Automatic identification method and system for laser positioning position in welded pipe cutting

Abstract

The present application belongs to the technical field of image analysis, and particularly relates to a laser positioning position automatic identification method and system in welded pipe cutting, which comprises the following steps: acquiring continuous image frames of the welded pipe cutting process, pre-processing the continuous image frames and preliminarily screening them with a first preset brightness threshold to generate a candidate laser point set; extracting gradient features in a preset neighborhood window for each candidate point in the candidate laser point set, constructing a structure tensor and performing eigenvalue decomposition to obtain a main eigenvalue and a secondary eigenvalue, and removing according to the main eigenvalue and a first preset energy threshold and a shape constraint index calculated from the main eigenvalue and the secondary eigenvalue and a first preset shape threshold to obtain a finely screened candidate point. The present application can distinguish the real laser point with the highest signal quality from multiple stable trajectories, and realize the identification of the laser positioning position in the welded pipe cutting in a strong interference industrial environment.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology. More specifically, this invention relates to a method and system for automatic laser positioning location recognition in welded pipe cutting. Background Technology

[0002] In modern industrial manufacturing, especially on continuous production lines for metal materials such as steel pipes and welded pipes, laser cutting technology has been widely adopted due to its high efficiency, high precision, and high flexibility. To ensure precise alignment of the cutting path, industrial sites commonly utilize machine vision systems combined with auxiliary laser emitters. A laser beam is projected onto the workpiece surface as a positioning guide point, and an industrial camera monitors the three-dimensional coordinates or two-dimensional image coordinates of this laser point on the workpiece surface in real time, thereby guiding the cutting head to perform dynamic trajectory compensation.

[0003] However, the cutting process of welded pipes generates intense sparks, high-temperature molten pool radiation, and arc flashes. Simultaneously, the oil, scratches, and oxide scale on the welded pipe's surface cause extremely complex specular and diffuse reflection phenomena. These visual interference sources create numerous noise points or bright spots in the images captured by the camera. The brightness, size, and shape of these pseudo-targets are highly similar to the actual laser positioning point. Traditional machine vision recognition algorithms typically rely solely on a single brightness threshold for image binarization and segmentation, supplemented by simple morphological features for target recognition. Therefore, when faced with these strong interferences, traditional machine vision recognition algorithms are prone to misidentifying sparks or strong reflections from the pipe wall as laser positioning points, or failing to recognize them when the laser point is briefly obscured by smoke or dust, affecting the continuity, accuracy, and final yield of laser cutting.

[0004] To overcome the limitations of single-frame image recognition, existing technologies have begun to combine local image feature analysis with temporal target tracking algorithms. This involves preliminary screening by extracting edge or texture features of candidate target points, followed by the use of filters to establish trajectories, predict states, and correlate data for the screened targets. This enhances the adaptability to short-term occlusion or uncertain motion of targets.

[0005] However, at the feature extraction level, this method uses a relatively simple feature representation, which cannot represent the unique energy concentration and Gaussian distribution structure of the laser spot. Its ability to distinguish between persistent, strong interference that is highly similar to the characteristics of the real laser spot is limited. At the temporal tracking level, existing filtering models often use fixed measurement noise covariance parameters, assuming that the reliability of external sensor observation data remains constant. However, in the actual working conditions of welded pipe cutting, the validity of laser point observation data can change drastically due to strong external interference such as sparks and smoke. Fixed filtering models cannot dynamically reduce the confidence level in extremely unreliable, interfered observation results. When the target is obscured or interfered with by strong light, the filter is easily misled by erroneous observation information, resulting in trajectory deviation or divergence, thus leading to tracking loss. Summary of the Invention

