Computer vision-based angle cutting method and device

Abstract

The application relates to the technical field of computer vision, in particular to a computer vision-based angle steel corner cutting method and device, which comprises the following steps: extracting edge features of the angle steel in a local region of interest window, mapping actual cutting parameters; inputting the actual cutting parameters into a preset data sandbox for nested verification; if the actual cutting parameters exceed a preset physical stroke threshold or the change amplitude of the actual cutting parameters relative to a pre-stored historical safe parameter is greater than a preset material continuous deformation threshold, an interception is triggered; otherwise, the nested verification is passed; if the interception is triggered, a compensation observation period is entered; after the actual cutting parameters regenerated in an expanded local region of interest window pass the nested verification for a preset number of times, the compensation observation period is exited. The application blocks destructive out-of-bound actions of machine tools caused by surface pseudo boundaries, thereby guaranteeing safe operation of the corner cutting device.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a computer vision-based method and apparatus for cutting angle steel corners. Background Technology

[0002] Angle steel, as an important basic profile, is widely used in many heavy industrial fields. In automated angle steel processing lines, the corner cutting process is a key step in ensuring the assembly accuracy and overall structural stability of the profiles. With the evolution of manufacturing towards intelligence, the introduction of computer vision technology to capture surface features and locate edge contours of angle steel in high-speed motion, thereby guiding servo mechanisms to complete adaptive processing, has become an inevitable trend for improving production efficiency. In this dynamic processing, the processing frame rate, anti-interference capability, and control of the underlying hardware physical security boundaries of the vision recognition system directly determine the survivability of the entire production line and the quality of the finished corner-cut product.

[0003] Currently, there are existing technological explorations in the industry based on machine vision for processing the morphological features of angle steel. For example, Chinese invention patent application CN104729426A discloses an automatic online inspection system and method for angle steel based on machine vision. This prior art equips each end of a mechanical transmission device with an encoder responsible for sending signals to an image recognition system. When the angle steel to be tested enters the acquisition range of a surface scan camera, the encoder, responsible for position detection and camera trigger sampling, sends a pulse signal to notify the surface scan camera to start acquiring images. The main program of the industrial control computer performs real-time analysis and processing such as edge detection on the images acquired by the surface scan camera, compares the actual detection values ​​with theoretical standard values, and judges the quality of parameters such as the geometric dimensions of the punched holes on the angle steel surface. This method based on encoder timing triggering and image feature comparison achieves, to a certain extent, the automated extraction of the geometric features of the angle steel surface.

[0004] In high-speed, continuous angle steel processing production lines, there is an inherent logical conflict between the local field-of-view tracking mechanism of vision recognition systems and the complex and abrupt changes in surface features in industrial settings. To match the extremely high production pace, vision systems must abandon time-consuming full-image scanning and use extremely small regions of interest for local edge tracking. However, the surfaces of angle steel in heavy industrial settings inevitably contain streaks of rust, large patches of oil, or highly reflective bands caused by cutting fluid. When a narrow strip of oil or highly reflective material traverses a very small local field-of-view window, the vision algorithm, lacking global contextual comparison, is highly susceptible to misjudging its high-contrast contour as the physical boundary of the angle steel. Existing industrial vision control architectures are mostly open-loop trust mechanisms that directly issue vision commands, lacking logic for questioning and intercepting the physical validity of the commands. Under such an architecture, the control system unilaterally trusts the pseudo-boundary extracted by vision and directly retrieves the pre-stored cutting parameters mapped to it. Due to the huge spatial positional deviation between the pseudo-boundary and the real boundary, the system is forced to retrieve extremely abnormal cutting heights and angle offsets. Once this command is issued, the servo mechanism will forcibly adjust its position with maximum torque. This blind execution logic not only prevents the equipment from achieving precise angle cutting, but also causes the cutting tool to directly impact the web of the angle steel or the machine tool chassis due to the movement far exceeding the actual physical clamping clearance of the angle steel, ultimately leading to destructive heavy equipment accidents such as spindle breakage and motor overload. Summary of the Invention

[0005] In order to prevent destructive out-of-bounds actions of machine tools caused by surface pseudo-boundaries and thus ensure the safe operation of corner cutting equipment, this application provides a computer vision-based method and equipment for corner cutting of angle steel.

[0006] Firstly, the angle steel corner-cutting method based on computer vision provided in this application adopts the following technical solution: A computer vision-based angle steel corner-cutting method is applied to a device including a conveying mechanism for conveying angle steel and an execution mechanism for performing the corner-cutting action, the method comprising: Based on the displacement variables of the conveying mechanism and preset reference parameters, a local region of interest window is dynamically generated and tracked. The edge features of the angle steel are extracted within the local region of interest window and mapped to generate actual cutting parameters. The actual cutting parameters are input into a preset data sandbox for nested verification: if the actual cutting parameters exceed the preset physical travel threshold, or the change in the pre-stored historical safety parameters is greater than the preset material continuous deformation threshold, then an interception is triggered; otherwise, the nested verification is passed, and the actual cutting parameters are sent to the actuator to perform the corner cutting action during the non-compensation observation period. If the interception is triggered, a compensation observation period is entered. The compensation parameters are deduced based on the historical security parameters to control the actuator and expand the local region of interest window. After the actual cutting parameters regenerated within the expanded local region of interest window pass the nested verification a preset number of times, the compensation observation period is exited and the local region of interest window is restored to its original size.

[0007] Optionally, the step of dynamically generating and tracking a local region of interest window based on the displacement variable of the conveying mechanism and preset reference parameters includes: Collect real-time encoder pulse data of the conveying mechanism and map it to the displacement variable; Based on the specifications of the angle steel to be processed, the pre-stored cutting parameters are extracted as the reference parameters; By combining the displacement variables and the reference parameters, the movement trajectory of the position to be processed is deduced, and the local region of interest window with its spatial position changing accordingly is generated; Based on the real-time encoder pulse data, image acquisition within the local region of interest window is synchronously triggered.

[0008] Optionally, the nested verification includes, in sequence: When the actual cutting parameters do not exceed the physical travel threshold, the material properties, thickness, and cumulative travel span of the angle steel are obtained. By combining the material property data, the thickness dimension, and the cumulative travel span, the maximum reasonable deformation tolerance is dynamically calculated. The maximum reasonable deformation tolerance is assigned as the continuous deformation threshold of the material. Extract the pre-stored historical safety parameters and calculate the change range between the current actual cutting parameters and the historical safety parameters; Based on the relationship between the material's continuous deformation threshold and the magnitude of the change, the interception or the issuance of the actual cutting parameters is triggered.

[0009] Optionally, the step of extracting the edge features of the angle steel within the local region of interest window and mapping them to generate actual cutting parameters includes: Extract the actual edge contour features of the angle steel within the local region of interest window, and use them as the edge features; The edge features are compared with the parameter feature space in the preset database to retrieve the corresponding target height coordinates and target deflection angle; The actual cutting parameters are generated by combining the target height coordinates and the target deflection angle.

