate is used for training. The initial learning rate is 0.01, and the network is trained 300 times.

[0080] The trained model is loaded with optimal weights, and the fragment images to be tracked and matched are detected. The predicted box output by each detection head is decoded into the bounding box in the actual image. Detection boxes with confidence scores below 0.5 are filtered out, and the remaining bounding boxes are filtered out using a non-maximum suppression algorithm. Then, the detection result is obtained through a classifier. The detection result of this invention is the fragment category. The network input is an image, and the fragment target prediction result is returned and the center coordinates (x, y) of the bounding box are output.

[0081] Step 3: Based on the debris scattering scenario, model the state equation and track the single-view debris target using Kalman filtering. The specific operations are as follows:

[0082] Creating the Kalman filter state equation X for the fragment target t+1 =AX t +Bu t The state vector of the fragment is represented as (x t ,y t ( ) represents the improved center coordinate position of the YOLOv4-tiny detection at time t. Let be the velocity of the target at time t in the x-axis and y-axis directions.

[0083] The debris dispersion process is modeled as a variable-speed motion affected only by air resistance, u t Let the acceleration at time t be denoted by A. The state transition matrix A is defined as follows:

[0084]

[0085] The control matrix B is defined as follows:

[0086]

[0087] In the above formula, T is the time interval between two consecutive frames.

[0088] (x) of several fragmented targets t ,y t The value is initialized to the center position coordinates detected by the improved YOLOv4-tiny algorithm in the first frame. Set to 0 and initialize an ID number for each fragment.

[0089] Using the Kalman filter algorithm, the predicted fragment center positions for the next frame are K1, K2…K. iLet i represent the total number of predicted results, and let the center positions of the ground truth detection boxes in the next frame be D1, D2...D... j Let j represent the total number of detection results. Calculate the Mahalanobis distance d(i, j) between the i predicted locations and the j detected locations, using this distance as the weight for the KM allocation algorithm. Calculate the most relevant matching pair, where:

[0090]

[0091] K i D represents the predicted location of the i-th debris target. j S represents the detection location of the j-th fragment target. i This represents the covariance matrix between the detected location and the average predicted location.

[0092] The association results are divided into three types: the predicted location and the detected location are successfully associated, the detected location is not associated, and the predicted location is not associated.

[0093] When the predicted location and the detected location are successfully associated, the same fragment target in two consecutive frames has the same ID number. The Kalman tracker is updated with the detected target state, and the prediction and association continue in the next frame.

[0094] If the detected locations are not associated, re-initialize these detected targets and continue prediction and association in the next frame;

[0095] If the predicted position is not associated, retain the filter parameters of the predicted position and continue the prediction for the next frame. Set a continuous frame variable n. If it is still not associated after n frames, delete this Kalman filter.

[0096] After all frames have been processed, the trajectory coordinates corresponding to each ID are output. The tracking algorithm flowchart is as follows: Figure 4 As shown.

[0097] Step 4: Based on the SURF algorithm, match fragment feature points from two camera perspectives and calculate their 3D coordinates. Combine this with single-view target tracking to complete the motion trajectory recovery in the spatial coordinate system. The specific operations are as follows:

[0098] Read the left and right camera images after tracking processing, detect fragment feature points in the images based on the SURF algorithm, and match the same fragment target associated with feature points in the two camera views using the nearest neighbor method.

[0099] The pixel coordinates of the left and right camera fragments need to be converted to 3D spatial coordinates in the world coordinate system. Assume the left camera coordinate system coincides with the world coordinate system; in this case, the rotation matrix from the left camera coordinate system to the world coordinate system is R. l =E, translation vector is t l=0. In the left and right cameras, the pixel coordinates of the target detection point are (u1,v1) and (u2,v2) respectively, and their corresponding world coordinates in space are P(x...). w ,y w ,z w Based on the principles of monocular camera imaging and coordinate transformation, we can conclude that:

[0100]

[0101] In the formula, A l and A r Here, R is the intrinsic parameter matrix for the left and right cameras, R is the rotation matrix, and t is the translation vector, obtained through calibration. The coefficients s1 and s2 can be eliminated. The four equations contain three unknowns, and the three-dimensional coordinates P(x) of point P can be obtained using the least squares method. w ,y w ,z w The camera intrinsic and extrinsic parameter matrices are simplified as follows:

[0102]

[0103] The three-dimensional coordinates P(x) of the fragment pixels were obtained using the least squares method. w ,y w ,z w The corresponding position in space is used to complete the motion trajectory recovery in the spatial coordinate system by combining the fragment ID obtained from target tracking.