[0006] To address the technical problem of accurately identifying the laser positioning position during welded pipe cutting in a strong interference environment, improving the anti-interference capability and tracking stability of the identification, and selecting the real laser point from multiple candidate trajectories, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides an automatic laser positioning position recognition method in welded pipe cutting, comprising: acquiring continuous image frames of the welded pipe cutting process; preprocessing the continuous image frames and initially screening them with a first preset brightness threshold to generate a candidate laser point set; extracting gradient features within a preset neighborhood window for each candidate point in the candidate laser point set, constructing a structure tensor and performing eigenvalue decomposition to obtain principal eigenvalues ​​and secondary eigenvalues; eliminating candidate points based on the principal eigenvalues ​​and a first preset energy threshold, as well as the morphological constraint index calculated from the principal eigenvalues ​​and secondary eigenvalues ​​and a first preset morphological threshold, to obtain finely screened candidate points; and performing temporal correlation on the finely screened candidate points to ensure successful laser positioning. The tracking trajectory is established or updated based on candidate points. The tracking trajectory includes a state vector composed of position coordinates, motion velocity, principal eigenvalues, and morphological constraint index. The historical principal eigenvalue sequence of the tracking trajectory is extracted and the spectral entropy is calculated. The spectral entropy is used as an adjustment parameter to dynamically update the measurement noise covariance matrix of the tracking model. The updated tracking model is used to predict the state at the next moment. The Mahalanobis distance between the observation vector of the current frame's finely selected candidate points and the predicted state is calculated and matched. The state of the successfully matched trajectory is updated or the continuously mismatched trajectory is terminated. Stable trajectories that meet the preset stability conditions are selected. The end point of the trajectory with the lowest global spectral entropy among the stable trajectories is identified as the laser positioning position.

[0008] This invention constructs a structural tensor and utilizes primary and secondary eigenvalues ​​to perform dual-constraint screening of candidate laser points from both energy and morphological dimensions. This eliminates interference from pseudo-target points such as sparks, metal scratches, and reflections in the welded pipe cutting process. During the time-series tracking stage, the spectral entropy of the trajectory's primary eigenvalue is introduced into the measurement noise covariance matrix update of the Kalman filter. This enables the tracking model to adapt to changes in brightness and morphology of the laser point caused by power fluctuations or smoke obstruction, ensuring the continuity and stability of tracking under unstable conditions. By using global spectral entropy as the criterion for trajectory quality judgment, the true laser point with the highest signal quality can be identified from multiple stable trajectories, guaranteeing the uniqueness and reliability of the positioning results. This achieves accurate identification of the laser positioning position for welded pipe cutting in a highly interfering industrial environment.

[0009] Preferably, the process of preprocessing the continuous image frames and performing preliminary screening with a first preset brightness threshold to generate a candidate laser point set includes: converting the acquired color image frames into grayscale images; analyzing the histogram of the grayscale image using the Otsu method to adaptively determine a global brightness segmentation threshold; binarizing the grayscale image based on the global brightness segmentation threshold to extract foreground pixels; performing connected component analysis on all the foreground pixels to extract independent foreground regions; calculating the geometric zeroth moment and first moment of each foreground region; obtaining the geometric centroid coordinates based on the zeroth moment and first moment; and constructing the candidate laser point set from all the geometric centroid coordinates.

[0010] This invention uses the Otsu method to adaptively determine the global brightness segmentation threshold and combines it with the calculation of the centroid of the connected components. It can dynamically adapt to the image binarization segmentation under different lighting conditions, avoid the failure of fixed thresholds, and improve the spatial positioning accuracy of candidate points through sub-pixel level centroid coordinates.

[0011] Preferably, the step of extracting gradient features within a preset neighborhood window and constructing a structure tensor includes: determining a pixel neighborhood window of a preset size centered on each candidate point; calculating the horizontal and vertical gradients of each pixel within the pixel neighborhood window using a Sobel operator of the preset size; combining the horizontal and vertical gradients into a two-dimensional gradient vector; calculating the outer product matrix of the two-dimensional gradient vectors of each pixel; calculating the distance weight coefficient of each pixel within the pixel neighborhood window relative to the center point using a two-dimensional Gaussian function; and multiplying the distance weight coefficients by the corresponding outer product matrix and performing a global summation to construct the second-order structure tensor.

[0012] This invention can highlight the gradient contribution of the central region and suppress edge noise, so that the principal eigenvalues ​​can more accurately characterize the Gaussian energy distribution characteristics of the laser spot and improve the sensitivity of screening for energy-concentrated targets.