[0010] Optionally, the nested verification includes: Extract the target height coordinates and the target deflection angle from the actual cutting parameters respectively; The target height coordinates and the target deflection angle are independently compared with the physical travel threshold; When either the target height coordinate or the target deflection angle exceeds the physical travel threshold, the interception is triggered to block the issuance of commands to the actuator. When neither the target height coordinates nor the target deflection angle exceed the physical travel threshold, the current actual cutting parameters are released to extract the historical safety parameters and calculate the change range between the current actual cutting parameters and the historical safety parameters.

[0011] Optionally, controlling the actuator based on the compensation parameters derived from the historical safety parameters includes: When the interception is triggered, the current actual cutting parameters are discarded; Extract the pre-stored historical security parameters, which are the actual cutting parameters that have passed the nested verification within a preset number of historical action cycles; Calculate the slope of the change in the historical safety parameters; The compensation parameters are generated by combining the change slope with the current displacement increment determined based on the displacement variable; The compensation parameters are sent to the executing agency.

[0012] Optionally, in expanding the local region of interest window, the method further includes: Obtain the area size of the enlarged local region of interest window, and the number of cycles for continuously issuing the compensation parameters; When the area size reaches a preset global field of view threshold and the interception is still triggered, or when the number of cycles reaches a preset compensation number threshold and the interception is still triggered, the generation of the compensation parameters is stopped and the expansion of the local region of interest window is stopped. An alarm signal is triggered, and an emergency stop braking command is issued to the actuator and the conveying mechanism.

[0013] Optionally, expanding the local region of interest window includes: The area of ​​the local region of interest window is gradually increased by a preset magnification, and the edge features are extracted after each level of expansion to regenerate the actual cutting parameters; When the regenerated actual cutting parameters pass the nested verification, the expansion of the local region of interest window is paused, and the real-time position coordinates corresponding to the current edge feature are obtained.

[0014] Optionally, restoring the local region of interest window to its original size includes: Using the latest acquired real-time location coordinates as the tracking anchor point, the area of ​​the local region of interest window is reset to its original size. Stop calculating the compensation parameters, update the verified actual cutting parameters to the historical safety parameters, and resume issuing the actual cutting parameters.

[0015] Secondly, the computer vision-based angle steel corner cutting device provided in this application adopts the following technical solution: the computer vision-based angle steel corner cutting device includes a conveying mechanism for conveying angle steel, an execution mechanism for performing the corner cutting action, and a controller, the controller including: The simulation and tracking module is used to dynamically generate and track a local region of interest window based on the displacement variables of the conveying mechanism and preset reference parameters. The extraction and mapping module is used to extract the edge features of the angle steel within the local region of interest window and map them to generate actual cutting parameters. The nested verification module is used to input the actual cutting parameters into a preset data sandbox for nested verification: if the actual cutting parameters exceed the preset physical travel threshold, or the change range of the pre-stored historical safety parameters is greater than the preset material continuous deformation threshold, then an interception is triggered; otherwise, the nested verification is deemed to pass, and the actual cutting parameters are sent to the actuator to perform the corner cutting action during the non-compensation observation period. The collaborative correction module is used to enter a compensatory observation period if the interception is triggered, control the execution mechanism based on the historical security parameters to deduce the compensatory parameters, and expand the local region of interest window; after the actual cutting parameters regenerated in the expanded local region of interest window pass the nested verification a preset number of times, the module exits the compensatory observation period and restores the local region of interest window to its original size.

[0016] In summary, this application includes the following beneficial technical effects: 1. This application constructs a two-layer nested verification mechanism that combines physical travel threshold and material continuous deformation threshold. It independently verifies the actual cutting parameters before they are issued, preventing abnormal parameters caused by false boundary detection by the vision system from entering the execution stage. This solves the problem of machine tool boundary collision caused by unconditional trust in visual recognition results in the prior art from the control architecture level, and significantly improves the operational safety of the corner cutting equipment.

[0017] 2. This application dynamically generates and tracks a local region of interest window based on the displacement variables and reference parameters of the conveying mechanism, so that the visual acquisition area moves synchronously with the movement trajectory of the angle steel. While ensuring high frame rate image processing efficiency, it accurately locks the expected appearance area of ​​the angle steel edge, overcoming the dual contradiction of low global scanning efficiency or easy loss of features in the existing technology, and achieving a balance between visual processing speed and anti-interference capability.

[0018] 3. After triggering parameter interception, this application enters a compensation observation period. By extrapolating compensation parameters from historical safety parameters, the actuator is controlled and the local region of interest window is expanded simultaneously. The processing action is maintained continuously under abnormal working conditions. Normal parameters are issued only after the vision system completes self-correction and passes multiple consecutive verifications. This avoids the passive handling mode of stopping the machine immediately when there is an abnormality in the prior art, and ensures the continuous operation capability and processing stability of the production line under complex working conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the physical hardware architecture of the angle steel corner cutting device according to an embodiment of this application; Figure 2 This is a logic flowchart of the angle steel corner cutting method according to an embodiment of this application; Figure 3 This is a comparison diagram of the abnormal magnification of the local region of interest window and the actual boundary finding process in an embodiment of this application; Figure 4 This is a waveform diagram showing the timing relationship of the synchronous triggering control of system actions in an embodiment of this application. Detailed Implementation

[0020] The following combination Figures 1-4 This application will be described in further detail.

[0021] This application discloses a computer vision-based angle steel corner cutting method, which is applied to a computer vision-based angle steel corner cutting device. The device includes a conveying mechanism, a vision acquisition device, an execution mechanism, and a controller.

[0022] Combination Figure 1 As shown, the conveying mechanism uses a servo motor-driven roller conveyor line to drive the angle steel to be processed along a preset horizontal path at a constant speed. An incremental encoder is coaxially mounted on the drive end of the conveying mechanism. The encoder rotates synchronously with the conveying roller to collect the angular displacement data of the conveying mechanism in real time. In this embodiment, the incremental encoder has a resolution of 1000 pulses per revolution, the roller circumference of the drive roller of the conveying mechanism is 200mm, and the linear conveying displacement of the angle steel corresponding to a single pulse is 0.2mm. The controller calculates the real-time movement displacement of the angle steel by accumulating the number of pulses output by the encoder per unit time. The update frequency of the displacement data is consistent with the encoder pulse output frequency to ensure the real-time nature of the displacement acquisition.

[0023] The vision acquisition device employs a global shutter area array industrial camera, fixedly mounted directly above the conveying plane of the conveyor mechanism. The camera's lens optical axis is perpendicular to the conveying plane, and the camera's global field of view covers the full lateral width of the angle steel conveying path, ensuring that the angle steel does not exceed the camera's shooting range during conveying. In this embodiment, the camera's lateral field of view is 200mm, corresponding to 2048 effective lateral pixels. The actual physical size of a single pixel is approximately 0.0977mm, providing a fixed calibration benchmark for subsequent spatial coordinate conversion. The camera's trigger port is electrically connected to the controller's synchronization signal output terminal, enabling it to receive trigger signals from the controller to complete image acquisition.