[0104] Step 5: Calculate the flight velocity of the fragments within the field of view based on their three-dimensional spatial coordinates. The specific steps are as follows:

[0105] Although the actual velocity change of the fragments is not linear, as long as the intercepted ballistic interval is very short, the velocity changes approximately linearly, and the instantaneous velocity at a certain point can be replaced by the average velocity of that segment.

[0106] With a fixed camera recording frequency, meaning a correlation between frame rate and time, the flight speed of debris within the field of view is determined by the pre-set frame rate of the high-speed camera and the actual spatial distance of the debris's flight.

[0107]

[0108] In the formula, v is the instantaneous velocity, i.e., the average velocity over a short interval, and s is the spatial distance of the debris within the flight time t.

[0109] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the fragment group parameter test image processing method based on sequential laser illumination to perform fragment group parameter test image processing based on sequential laser illumination.

[0110] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for processing fragment group parameters under sequential laser illumination is implemented to perform fragment group parameter testing image processing under sequential laser illumination.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An image processing method for fragment group parameter testing based on sequential laser illumination, characterized in that, Includes the following steps: Step 1: Build a binocular intersection test system that combines two sequential multipulse lasers with two high-speed cameras, calibrate the two cameras, and simultaneously acquire fragmented sequence images from the perspectives of the two cameras and perform preprocessing. Step 2: Add an ECA attention module after each CSP module. In addition to the two detection heads, add one extra detection head for result detection. Replace all activation functions in the backbone network with the Mish activation function. Based on this, construct an improved YOLOv4-tiny network model to detect fragment targets in fragmented image sequences, obtaining the target bounding box and center coordinates. This includes the following steps: Step S21: Data annotation is performed on the preprocessed fragmented image, including target category and target bounding box, and the image is divided into training set and validation set in a 9:1 ratio. Step S22, construct the improved YOLOv4-tiny network structure as follows: 1) Backbone Feature Extraction Network Module The backbone feature extraction network module consists of two connected components. Modules and 3 The module first inputs an image, which then passes through two sequentially. Module, output feature map ,in, The module consists of convolutional layers, batch normalization layers, and Mish activation function layers, in that order. The feature maps F1 are then input into these three concatenated layers. The module obtains the feature map ,in, The module includes two branches, the first of which is... , , , The modules are connected sequentially, and at the same time The output is also Input, The output is connected to the ECA attention mechanism module; the feature map of the second branch only passes through... The module processes the two branches, which converge and output through a max-pooling layer. 2) Multi-scale feature fusion network module As input to the multi-scale feature fusion network, it passes through the module , Output feature map , go through and an upsampling layer output feature map , go through and and an upsampling layer output feature map ,in, The modules consist of convolutional layers, batch normalization layers, and LeakyReLU activation function layers, in that order. 3) Prediction Output Module Feature map after With a convolutional layer output detection head , go through Module with a convolutional layer output detection head , go through Module with a convolutional layer output detection head ; Step S23: The network training uses CIOU as the loss function, optimizes the loss function using cosine annealing algorithm with momentum factor of 0.937, trains the network with a batch size of 4, uses a variable learning rate, with an initial learning rate of 0.01, trains the network 300 times, trains the YOLOv4-tiny network model, and obtains the optimal weights. Step S24: Detect the fragment image to be tracked and matched, return the fragment target prediction result and output the center coordinates of the bounding box. ; Step 3: Model the state equation based on the debris scattering scene, and track the single-view debris target based on Kalman filtering; Step 4: Based on the SURF algorithm, match the fragment feature points from the two camera perspectives and calculate the three-dimensional coordinates. Combine this with single-view target tracking to complete the motion trajectory recovery in the spatial coordinate system. Step 5: Calculate the flight speed of the fragments within the field of view based on their three-dimensional spatial coordinates.