[0013] Preferably, the step of eliminating candidates based on the principal eigenvalue and the first preset energy threshold, and the morphological constraint index calculated from the principal eigenvalue and the secondary eigenvalue and the first preset morphological threshold, includes: dividing the difference between the principal eigenvalue and the secondary eigenvalue by the sum of the principal eigenvalue and the secondary eigenvalue to calculate the morphological constraint index; determining whether the principal eigenvalue is less than the first preset energy threshold, and if so, determining that the candidate point has insufficient energy and eliminating it; determining whether the morphological constraint index is greater than the first preset morphological threshold, and if so, determining that the candidate point is an anisotropic distortion point and eliminating it.

[0014] This invention utilizes the morphological constraint index for anisotropy discrimination, which can effectively distinguish between circular laser spots and linear or corner-shaped interference, reducing the false detection rate.

[0015] Preferably, the step of extracting the historical principal feature value sequence of the tracking trajectory and calculating the spectral entropy, and using the spectral entropy as an adjustment parameter to dynamically update the measurement noise covariance matrix of the tracking model, includes:

[0016] Extract the historical main feature values ​​of the most recent first preset number of frames of the tracking trajectory to form a time series; perform a fast Fourier transform on the time series to calculate the power spectral density function in the frequency domain, and normalize the power spectral density function into a probability distribution sequence; calculate the Shannon entropy of the probability distribution sequence as the spectral entropy; multiply the preset basic measurement noise covariance matrix by a linear scaling factor composed of a scaling factor and the spectral entropy to obtain the measurement noise covariance matrix updated at the current time.

[0017] Preferably, the step of matching and associating the observation vector of the current frame's finely selected candidate point with the Mahalanobis distance of the predicted state includes: constructing a four-dimensional observation vector containing position coordinates, observation principal eigenvalues, and observation morphological constraint indices; extracting the corresponding prediction components from the predicted state vector of the tracking trajectory using the observation matrix, and calculating the four-dimensional residual innovation vector between the four-dimensional observation vector and the prediction components; calculating the residual covariance matrix by combining the observation matrix, the prediction covariance matrix, and the measurement noise covariance matrix; calculating the transpose matrix of the four-dimensional residual innovation vector, the inverse matrix of the residual covariance matrix, and the continuous product of the four-dimensional residual innovation vector to obtain the squared Mahalanobis distance; and determining that the finely selected candidate point and the tracking trajectory are successfully matched when the squared Mahalanobis distance is less than a preset chi-square distribution matching threshold.

[0018] Preferably, the step of filtering stable trajectories that meet the preset stability conditions includes: traversing all unterminated tracking trajectories, determining whether the total number of frames in the lifecycle of the tracking trajectory from its establishment to the current time is greater than or equal to a second preset number, and whether it has maintained continuous matching success within the most recent third preset number of frames; if both conditions are met, the corresponding tracking trajectory is stored in the stable trajectory set.

[0019] This invention uses four-dimensional observation vectors combined with Mahalanobis distance for data association, and eliminates invalid associations by setting a chi-square threshold, thereby improving the association stability under complex interference environments.

[0020] Preferably, the step of predicting the state at the next moment using the updated tracking model includes: constructing a state transition matrix in the form of a sixth-order square matrix, wherein the state transition matrix adopts a uniform linear motion model including a sampling time interval for the position coordinates and the motion velocity, and adopts a zero-order hold constant model without time parameter coupling for the principal eigenvalues ​​and the morphological constraint index; and using the state transition matrix to perform a linear transformation on the posterior state vector of the previous moment to calculate the prior predicted state vector of the current moment.

[0021] This invention filters trajectories based on the dual stability conditions of total lifetime frames and recent consecutive successful matching, and then uses the lowest global spectral entropy as the final decision criterion to automatically select the real laser point from multiple coexisting trajectories, thus solving the problem of indecisiveness due to multiple candidate trajectories.