[0024] The actuator is a servo-driven corner-cutting tool post, which is mounted on a follower mechanism that can move synchronously along the angle steel conveying direction (Y-axis). The tool post carries carbide cutting tools and can be adjusted in height along the Z-axis perpendicular to the conveying plane, and in deflection angle around the X-axis perpendicular to the angle steel conveying direction. The tool post's mechanical structure has fixed physical limits: a maximum downward physical limit of 150mm and a minimum upward physical limit of 0mm in the Z-axis direction; a maximum positive deflection limit of 15° and a maximum negative deflection limit of -15° in the X-axis direction. These physical limits restrict the maximum adjustment range of the tool post and prevent hard collisions in the mechanical structure.

[0025] The controller is electrically connected to the encoder, vision acquisition device, and actuator to perform timing synchronization control, data processing, parameter verification, and command issuance. The controller has a built-in preset database for storing angle steel specifications, standard contour features, equipment limit parameters, and historical safety parameters. The controller's memory contains a preset data sandbox, an independent verification and execution space isolated from the main control program logic. All parameters to be issued to the actuator must undergo a complete verification process within the data sandbox. Only parameters that pass verification can be retrieved and issued by the main control program, isolating abnormal parameters at the software level and preventing them from directly entering the execution phase.

[0026] There are two core technical defects in the existing automated angle steel cutting process. First, there is an inherent contradiction in the vision acquisition mode. To match the high-speed production cycle, the vision system needs to abandon time-consuming full-image scanning and use fixed-size local regions of interest for edge tracking. However, as a basic material in heavy industry, angle steel inevitably has rust spots, oil stains, and high-contrast pseudo-contours formed by cutting fluid reflections on its surface. When pseudo-contours cross the local field of view, the vision system, without global context reference, is very likely to misjudge the pseudo-contours as the real edges of the angle steel, leading to incorrect feature extraction. Second, there are safety risks in the control architecture. Existing technologies mostly adopt an open-loop trust mechanism that directly maps visual recognition results to control parameters and directly issues parameters for execution. This lacks the logic to question and intercept the physical rationality of the instructions. Incorrect parameters can cause the tool holder to overstep its limits, leading to equipment accidents such as tool collisions with the angle steel web, machine tool chassis, spindle breakage, and motor overload. At the same time, existing technologies often directly trigger emergency stops after detecting anomalies, which cannot maintain continuous operation of the production line and seriously affects production efficiency.

[0027] like Figure 2 As shown, this embodiment addresses the core deficiencies of the existing technology by constructing a complete technical chain encompassing timing tracking, feature extraction, parameter verification, fault tolerance correction, and closed-loop execution. It balances visual processing efficiency with feature extraction accuracy through a dynamically adjusted region of interest window that moves with the angle steel. A double-nested parameter verification mechanism blocks the transmission of erroneous parameters, mitigating equipment failure risks at the control architecture level. Furthermore, a smooth compensation and visual correction coordination mechanism under abnormal operating conditions enables system self-correction without triggering shutdowns, maintaining continuous production line operation and achieving a balance between processing efficiency, processing accuracy, and equipment operational safety. The following steps will be described in detail: S1 Dynamic Local Region of Interest Window Generation and Tracking S11 Real-time acquisition and mapping of displacement variables The controller collects pulse data output from the encoder of the conveying mechanism in real time. To avoid the infinite accumulation of macroscopic displacement data, when the corresponding cut corner segment of the angle steel to be processed enters the global field of view of the vision acquisition device, this is used as a reference starting point to perform relative cumulative counting of the subsequently collected pulse data. Combined with the encoder resolution and the circumference parameters of the conveying mechanism's drive roller, this relatively cumulative pulse count is mapped to the real-time relative displacement of the cut corner segment within the field of view, and this real-time relative displacement is set as the displacement variable. The encoder is coaxially mounted on the drive roller end of the conveying mechanism, rotating synchronously with the conveying roller. In this embodiment, the incremental encoder has a resolution of 1000 pulses per revolution, the circumference of the conveying mechanism's drive roller is 200mm, and the linear conveying displacement of the angle steel corresponding to a single pulse is 0.2mm. The update frequency of the displacement data is completely consistent with the encoder's pulse output frequency, ensuring the real-time nature of the displacement acquisition.

[0028] Matching and retrieving S12 reference parameters The controller first obtains the specifications of the angle steel to be processed, then extracts the corresponding pre-stored cutting parameters from the preset database based on these specifications, and sets these pre-stored cutting parameters as the reference parameters. The reference parameters include the standard edge initial coordinates, standard cutting height, and standard deflection angle of the corresponding angle steel specifications. This clearly defines the expected area where the angle steel edge will appear in the camera's global field of view, providing a stable spatial position reference for the generation of the region of interest window. It can also adapt to the processing needs of angle steel of different specifications, eliminating the need to recalibrate the field of view every time the processing batch is changed.

[0029] S13 Dynamic generation and tracking of the local region of interest window The controller combines real-time updated displacement variables with baseline parameters to deduce the real-time movement trajectory of the angle steel edge to be processed within the camera's global field of view. Using the initial coordinates of the standard edge in the baseline parameters as the origin, and superimposing the relative pixel offsets corresponding to the displacement variables within the current processing cycle, the expected spatial coordinates of the angle steel edge at each moment are calculated. This cleverly avoids coordinate addressing out-of-bounds errors caused by the absolute accumulation of macroscopic displacements. The pixel offset conversion is based on the camera's calibrated physical equivalent of 0.0977 mm / pixel, converting the displacement variable in mm units to the offset value in pixels, ensuring the accuracy of the coordinate deduction. Then, along the trajectory of this coordinate change, a local region of interest window with a dynamically changing spatial position is generated. Initially, the size of the region of interest window is 200×100 pixels. This size is set to 1.2 times the edge width of the largest processing specification angle steel. This ensures complete coverage of the expected area where the angle steel edge will appear, accommodates minor lateral deviations during angle steel transport, and compresses the data volume of a single frame image processing to less than 0.48% of the full-frame image, significantly reducing the controller's computational load, improving image processing response speed, and perfectly adapting to the processing cycle of high-speed production lines. The center position of the window always coincides with the calculated expected coordinates of the angle steel edge and moves synchronously within the field of view as the angle steel moves, locking the expected area where the angle steel edge will appear throughout the process. This eliminates the need for time-consuming full-image scanning and successfully solves the problem of traditional fixed windows easily losing edge features during high-speed transport.

[0030] Synchronous Triggering of S14 Local Window Image Acquisition The controller presets the pulse interval between two adjacent image acquisitions based on the encoder's real-time pulse data. Whenever the accumulated pulse count reaches the preset pulse interval, it sends a synchronization trigger signal to the vision acquisition device via the hardware I / O port, triggering the vision acquisition device to complete image acquisition only within the current local region of interest window. In this embodiment, the pulse interval between two adjacent image acquisitions is set to 50 pulses, corresponding to a 10mm travel displacement of the angle steel. This interval setting matches the normal conveying speed of the angle steel, ensuring continuous connection of the edge areas of adjacent acquisition frames, and also ensuring that each corner cutting position of the angle steel can acquire at least 3 effective images, avoiding the omission of edge features, while also preventing excessive processing load on the controller due to excessively high acquisition frequency.