2. The image processing method for fragment group parameter testing based on sequential laser illumination according to claim 1, characterized in that, Step 1: Build a binocular intersection testing system using two sequential multipulse lasers and two high-speed cameras. Calibrate the two cameras and simultaneously acquire fragmented image sequences from both camera perspectives, then preprocess them. This includes the following steps: Step S11: Place two high-speed cameras of the same model outside the blast radius, adjust the field of view to converge in the test area and focus, place two sequential multi-pulse lasers and the high-speed cameras on the same side, adjust the laser optical path to align with the target and the location of the debris scattering, and place a soft light reflector behind the field of view. Step S12: Before the experiment, camera calibration is performed using a checkerboard pattern. The checkerboard pattern is placed at various angles in the field of view. Two high-speed cameras simultaneously capture images of the same checkerboard pattern, and the parameters of the two cameras are solved using the checkerboard pattern images. Step S13: The high-speed camera continuously captures instantaneous images of the debris target when it is illuminated by the laser to obtain a fragment sequence image; Step S14: Histogram equalization is used to preprocess the fragmented image.

3. The image processing method for fragment group parameter testing based on sequential laser illumination according to claim 1, characterized in that, Step 3, based on the state equation modeling of the debris scattering scene, and using Kalman filtering to track single-view debris targets, includes the following steps: Step S31: Model the debris dispersion process as a variable-speed motion affected only by air resistance, and create the Kalman filter state equation for the debris target. The state vector of the fragment is represented as , for Continuously improve the center coordinate position of YOLOv4-tiny detection. For the goal Always shaft and Velocity in the axial direction; for Acceleration at any moment, state transition matrix Defined as: (1)； Control Matrix Defined as: (2)； In the above formula, The time interval between two consecutive frames; Step S32, to separate several fragment targets The value is initialized to the center position coordinates detected by the improved YOLOv4-tiny algorithm in the first frame. Set to 0 and initialize an ID number for each fragment; Step S33: Using the Kalman filter algorithm, predict the center position of the fragment in the next frame. , The total number of predicted results is represented by the center position of the ground truth detection box in the next frame. , The total number of test results is used for calculation. Each predicted location and The Mahalanobis distance at each detection location is used as the weight for the KM assignment algorithm to calculate the most relevant matching pair; The association results are divided into three types: the predicted location and the detected location are successfully associated, the detected location is not associated, and the predicted location is not associated. When the predicted location and the detected location are successfully associated, the same fragment target in two consecutive frames has the same ID number. The Kalman tracker is updated with the detected target state, and the prediction and association continue in the next frame. If the detected locations are not associated, re-initialize these detected targets and continue prediction and association in the next frame; If the predicted location is not associated, retain the filter parameters for the predicted location and continue prediction for the next frame, setting a continuous frame variable. ,when If the frame is still not associated, delete this Kalman filter; Once all frames have been processed, output the trajectory coordinates corresponding to each ID.

4. The image processing method for fragment group parameter testing based on sequential laser illumination according to claim 1, characterized in that, Step 4: Based on the SURF algorithm, match the fragment feature points from two camera perspectives and calculate their 3D coordinates. Combine this with single-view target tracking to complete the motion trajectory recovery in the spatial coordinate system, including the following steps: Step S41: Read the left and right camera images after tracking processing, detect fragment feature points in the images based on the SURF algorithm, and match the same fragment target associated with feature points in the two camera views using the nearest neighbor method. Step S42: The pixel coordinates of the left and right camera fragments need to be converted to three-dimensional spatial coordinates in the world coordinate system. Based on the principles of monocular camera imaging and coordinate transformation: (3)； In the formula, the coordinates of the fragment pixels are respectively and The corresponding world coordinates in space are , and These are the intrinsic parameter matrices for the left and right cameras, respectively. For rotation matrix, The translation vector is obtained through calibration. and For coefficients; The three-dimensional coordinates of the fragment pixels were obtained using the least squares method. Based on the corresponding position in space, the fragment ID obtained from target tracking is used to complete the motion trajectory recovery in the spatial coordinate system.

5. The image processing method for fragment group parameter testing based on sequential laser illumination according to claim 1, characterized in that, Step 5: Calculate the flight velocity of the fragments within the field of view based on their three-dimensional spatial coordinates, including the following steps: With a fixed camera recording frequency, meaning a correlation between frame rate and time, the flight speed of debris within the field of view is determined by the pre-set frame rate of the high-speed camera and the actual spatial distance of the debris's flight. (4)； In the formula, Instantaneous velocity, i.e., average velocity over a short interval. For fragment flight time Spatial distance within.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the fragment group parameter testing image processing system based on sequential laser illumination as described in any one of claims 1-5, thereby realizing fragment group parameter testing image processing based on sequential laser illumination.

7. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the fragment group parameter testing image processing system based on sequential laser illumination as described in any one of claims 1-5, thereby realizing fragment group parameter testing image processing based on sequential laser illumination.

Images