[0022] Preferably, the step of calculating the Mahalanobis distance between the observation vectors of the current frame's finely selected candidate points and the predicted state, and updating the state of successfully matched trajectories or terminating continuously mismatched trajectories, includes: calculating the squared Mahalanobis distance between the observation vectors of all finely selected candidate points in the current frame and the predicted states of all tracking trajectories, and constructing a two-dimensional matching cost matrix; replacing matrix elements in the matching cost matrix whose values ​​are greater than a preset chi-square distribution matching threshold with infinity, and using the Hungarian bipartite graph matching algorithm to solve the global minimum cost allocation scheme of the matching cost matrix, thereby completing the one-to-one association between the observation vectors and the predicted states; for tracking trajectories that have not matched observation vectors, accumulating their mismatch counters, and terminating and deleting the tracking trajectory when the value of the mismatch counters continuously exceeds a set threshold.

[0023] Secondly, the present invention provides an automatic laser positioning position recognition system for welded pipe cutting, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned automatic laser positioning position recognition method for welded pipe cutting is implemented.

[0024] By adopting the above technical solution, the automatic laser positioning position recognition method in the above-mentioned welded pipe cutting is generated into a computer program and stored in a memory so that it can be loaded and executed by the processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0025] The beneficial effects of this invention are as follows:

[0026] This invention constructs a structural tensor and utilizes primary and secondary eigenvalues ​​to perform dual-constraint screening of candidate laser points not only from the energy dimension but also from the morphological dimension, eliminating interference from various pseudo-target points in the welded pipe cutting site. During the time-series tracking stage, the spectral entropy of the trajectory's primary eigenvalue is used in the state update process of the tracking model, enabling the tracking process to closely match the actual changing characteristics of the laser point signal. This addresses brightness and morphological changes caused by power fluctuations or occlusion of the laser point, ensuring the continuity and stability of tracking under unstable conditions. Using global spectral entropy as the basis for trajectory judgment, the true laser point with the highest signal quality can be identified from multiple stable trajectories, guaranteeing the uniqueness and reliability of the positioning results and achieving the identification of the laser positioning position for welded pipe cutting in a highly interfering industrial environment. Attached Figure Description

[0027] Figure 1 This is a flowchart of the automatic laser positioning position recognition method in welded pipe cutting according to the present invention;

[0028] Figure 2 This is a diagram illustrating the effect of the structural tensor dual-constraint sieving method of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] This invention discloses an automatic laser positioning position recognition method for welded pipe cutting, referring to... Figure 1 This includes steps S1-S4:

[0032] S1. Obtain continuous image frames of the welded pipe cutting process, preprocess the continuous image frames and perform preliminary screening with a first preset brightness threshold to generate a candidate laser point set.

[0033] In an optional embodiment, an industrial camera is rigidly mounted to the side above the high-energy laser cutting head to continuously acquire color image frames of the welded pipe surface at a high rate of 100 frames per second.

[0034] When the processor receives the current color image frame, it first iterates through each pixel in the image, executes a color space physical dimensionality reduction transformation algorithm, converts the color image containing three channels (red (R), green (G), and blue (B)) into a single-channel 8-bit grayscale image, and uses the weighted average transformation formula based on human visual perception, as standard by the International Commission on Illumination (ICI) for coordinates... Substitute the pixels into the formula:

[0035] ;

[0036] For example, if the RGB values ​​of a pixel are R=210, G=30, and B=20, then After rounding down, the grayscale value of this pixel is 82.

[0037] Subsequently, the processor performs pixel intensity histogram statistics on the entire generated grayscale image and applies Otsu's method to automatically traverse every potential grayscale threshold between 0 and 255. It adaptively searches for the brightness segmentation threshold that maximizes the inter-class variance as the first preset brightness threshold for the current frame. If the optimal segmentation threshold for the current frame is calculated to be 195, the algorithm traverses the grayscale image. If the grayscale value at the coordinate is strictly greater than 195, then the corresponding position in the binarized mask image is assigned a value of 255; if it is less than or equal to 195, then the value is assigned to 0.