[0031] Combination Figure 4 As can be seen from the correspondence between the time axis and the waveform, the operation of this application is strictly controlled by the hardware-level low-level timing. The continuous mechanical displacement pulses output by the encoder directly trigger the camera's IO port for synchronous shooting, ensuring that the moment of freezing of each frame is absolutely bound to the spatial position of the angle steel. At the same time, the image processing time, the control command issuance nodes, and the execution actions of the servo mechanism are precisely coordinated within the continuous waveform gaps. Figure 4 This study clearly reveals the seamless collaboration mechanism between visual algorithms and mechanical movements, proving that this application can completely eliminate image lag, motion blur, and position misalignment problems commonly found on high-speed production lines. This hardware triggering mode, synchronized with encoder pulses, ensures that each captured image frame corresponds perfectly to the actual movement position of the angle steel, avoiding motion blur or position misalignment problems that occur during high-speed conveying, and providing a clear and synchronized image foundation for subsequent image processing.

[0032] S2 Angle Steel Edge Feature Extraction and Actual Cutting Parameter Mapping S21 Edge contour feature extraction within the local window The controller first receives the image within the dynamically moving region of interest (ROI) window, synchronously triggered by the encoder in the S1 vision acquisition device. Gaussian filtering is applied to the image to remove noise from industrial electrical interference and surface reflections, as well as random noise. Then, grayscale normalization is performed on the filtered image, unifying the grayscale values ​​of all pixels to the range of 0 to 255. Next, a fixed threshold is used for segmentation, converting the image into a clear binary image. Finally, the Canny edge detection operator is used to extract continuous edge line contours from the binary image, and these contours are used as the edge features of the angle steel. A 3×3 Gaussian filter kernel is chosen because it is compatible with the 200×100 pixel ROI window size set in S1, effectively smoothing noise without weakening the grayscale gradient changes of the angle steel edges and avoiding blurring of contour details. The fixed threshold for segmentation is set to 127 because the image has already undergone grayscale normalization in the range of 0-255, and 127 is the median value of the range. The local region of interest window of the dynamic tracking only contains two main grayscale levels: the edge of the angle iron and the transport background. Median segmentation can stably separate the foreground and background, and is not affected by global illumination fluctuations. The processing speed is also much faster than the adaptive threshold algorithm, and can keep up with the processing pace of the high-speed pipeline.

[0033] Feature space comparison of S22 edge features The controller compares the extracted edge features with the parameter feature space of the same specification as the reference parameters retrieved by S1 in the preset database. This parameter feature space is pre-constructed for each specification of angle steel under calibrated working conditions. Specifically, it does this by pre-fine-tuning the lateral position and deflection posture of the angle steel of the same specification during clamping and positioning to simulate various edge position deviations that may occur on the actual production line. Whenever the angle steel edge experiences a positional offset of 0.1mm or an angular deflection of 0.1°, the controller uses a vision acquisition device to capture the actual edge contour of the corresponding angle steel in the current field of view. At the same time, it calculates and records the target height coordinates and target deflection angle that the actuator tool holder must adjust accordingly in order to accurately cut the angle steel in this posture.

[0034] Then, the extracted standard edge contour feature point sets are bound to the corresponding tool holder height and angle parameters and stored in the database, ultimately forming a complete parameter feature space for the angle steel of this specification. This ensures that each set of feature shapes corresponds to a unique and correct machining coordinate. The adjustment interval of 0.1mm and 0.1° is set in accordance with the conventional accuracy requirements for angle steel machining. This ensures that the coverage density of the feature space meets the machining accuracy requirements, while preventing the feature space data from becoming too large due to excessively small intervals, which would affect the comparison efficiency.

[0035] During the comparison, the controller first samples the extracted actual edge contour and the standard edge contour in the feature space at equal intervals, taking 100 evenly distributed contour feature points from each. The setting of 100 sampling points is based on the contour length within the local region of interest window, which can completely restore the shape features of the contour without increasing the amount of computation. Then, the Euclidean distance between corresponding points in the two sets of feature points is calculated, and the average value of the Euclidean distance of all points is taken. The set of standard edge contour features with the smallest average value is selected, and then the target height coordinates and target deflection angle bound to this set of features are retrieved.

[0036] Using this method for feature matching can quickly lock in the standard parameters that best match the actual edge, reducing the intermediate error steps of traditional methods that first measure the size and then convert the cutting parameters. It also meets the fast processing needs of local window feature extraction and will not slow down the overall processing pace.

[0037] Generation of actual cutting parameters for S23 The controller combines the obtained target height coordinates and target deflection angle to generate the actual cutting parameters for the current action cycle. These parameters include the target height coordinates and target deflection angle of the actuator's tool holder. The numerical format and units of these parameters perfectly match the communication protocol of the actuator's servo drive; the height coordinates are in mm, and the deflection angle is in °, allowing direct control of the actuator's tool holder's position adjustment. The generated actual cutting parameters correspond perfectly with the angle steel specifications and reference parameters determined in S1, ensuring the parameters' compatibility with the current machining object and providing precise numerical data for subsequent tool holder motion control.

[0038] Double-layer nested verification and command control of actual S3 cutting parameters S31 Physical Travel Threshold Preset and Parameter Decomposition After receiving the actual cutting parameters generated in S2, the controller first sends the parameters into a preset data sandbox in memory. Then, based on the mechanical limit parameters of the tool holder, it presets the physical travel thresholds. The physical travel threshold for the height coordinate corresponds exactly to the tool holder's Z-axis mechanical limit calibrated in S1, set to 0mm to 150mm. The physical travel threshold for the deflection angle corresponds exactly to the tool holder's X-axis mechanical limit, set to -15° to 15°. These two thresholds are entirely defined by the mechanical hard limit boundaries of the actuator, ensuring that parameters exceeding the thresholds will inevitably trigger a hard collision with the mechanical structure, providing an absolute physical boundary benchmark for the entire verification logic. Next, the controller breaks down the actual cutting parameters of the current action cycle, extracting the target height coordinate and target deflection angle separately for subsequent independent comparison verification. This independent separation of the two parameters avoids the problem of a single parameter anomaly being masked by the overall parameters, improving the accuracy of subsequent verification steps.

[0039] S32 outer layer physical limit boundary verification The controller compares the extracted target height coordinates with the physical travel threshold for height coordinates, and then compares the extracted target deflection angle with the physical travel threshold for deflection angle. If either the target height coordinates or the target deflection angle exceeds the corresponding physical travel threshold, the controller determines that the current actual cutting parameters have a serious risk of exceeding the limit and directly triggers an interception action, completely blocking the path of the actual cutting parameters to the actuator. If neither the target height coordinates nor the target deflection angle exceeds the corresponding physical travel threshold, the controller allows the current actual cutting parameters to proceed to the subsequent verification stage. This step uses the mechanical physical limits of the equipment itself as the absolute verification boundary, which can directly intercept extreme abnormal parameters that would cause a hard collision between the tool holder and the mechanical structure, setting up the first line of defense for equipment operation safety. At the same time, the design of completing the outer layer verification before entering the inner layer logic can directly filter out extreme abnormal parameters without performing subsequent redundant calculations, adapting to the rapid processing requirements of high-speed pipelines.

[0040] Dynamic calculation of the continuous deformation threshold of S33 material The controller extracts the material properties and thickness of the angle steel to be processed from a preset database. Combining this data with the cumulative travel distance from the latest historical safety parameter, it dynamically calculates the maximum reasonable deformation tolerance of the angle steel and assigns this maximum reasonable deformation tolerance as the material's continuous deformation threshold. Angle steel is a rigid metallic material; during continuous conveying, without external force forcibly breaking it, the edge spatial position within adjacent action cycles will not undergo abrupt changes exceeding the material's own mechanical elasticity.

[0041] In specific calculations, based on the elastic bending deformation law of simply supported beams in mechanics of materials, and combined with the elastic modulus, moment of inertia of the section, and cumulative travel span of the material, the maximum elastic bending deformation of the angle steel under this working condition is calculated, and then the maximum change in deflection angle is calculated accordingly. In this embodiment, for rigid angle steel of Q235 material with a thickness of 5mm, under the minimum conventional travel span (i.e., 10mm), combined with the rigid mechanical characteristics of Q235 steel and the above deformation calculation logic, the maximum reasonable pure elastic height bending deformation of this specification of heavy angle steel within an extremely short span of only 10mm is extremely small, not exceeding 0.08mm. However, considering the physical equivalent of the underlying single pixel of the visual acquisition device of approximately 0.0977mm, and the feature matching step size of 0.1mm in the preset database, in order to accommodate the reasonable quantization error of the underlying system and avoid false interception caused by normal numerical discrepancies, the controller adds the basic measurement tolerance to the theoretical elastic deformation, dynamically setting the height coordinate change under a 10mm span to not exceed 0.25mm. Next, based on trigonometric function conversion, the corresponding change in deflection angle is approximately 1.4°. Therefore, the final dynamic setting of the continuous material deformation threshold for this benchmark span is a height coordinate change of no more than 0.25mm and a deflection angle change of no more than 1.4°, and it is adaptively amplified proportionally as the span increases. For angle steel of different materials and thicknesses, the controller can dynamically adjust the continuous material deformation threshold according to the same mechanical tolerance logic, ensuring the scientific adaptability of the verification logic to different processing objects and avoiding the system verification deadlock problem caused by rigid fixed thresholds.

[0042] Verification of the deformation law of S34 inner layer material The controller retrieves pre-stored historical safety parameters from a preset storage queue. These historical safety parameters are the actual cutting parameters that have passed the most recent five consecutive complete verification processes. It then calculates the change in the current actual cutting parameter compared to the historical safety parameters from the previous action cycle. This change includes the change in height coordinates and the change in deflection angle. If the change in height coordinates exceeds the height tolerance within the material continuous deformation threshold, or if the change in deflection angle exceeds the angle tolerance within the material continuous deformation threshold, the controller determines that the currently extracted edge features have a risk of false boundary detection, triggers an interception action, and blocks the path of the actual cutting parameter to the actuator. If neither the change in height coordinates nor the change in deflection angle exceeds the corresponding material continuous deformation threshold, the controller determines that the current actual cutting parameter is valid, completes the entire nested verification process, and releases the actual cutting parameter.

[0043] This step uses the continuous deformation law of rigid metal materials as the verification benchmark. It can identify small abnormal parameters that cannot be intercepted by the outer verification, accurately block erroneous actions caused by false boundary false detection, set up a second line of defense for equipment operation safety, and break the technical bias of unconditionally trusting the results of visual recognition in the existing technology. It can solve the equipment safety problem caused by visual false detection from the control architecture level without relying on complex visual algorithm optimization.

[0044] Issuance of S35 valid parameters and updating of historical data When the controller confirms that the system is within the non-compensatory observation period, it directly sends the valid actual cutting parameters, which have completed all nested verification processes, to the actuator. Simultaneously, it updates these valid actual cutting parameters to the historical safety parameter storage queue. The storage queue uses a first-in, first-out update rule, always retaining the five most recent valid historical safety parameters. The queue length is set to five because this length fully reflects the continuous changing trend of the angle steel edge position without increasing the computational load due to excessive historical data. It perfectly matches the processing cycle of a high-speed pipeline and provides stable baseline data for the inner verification stages of subsequent action cycles.

[0045] Compensatory control and visual collaborative correction after S4 anomaly interception S41 Abnormal Parameter Discard and Compensation Benchmark Extraction When the controller triggers the interception action in step S3, it directly discards the current abnormal actual cutting parameters and does not issue any instructions related to the abnormal parameters to the actuator, thus preventing erroneous parameters from entering the execution stage at the source. Then, the controller extracts all five consecutive historical safety parameters from the continuously updated historical safety parameter storage queue in step S3, using them as benchmark data for subsequent compensation parameter deduction. The queue length is fixed at five parameters because this length can fully reflect the linear change trend of the angle steel edge position during continuous conveying, while avoiding the introduction of early interference deviations due to excessive historical data, thus effectively balancing the accuracy and computational efficiency of subsequent compensation deduction.

[0046] Derivation and Issuance of S42 Compensation Parameters The controller first calculates the slope of change for five consecutive historical safety parameters in chronological order. This slope includes the average slope of change for the height coordinate and the average slope of change for the deflection angle. Specifically, it first calculates the single change in height coordinate and deflection angle within two adjacent historical periods. Then, combining this with the angle steel displacement increment at each period interval, it calculates the parameter change corresponding to a unit displacement. The average value of all unit displacement changes is then taken to obtain the corresponding average slope, ensuring that the dimensions of the slope perfectly match the subsequent displacement increment. Next, the controller combines the calculated slope with the cumulative displacement increment of the angle steel in the current action period determined in step S1 to deduce the expected change trend of the angle steel edge in the current action period. It then generates compensation parameters that perfectly match this change trend. During the deduction process, the values ​​of the compensation parameters are simultaneously verified to ensure that they always fall within the physical travel threshold range, avoiding the risk of the compensation parameters themselves exceeding the limits. The generated compensation parameters closely match the actual movement trend of the angle steel, ensuring smooth and continuous movement of the actuator's tool holder, preventing sudden changes or pauses, and preventing equipment vibration caused by parameter jumps. Finally, the controller sends the generated compensation parameters to the actuator, controlling the actuator to complete the position adjustment for the current action cycle. To ensure the authenticity and purity of the subsequent verification benchmark, these compensation parameters are only issued as control commands and are not updated in the pre-stored historical safety parameter queue.