[0038] After binarization, the processor applies a breadth-first search eight-neighbor connected component analysis algorithm based on a queue structure to all foreground pixels in the mask image to extract individual foreground connected regions. Let a given foreground connected region be denoted as a Region. Given the two-dimensional coordinates of each foreground pixel contained within the foreground connected region, calculate its zeroth-order spatial moment for each foreground connected region. :

[0039] ;

[0040] Next, calculate the first-order horizontal spatial moments of the region. Orthogonal space moment :

[0041] ; ;

[0042] Divide the first-order spatial moment by the zeroth-order spatial moment to obtain the precise geometric centroid coordinates of each foreground connected region, which will serve as the coordinates of the candidate center point. :

[0043] ;

[0044] For example, a certain highlighted connected component consists of 4 pixels with coordinates (1500, 800), (1501, 800), (1500, 801), and (1501, 801), which are calculated as follows: , , The centroid coordinates are obtained. All the calculated sub-pixel level geometric centroid coordinates together constitute the candidate laser point set output in this step.

[0045] S2. Extract gradient features within a preset neighborhood window for each candidate point in the candidate laser point set, construct a structure tensor and perform eigenvalue decomposition to obtain principal eigenvalues ​​and secondary eigenvalues. Eliminate candidate points based on the principal eigenvalues ​​and the first preset energy threshold, as well as the morphological constraint index calculated from the principal eigenvalues ​​and secondary eigenvalues ​​and the first preset morphological threshold, to obtain finely screened candidate points.

[0046] In an optional embodiment, the processor delineates a neighborhood window of a preset size of 7×7 pixels on the unbindified original single-channel grayscale image, centered on each candidate point. For each pixel within this neighborhood window... The horizontal gradient is calculated by convolution using a Sobel difference operator with a preset size of 3×3. and vertical gradient Calculated This constitutes the gradient feature vector of the pixel. .

[0047] Calculate the outer product matrix of the gradient eigenvector and its transpose row vector:

[0048]

[0049] To highlight the Gaussian energy distribution characteristics of the laser spot, a two-dimensional Gaussian function is used to calculate the coordinates of the pixel relative to the center candidate point. Distance weight coefficient :

[0050]

[0051] In this embodiment, the standard deviation of the Gaussian distribution is shown. The optimal value is 1.5. The 49 distance weight coefficients are multiplied by their corresponding outer product matrices and then summed globally to construct a structure tensor matrix in the form of a second-order diagonal matrix. :

[0052]

[0053] By solving the characteristic equation Obtain the principal eigenvalues ​​characterizing the absolute intensity of local energy. and secondary eigenvalues ​​characterizing energy in orthogonal directions ,and .

[0054] The processor then executes the double-elimination logic:

[0055] First, it is determined whether the principal feature value is less than the first preset energy threshold. In this embodiment, the first preset energy threshold is set to 2500. If the principal feature value is less than 2500, the candidate point is determined to have insufficient energy, which is weak reflection or background noise, and is therefore rejected.

[0056] Secondly, for the retained candidate points, the morphological constraint index is calculated. The algebraic difference between the principal eigenvalues ​​and the secondary eigenvalues ​​is divided by the algebraic sum to obtain:

[0057]

[0058] Determine whether the shape constraint index is greater than the first preset shape threshold. In this embodiment, the first preset shape threshold is set to 0.5. If the shape constraint index is greater than 0.5, it indicates that the gradient of the target in each direction is extremely uneven, such as extremely narrow metal scratches and reflections, and is determined to be an anisotropic distortion point and is removed.

[0059] For example, the tensor decomposition of a candidate point yields The value is 6185.3. The value is 2314.7, and its principal eigenvalue is greater than 2500, satisfying the condition. (This is the morphological constraint index.) Since 0.455 is less than 0.5, it also meets the condition, and this candidate point is retained as a candidate point for fine screening.

[0060] S3. Perform temporal correlation on the finely selected candidate points, and establish or update the tracking trajectory for the successfully correlated candidate points. The tracking trajectory contains a state vector consisting of position coordinates, motion velocity, principal eigenvalues ​​and morphological constraint index. Extract the historical principal eigenvalue sequence of the tracking trajectory and calculate the spectral entropy. Use the spectral entropy as an adjustment parameter to dynamically update the measurement noise covariance matrix of the tracking model.

[0061] In an optional embodiment, for the first finely screened candidate point, the processor allocates a new Kalman filter tracking trajectory in memory and defines a six-dimensional state vector column matrix. Its elements are, in order: horizontal position coordinates Vertical position coordinates Horizontal speed Vertical motion speed Principal eigenvalues Form constraint index :

[0062]

[0063] Because laser-guided motion is extremely slow for relatively high frame rates, the processor constructs a state transition matrix in the form of a sixth-order square matrix. For position coordinates and motion velocity, a uniform linear motion model incorporating sampling time intervals is adopted; for the principal eigenvalues ​​and shape constraint exponents of purely optical properties, a zero-order hold constant model without time parameter coupling is used.

[0064]

[0065] in The inter-frame sampling time interval is set to 0.01s based on 100fps, using the state transition matrix. For the posterior state vector of the previous time step By performing a linear matrix multiplication transformation, the prior predicted state vector reflecting the internal deduction of the system at the current moment can be calculated. .

[0066] The processor extracts the historical main feature values ​​of the most recent first preset number of frames of the tracked trajectory. In this embodiment, 16 frames are set to form a one-dimensional time series in the time domain. After de-meaning, a 16-point discrete fast Fourier transform is performed to calculate the square of the complex amplitude in the frequency domain, thus obtaining the power spectral density function sequence. Normalization is achieved by dividing each discrete value in the power spectral density function sequence by the sum of the values ​​in the entire sequence, yielding the probability distribution sequence. Then, the frequency domain spectral entropy was calculated based on Shannon's information entropy formula. :

[0067]

[0068] The system's preset basic measurement noise covariance matrix Multiplied by the scaling factor and frequency domain spectral entropy The linear scaling factor constituted in this embodiment The calibration is set to 2, resulting in the updated measurement noise covariance matrix. :

[0069]

[0070] When the target's energy flickers violently due to smoke or spark interference, the frequency domain spectral entropy value soars, causing the measurement noise covariance matrix to increase dramatically. The Kalman filter adaptively reduces its trust in extremely unreliable external observation data and instead relies more on the prediction model established within the system for smooth inference and filtering, thereby effectively resisting sudden strong interference and preventing tracking divergence.

[0071] S4. Predict the state at the next moment using the updated tracking model, calculate the Mahalanobis distance between the observation vector of the candidate points selected in the current frame and the predicted state, update the state of the successfully matched trajectory or terminate the trajectory that is continuously mismatched, select stable trajectories that meet the preset stability conditions, and identify the end point of the trajectory with the lowest global spectral entropy among the stable trajectories as the laser positioning position.

[0072] In an optional embodiment, for the finely selected candidate points extracted in the current frame, an observation vector is constructed that contains only the position coordinates, principal feature values, and morphological constraint indices. Since velocity cannot be directly measured, a linear observation matrix of four rows and six columns is constructed. When mapping the state vector to the observation vector, the corresponding matrix elements , , , Set the elements to 1 and the rest to 0 to calculate the four-dimensional residual information vector:

[0073]

[0074] Combining Kalman prior prediction covariance matrix Covariance matrix of measurement noise Calculate the residual covariance matrix :

[0075]

[0076] Using all the finely selected candidate points detected in the current frame as the set of observations, for each predicted state and each observation, multi-dimensional observation information including position coordinates, principal eigenvalues, and morphological constraint indices is extracted, and the Mahalanobis distance between the observed state and the predicted state is calculated:

[0077]

[0078] The processor calculates Mahalanobis distance pairwise for all carefully selected candidate points and all surviving tracking trajectories in the current frame to construct a two-dimensional matching cost matrix. A preset chi-square distribution matching threshold of 13.28 is set, and elements in the cost matrix greater than 13.28 are replaced with infinity. The Hungarian bipartite graph matching algorithm is used to solve the global minimum cost allocation scheme of the matching cost matrix to complete the one-to-one association. A posteriori update is performed on the successfully matched trajectories. For tracking trajectories that are left out and have not matched the observation vector, their internal mismatch counter is incremented by 1. When the mismatch counter continuously exceeds the set threshold, the processor terminates and deletes the tracking trajectory from memory.