[0047] S43 Gradual Enlargement of the Region of Interest Window While generating compensation parameters, the controller sends a window expansion command to the vision acquisition device, progressively expanding the area of ​​the local region of interest window according to a preset magnification. In this embodiment, the preset magnification is set to 2 times. This magnification setting ensures that each window expansion can escape the current local interference area and obtain more global context information, while preventing a significant decrease in image processing efficiency due to an excessively large single window expansion range, thus achieving a good balance between anti-interference capability and processing speed.

[0048] Combination Figure 3 A top-down comparison of the three stages—(a) normal tracking, (b) abnormal triggering, and (c) window expansion search—shows that when a small local region of interest at a high frame rate is covered by pseudo-boundaries such as oil stains or strong reflections, causing inner-layer verification to block it, the system does not fall into deadlock and stop. The controller actively jumps out of the local interference area by dynamically expanding the window area according to the magnification, in order to obtain broader global context features, thereby re-locking the true physical boundary of the angle steel far away from the oil stains.

[0049] The window expansion operation is performed only once per action cycle. After expansion, the controller re-executes step S2 for edge feature extraction and actual cutting parameter generation within the expanded region of interest window, and sends the newly generated actual cutting parameters into the nested verification process of step S3. If the actual cutting parameters generated within the expanded window still trigger interception, the controller will continue to expand the window area by a preset ratio in the next action cycle. If the actual cutting parameters generated within the expanded window pass all nested verification processes, the controller pauses the window expansion operation and continuously updates the real-time position coordinates corresponding to the current edge features based on the displacement variable, providing dynamic tracking anchor points for subsequent window restoration.

[0050] S44 extreme condition backup monitoring As the window expands incrementally, the controller continuously acquires the area size of the expanded region of interest (ROI) window and the number of cycles for issuing compensation parameters. The controller pre-sets the global field-of-view threshold to the camera's full-frame size, which is 2048×2048 pixels. This threshold corresponds to the maximum shooting range of the visual acquisition device; beyond this range, it's impossible to acquire more effective feature information through window expansion. Simultaneously, the controller pre-sets a compensation frequency threshold of 6 times. This threshold not only matches the reasonable blind walking span of the angle iron under abnormal interference but also is compatible with the physical limits of the geometric progression of the window area's incremental expansion, ensuring normal counting before triggering the full-frame global field-of-view threshold. If the verification fails after more than 6 compensations, it indicates that the interference is not temporary, such as localized oil stains or reflections, but rather a persistent anomaly such as lens obstruction or severe deformation of the angle iron. If the window area reaches the global field of view threshold and still triggers interception, or if the number of cycles of continuously issuing compensation parameters reaches the compensation number threshold and still triggers interception, the controller will immediately stop generating compensation parameters, stop expanding the local region of interest window, trigger an audible and visual alarm signal, and issue an emergency stop braking command to the actuator and conveyor to prevent the equipment from blindly operating under continuous abnormal conditions and causing safety accidents.

[0051] S45 window restoration stability determination The controller continuously verifies the actual cutting parameters generated within the enlarged window after the window expansion is paused. When the newly generated actual cutting parameters pass all nested verification processes three times consecutively, the controller determines that the current vision system has completed self-correction, false boundary interference has been eliminated, and prepares to initiate window recovery and exit the compensation observation period. During this continuous verification observation period, the system remains in the compensation observation period and does not issue the current actual cutting parameters, but continues to issue compensation parameters to control the actuators. Simultaneously, the actual cutting parameters that pass nested verification once are synchronously updated to the historical safety parameter storage queue. The setting of three consecutive verifications effectively avoids accidental window recovery due to a single successful verification, ensuring that the corrected edge features are stable and reliable, preventing re-triggered interception after window shrinking, and also preventing window recovery lag due to excessive verification attempts, thus avoiding impact on processing efficiency.

[0052] S46 Local Interest Region Window Size Restoration The controller uses the latest real-time position coordinates acquired at the end of the continuous verification observation period as the tracking anchor point, resets the area of ​​the local region of interest window to its initial size before expansion, which is 200×100 pixels set in step S1, and returns to the high frame rate local tracking mode. At the same time, the controller stops the derivation of compensation parameters, exits the compensation observation period, and fully resumes the normal actual cutting parameter distribution process, completing the self-correcting closed loop of the entire abnormal working condition.

[0053] S5 corner cutting action execution and continuous machining cycle Precise execution of the S51 chamfering action The actuator receives the valid actual cutting parameters or compensation parameters from the controller, first driving the tool holder to adjust along the Z-axis to the target height coordinate, then adjusting around the X-axis to the target deflection angle. After the position is adjusted, it enters the ready-to-trigger state. The controller, combined with the encoder pulse data collected in real time in step S1, continuously judges whether the angle steel cutting position has moved directly below the cutting tool. When the cumulative number of pulses reaches the preset cutting trigger threshold, it immediately sends a cutting execution command to the actuator. The setting of the cutting trigger threshold is perfectly matched with the standard edge initial coordinates calibrated in the reference parameters in step S1, and also takes into account the fixed physical distance between the camera installation position and the tool installation position. Based on the displacement of 0.2mm corresponding to a single pulse, the corresponding pulse offset is calculated to ensure that the angle steel to be cut is completely aligned with the cutting tool, and there will be no problem of cutting position offset.

[0054] After receiving the corner-cutting command, the actuator first drives the tool holder to accelerate along the Y-axis until it synchronizes with the current travel speed of the angle steel. In a relatively static synchronized follow-up state, it drives the cutting tool to press down at a preset uniform speed to complete the corner-cutting action of the current angle steel. After pressing down, it quickly lifts the tool upward to reset to the preset safe position, and the follow-up mechanism drives the tool holder to quickly return to the initial standby position. Since the minimum upward physical limit in the Z-axis direction is 0mm, the safe position is set at 10mm in the Z-axis coordinate, placing it within the absolute high safe zone. This ensures that the tool is completely away from the conveyor belt below, avoiding lateral scraping or destructive impact caused by relative displacement with the continuously traveling rigid angle steel, and also prevents the tool holder position adjustment efficiency of the next cycle from being affected by an excessively long reset stroke.

[0055] Starting the S52 continuous machining cycle During the continuous conveying of angle steel, the controller maintains the complete processing flow of steps S1 to S4 as an independent high-frequency main cycle. This main cycle is strictly triggered by encoder pulse synchronization and will not be blocked or interrupted by the long and time-consuming mechanical actions of the actuator, such as pressing down to cut or lifting the tool to reset.

[0056] After the controller completes the single-cycle corner cutting action and confirms that the tool holder has returned to a safe position, the actuator immediately re-enters the standby state, seamlessly connecting with the latest actual cutting parameters or compensation parameters continuously issued by the main cycle, and initiating follow-up cutting preparation for the next corner cutting position. During the continuous processing cycle, the controller continuously uses the reference parameters of the current batch of angle steel, without repeating the retrieval operation. Only when the angle steel processing specifications are changed will the reference parameter matching and retrieval in step S12 be re-executed, ensuring the smoothness of continuous processing.