[0079] After the cutting and processing task is completed, the processor traverses all unterminated tracking trajectories and checks whether they simultaneously meet two preset stability conditions: first, whether the total number of frames in the lifetime is greater than or equal to a second preset number (60 frames in this embodiment); second, whether continuous successful matching without mismatch has been maintained within the most recent third preset number of frames (10 frames in this embodiment). If both conditions are met, the trajectory is extracted and stored in a stable trajectory set. All principal feature values ​​generated by each trajectory in the stable trajectory set throughout its entire lifetime are extracted to form an extremely long global time series. A global fast Fourier transform and probability normalization are performed on this global time series, and the global frequency domain spectral entropy representing the randomness of the global energy fluctuation of the trajectory is calculated again.

[0080] The randomness of the sparks results in extremely high global frequency domain spectral entropy, while the real laser point, even if disturbed, has a constant underlying physical energy source and extremely low global frequency domain spectral entropy. The processor selects the trajectory with the lowest global frequency domain spectral entropy as the final and unique true trajectory, extracts the coordinates of the last frame's endpoint, and outputs them to the laser cutting servo mechanism through the communication interface to complete the final industrial physical positioning compensation.

[0081] Reference Figure 2 This demonstrates the effect of eliminating real laser points and interference points based on principal eigenvalues ​​and morphological constraint indices.

[0082] This invention also discloses an automatic laser positioning position recognition system for welded pipe cutting, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the automatic laser positioning position recognition method for welded pipe cutting according to this invention is implemented.

[0083] The laser positioning automatic identification system for welded pipe cutting also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0084] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for automatic laser positioning position recognition during welded pipe cutting, characterized in that, include: Acquire continuous image frames of the welded pipe cutting process, preprocess the continuous image frames and perform preliminary screening with a first preset brightness threshold to generate a candidate laser point set; For each candidate point in the candidate laser point set, the gradient features within a preset neighborhood window are extracted, a structural tensor is constructed, and eigenvalue decomposition is performed to obtain the principal eigenvalue and the secondary eigenvalue. Based on the principal eigenvalue and the first preset energy threshold, as well as the morphological constraint index calculated from the principal eigenvalue and the secondary eigenvalue and the first preset morphological threshold, the candidate points are eliminated to obtain the finely screened candidate points. The selected candidate points are temporally correlated, and a tracking trajectory is established or updated for the successfully correlated candidate points. The tracking trajectory includes a state vector consisting of position coordinates, motion velocity, principal eigenvalues, and morphological constraint index. The historical principal eigenvalue sequence of the tracking trajectory is extracted and the spectral entropy is calculated. The spectral entropy is used as an adjustment parameter to dynamically update the measurement noise covariance matrix of the tracking model. The updated tracking model is used to predict the state at the next moment. The observation vector of the candidate point selected in the current frame is calculated and matched with the Mahalanobis distance of the predicted state. The state of the successfully matched trajectory is updated or the trajectory with continuous mismatch is terminated. Stable trajectories that meet the preset stability conditions are selected. The end point of the trajectory with the lowest global spectral entropy among the stable trajectories is identified as the laser positioning position.

2. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The continuous image frames are preprocessed and initially screened using a first preset brightness threshold to generate a candidate laser point set, including: The acquired color image frames are converted into grayscale images. The Otsu method is used to analyze the histogram of the grayscale image to adaptively determine the global brightness segmentation threshold. Based on the global brightness segmentation threshold, the grayscale image is binarized to extract foreground pixels. Connected component analysis is performed on all the foreground pixels to extract independent foreground regions. The zeroth and first moments of the geometry of each foreground region are calculated, and the geometric centroid coordinates are obtained based on the zeroth and first moments. The candidate laser point set is formed by all the geometric centroid coordinates.

3. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The step of extracting gradient features within a preset neighborhood window and constructing a structure tensor includes: A pixel neighborhood window of a preset size is determined centered on each candidate point. The horizontal and vertical gradients of each pixel within the neighborhood window are calculated using a Sobel operator of the preset size. The horizontal and vertical gradients are combined into a two-dimensional gradient vector, and the outer product matrix of the two-dimensional gradient vectors of each pixel is calculated. A two-dimensional Gaussian function is used to calculate the distance weight coefficient of each pixel within the neighborhood window relative to the center point. The distance weight coefficient is multiplied by the corresponding outer product matrix and then summed globally to construct the second-order structure tensor.

4. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The step of eliminating features based on the principal feature value and the first preset energy threshold, and the form constraint index calculated from the principal feature value and the secondary feature value and the first preset form threshold, includes: dividing the difference between the principal feature value and the secondary feature value by the sum of the principal feature value and the secondary feature value to calculate the form constraint index; Determine whether the main feature value is less than the first preset energy threshold. If it is less, determine that the candidate point has insufficient energy and remove it. Determine whether the morphological constraint index is greater than the first preset morphological threshold. If it is greater, determine that the candidate point is an anisotropic distortion point and remove it.

5. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The step of extracting the historical principal feature value sequence of the tracking trajectory and calculating the spectral entropy, and using the spectral entropy as an adjustment parameter to dynamically update the measurement noise covariance matrix of the tracking model, includes: Extract the historical main feature values ​​of the most recent first preset number of frames of the tracking trajectory to form a time series; perform a fast Fourier transform on the time series to calculate the power spectral density function in the frequency domain, and normalize the power spectral density function into a probability distribution sequence; calculate the Shannon entropy of the probability distribution sequence as the spectral entropy; multiply the preset basic measurement noise covariance matrix by a linear scaling factor composed of a scaling factor and the spectral entropy to obtain the measurement noise covariance matrix updated at the current time.

6. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The process of matching and associating the observation vector of the current frame's finely selected candidate point with the Mahalanobis distance of the predicted state includes: constructing a four-dimensional observation vector containing position coordinates, observation principal eigenvalues, and observation morphological constraint indices; extracting the corresponding prediction components from the predicted state vector of the tracking trajectory using the observation matrix; calculating the four-dimensional residual innovation vector between the four-dimensional observation vector and the prediction components; calculating the residual covariance matrix by combining the observation matrix, the prediction covariance matrix, and the measurement noise covariance matrix; calculating the transpose matrix of the four-dimensional residual innovation vector, the inverse matrix of the residual covariance matrix, and the continuous product of the four-dimensional residual innovation vector to obtain the squared Mahalanobis distance; and determining that the finely selected candidate point and the tracking trajectory are successfully matched when the squared Mahalanobis distance is less than a preset chi-square distribution matching threshold.

7. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The selection of stable trajectories that meet preset stability conditions includes: Traverse all unterminated tracking trajectories and determine whether the total number of frames in the lifecycle of the tracking trajectory from its establishment to the current time is greater than or equal to the second preset number, and whether it has maintained continuous successful matching within the most recent third preset number of frames; If all conditions are met, the corresponding tracking trajectory is stored in the stable trajectory set.

8. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The step of predicting the state at the next time step using the updated tracking model includes: A state transition matrix in the form of a sixth-order square matrix is ​​constructed. In the state transition matrix, a uniform linear motion model including sampling time intervals is adopted for the position coordinates and the motion velocity, and a zero-order hold constant model without time parameter coupling is adopted for the principal eigenvalues ​​and the morphological constraint index. The posterior state vector of the previous moment is linearly transformed using the state transition matrix to calculate the prior predicted state vector of the current moment.

9. The automatic laser positioning position recognition method for welded pipe cutting according to claim 1, characterized in that, The process of calculating the Mahalanobis distance between the observation vectors of the current frame's finely selected candidate points and the predicted state, and updating the state of successfully matched trajectories or terminating continuously mismatched trajectories, includes: calculating the squared Mahalanobis distance between the observation vectors of all finely selected candidate points in the current frame and the predicted states of all tracking trajectories, constructing a two-dimensional matching cost matrix; replacing matrix elements in the matching cost matrix whose values ​​are greater than a preset chi-square distribution matching threshold with infinity, and using the Hungarian bipartite graph matching algorithm to solve for the global minimum cost allocation scheme of the matching cost matrix, completing the one-to-one association between the observation vectors and the predicted states; for tracking trajectories that have not matched observation vectors, accumulating their mismatch counters, and terminating and deleting the tracking trajectory when the value of the mismatch counter continuously exceeds a set threshold.

10. An automatic laser positioning system for welded pipe cutting, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the automatic laser positioning position recognition method for welded pipe cutting according to any one of claims 1-9.

Images