[0057] Meanwhile, the high-frequency main loop, which is not affected by mechanical movements, will continuously update the historical safety parameter storage queue according to the update rules in step S35. This ensures that the inner verification link of each action cycle can use the latest legal parameters with continuous spatial position as the verification benchmark, so that the entire processing flow forms a stable closed loop, realizes high-speed automated continuous corner cutting processing of angle steel, completely avoids visual feature discontinuity and verification deadlock caused by waiting for mechanical movements, and well balances processing safety and production continuity.

[0058] In summary, this embodiment solves the inherent contradictions of low global scanning efficiency and susceptibility to interference of fixed windows in existing technologies by using dynamically following local regions of interest windows. While ensuring high frame rates in visual processing, it accurately locks the detection area of ​​the angle steel edge. Through a double-nested parameter verification mechanism, it breaks the technical bias of unconditionally trusting visual recognition results in existing technologies. Using the physical limits of the equipment and the deformation law of the material as dual benchmarks, it blocks the path of erroneous parameters from the control architecture level, thereby avoiding equipment collision accidents caused by false boundary misdetection at the root. Through a smooth compensation and visual correction coordination mechanism under abnormal operating conditions, it abandons the passive handling mode of stopping the machine immediately when there is an abnormality in existing technologies. It completes system self-correction without interrupting production and maintains the continuous operation capability of the production line.

[0059] This embodiment also discloses a computer vision-based angle steel corner cutting device, which is used to implement the aforementioned computer vision-based angle steel corner cutting method and is suitable for use in automated continuous processing lines for angle steel. The core of the device includes a conveying mechanism, a vision acquisition device, an execution mechanism, and a controller. The components are synchronized in time and electrically interconnected to collaboratively complete the fully automated corner cutting operation of angle steel.

[0060] The conveying mechanism employs a servo motor-driven roller conveyor line to propel the angle steel to be processed along a preset horizontal path at a uniform speed. An incremental encoder is coaxially mounted on the drive end, rotating synchronously with the conveyor rollers. The encoder has a resolution of 1000 pulses per revolution, and the circumference of the conveyor roller is 200mm. A single pulse corresponds to a linear conveying displacement of 0.2mm for the angle steel. The controller accumulates the number of pulses output by the encoder per unit time, mapping the angle steel's displacement variable in real time. This provides a precise timing reference for subsequent window tracking and image synchronization triggering.

[0061] The vision acquisition device employs a global shutter area array industrial camera, fixedly mounted directly above the conveying plane of the conveyor mechanism. The lens optical axis is perpendicular to the conveying plane, and the camera's horizontal field of view is 200mm, corresponding to 2048 effective horizontal pixels. Each pixel corresponds to an actual physical size of approximately 0.0977mm. The global field of view completely covers the full horizontal width of the angle steel conveying path, ensuring that the angle steel does not exceed the shooting range during conveyance. The camera's trigger port is electrically connected to the controller's synchronization signal output, allowing it to receive trigger commands from the controller and accurately acquire images within a specified local region of interest window. This provides a clear, positionally synchronized image foundation for subsequent angle steel edge feature extraction.

[0062] The actuator is a servo-driven corner-cutting tool holder, mounted on a follower mechanism that moves synchronously along the angle steel conveying direction. The tool holder carries a carbide cutting tool and can be adjusted in height along the Z-axis perpendicular to the conveying plane, and in deflection angle around the X-axis perpendicular to the angle steel conveying direction. The tool holder's mechanical structure has fixed physical limits: the Z-axis travel range is 0mm to 150mm, and the X-axis deflection range is -15° to 15°, limiting the maximum adjustment range of the tool holder and preventing hard collisions in the mechanical structure. The actuator can receive valid or compensating parameters from the controller to precisely adjust the tool holder position and perform the corner-cutting action in sync with the angle steel's travel speed.

[0063] The controller is electrically connected to the encoder, vision acquisition device, and actuator respectively. It has a built-in preset database that can store the specification parameters, standard contour features, equipment limit parameters, and historical safety parameters of the corresponding angle steel. Its memory has an independent data sandbox that is isolated from the main control program logic. All parameters to be sent to the actuator must complete a complete verification process in the data sandbox. Only the parameters that pass the verification can be retrieved and sent by the main control program, thus isolating abnormal parameters at the software level and preventing them from directly entering the execution stage.

[0064] The controller incorporates built-in simulation and tracking logic. Based on the real-time displacement variables of the conveyor mechanism and pre-retrieved reference parameters, it simulates the real-time movement trajectory of the angle steel edge to be processed within the camera's global field of view, generating a local region of interest window whose spatial position changes dynamically. The initial window size is 200×100 pixels, and its center position always coincides with the expected coordinates of the angle steel edge. It moves synchronously with the angle steel to lock the detection area. Simultaneously, based on the encoder's real-time pulse data, it synchronously triggers the camera to complete image acquisition within the corresponding window, balancing image processing efficiency and feature locking stability. The controller can preprocess and extract edge features from the acquired local window images. It compares the extracted edge features with the parameter feature space of angle steel of the corresponding specification in the preset database, retrieves the target height coordinates and target deflection angle with the highest matching degree, and generates the actual cutting parameters for the current action cycle, providing accurate numerical basis for the tool holder action.

[0065] The controller core employs a dual-layer nested verification mechanism. The generated actual cutting parameters are sent to an isolated data sandbox for verification. First, based on the physical stroke threshold defined by the physical limit of the tool holder, extreme abnormal parameters exceeding mechanical limits are intercepted. Then, combining the material properties, thickness, and cumulative travel span of the angle steel, the continuous deformation threshold of the material is dynamically calculated. The changes in current parameters are compared with pre-stored historical safety parameters to intercept abnormal parameters exceeding the reasonable deformation range of the material. The historical safety parameters are the actual cutting parameters that have passed the most recent five consecutive complete verification processes. Only after both verifications pass will the controller issue the valid parameters to the actuator during the non-compensatory observation period, simultaneously updating the historical safety parameter queue. This prevents equipment collision risks caused by false boundary detections at the control architecture level.

[0066] When parameter verification triggers an interception, the controller discards the current abnormal parameters and does not issue any related instructions to the actuator, simultaneously entering a compensation observation period. The controller extracts pre-stored historical safety parameters, deduces compensation parameters that conform to the angle steel's movement trend, and issues them to the actuator to ensure smooth, continuous, and uninterrupted processing. At the same time, it progressively expands the local region of interest window at a preset magnification of 2x to obtain more global context information to eliminate false boundary interference. During window expansion and compensation execution, the controller monitors the window size and the number of compensation cycles in real time. If the window area reaches the camera's full-frame global field of view threshold and still triggers an interception, or if the number of consecutive compensation parameter issuance cycles reaches the preset threshold of 6 times and still fails to pass verification, the compensation and window expansion operations are immediately stopped, an audible and visual alarm is triggered, and an emergency stop command is issued to the conveyor and actuator to prevent the equipment from blindly operating in a continuously abnormal state. When the actual cutting parameters regenerated within the expanded window pass the complete nested verification three times consecutively, the controller determines that the false boundary interference has been eliminated. Using the current real-time edge position as the tracking anchor point, the window is restored to its initial size, the compensation parameter derivation is stopped, the compensation observation period is exited, the normal parameter distribution process is fully restored, and the self-correction closed loop of the abnormal working condition is completed.

[0067] The various components of this equipment work together to form a complete technical chain of timing tracking, feature extraction, security verification, fault tolerance correction, and closed-loop execution. While adapting to the processing cycle of high-speed production lines, it also takes into account the cutting accuracy and equipment operation safety, and stably realizes automated continuous processing of angle steel.

[0068] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for cutting angle steel corners based on computer vision, characterized in that, The method, applied to equipment comprising a conveying mechanism for conveying angle steel and an actuator for performing a cutting action, includes: Based on the displacement variables of the conveying mechanism and preset reference parameters, a local region of interest window is dynamically generated and tracked. The edge features of the angle steel are extracted within the local region of interest window and mapped to generate actual cutting parameters. The actual cutting parameters are input into a preset data sandbox for nested verification: if the actual cutting parameters exceed the preset physical travel threshold, or the change in the pre-stored historical safety parameters is greater than the preset material continuous deformation threshold, then an interception is triggered; otherwise, the nested verification is passed, and the actual cutting parameters are sent to the actuator to perform the corner cutting action during the non-compensation observation period. If the interception is triggered, a compensation observation period is entered. The compensation parameters are deduced based on the historical security parameters to control the actuator and expand the local region of interest window. After the actual cutting parameters regenerated within the expanded local region of interest window pass the nested verification a preset number of times, the compensation observation period is exited and the local region of interest window is restored to its original size.

2. The computer vision-based angle steel cutting method according to claim 1, characterized in that, The process of dynamically generating and tracking a local region of interest window based on the displacement variables of the conveying mechanism and preset reference parameters includes: Collect real-time encoder pulse data of the conveying mechanism and map it to the displacement variable; Based on the specifications of the angle steel to be processed, the pre-stored cutting parameters are extracted as the reference parameters; By combining the displacement variables and the reference parameters, the movement trajectory of the position to be processed is deduced, and the local region of interest window with its spatial position changing accordingly is generated; Based on the real-time encoder pulse data, image acquisition within the local region of interest window is synchronously triggered.

3. The computer vision-based angle steel cutting method according to claim 1, characterized in that, The nested verification includes, in sequence: When the actual cutting parameters do not exceed the physical travel threshold, the material properties, thickness, and cumulative travel span of the angle steel are obtained. By combining the material property data, the thickness dimension, and the cumulative travel span, the maximum reasonable deformation tolerance is dynamically calculated. The maximum reasonable deformation tolerance is assigned as the continuous deformation threshold of the material. Extract the pre-stored historical safety parameters and calculate the change range between the current actual cutting parameters and the historical safety parameters; Based on the relationship between the material's continuous deformation threshold and the magnitude of the change, the interception or the issuance of the actual cutting parameters is triggered.

4. The computer vision-based angle steel cutting method according to claim 1, characterized in that, The step of extracting the edge features of the angle steel within the local region of interest window and mapping them to generate actual cutting parameters includes: Extract the actual edge contour features of the angle steel within the local region of interest window, and use them as the edge features; The edge features are compared with the parameter feature space in the preset database to retrieve the corresponding target height coordinates and target deflection angle; The actual cutting parameters are generated by combining the target height coordinates and the target deflection angle.

5. The computer vision-based angle steel cutting method according to claim 1, characterized in that, The nested verification includes: Extract the target height coordinates and the target deflection angle from the actual cutting parameters respectively; The target height coordinates and the target deflection angle are independently compared with the physical travel threshold; When either the target height coordinate or the target deflection angle exceeds the physical travel threshold, the interception is triggered to block the issuance of commands to the actuator. When neither the target height coordinates nor the target deflection angle exceed the physical travel threshold, the current actual cutting parameters are released to extract the historical safety parameters and calculate the change range between the current actual cutting parameters and the historical safety parameters.

6. The computer vision-based angle steel cutting method according to claim 1, characterized in that, The step of controlling the actuator by deriving compensation parameters based on the historical safety parameters includes: When the interception is triggered, the current actual cutting parameters are discarded; Extract the pre-stored historical security parameters, which are the actual cutting parameters that have passed the nested verification within a preset number of historical action cycles; Calculate the slope of the change in the historical safety parameters; The compensation parameters are generated by combining the change slope with the current displacement increment determined based on the displacement variable; The compensation parameters are sent to the executing agency.

7. The computer vision-based angle steel cutting method according to claim 1, characterized in that, In the process of expanding the local region of interest window, the method further includes: Obtain the area size of the enlarged local region of interest window, and the number of cycles for continuously issuing the compensation parameters; When the area size reaches a preset global field of view threshold and the interception is still triggered, or when the number of cycles reaches a preset compensation number threshold and the interception is still triggered, the generation of the compensation parameters is stopped and the expansion of the local region of interest window is stopped. An alarm signal is triggered, and an emergency stop braking command is issued to the actuator and the conveying mechanism.

8. The computer vision-based angle steel cutting method according to claim 6, characterized in that, Expanding the local region of interest window includes: The area of ​​the local region of interest window is gradually increased by a preset magnification, and the edge features are extracted after each level of expansion to regenerate the actual cutting parameters; When the regenerated actual cutting parameters pass the nested verification, the expansion of the local region of interest window is paused, and the real-time position coordinates corresponding to the current edge feature are obtained.

9. The computer vision-based angle steel cutting method according to claim 8, characterized in that, Restoring the local region of interest window to its original size includes: Using the latest acquired real-time location coordinates as the tracking anchor point, the area of ​​the local region of interest window is reset to its original size. Stop calculating the compensation parameters, update the verified actual cutting parameters to the historical safety parameters, and resume issuing the actual cutting parameters.

10. A computer vision-based angle steel cutting device, characterized in that, For implementing the method of any one of claims 1-9, the device comprises a conveying mechanism for conveying angle steel, an actuator for performing a cutting action, and a controller, the controller comprising: The simulation and tracking module is used to dynamically generate and track a local region of interest window based on the displacement variables of the conveying mechanism and preset reference parameters. The extraction and mapping module is used to extract the edge features of the angle steel within the local region of interest window and map them to generate actual cutting parameters. The nested verification module is used to input the actual cutting parameters into a preset data sandbox for nested verification: if the actual cutting parameters exceed the preset physical travel threshold, or the change range of the pre-stored historical safety parameters is greater than the preset material continuous deformation threshold, then an interception is triggered; otherwise, the nested verification is deemed to pass, and the actual cutting parameters are sent to the actuator to perform the corner cutting action during the non-compensation observation period. The collaborative correction module is used to enter a compensatory observation period if the interception is triggered, control the execution mechanism based on the historical security parameters to deduce the compensatory parameters, and expand the local region of interest window; after the actual cutting parameters regenerated in the expanded local region of interest window pass the nested verification a preset number of times, the module exits the compensatory observation period and restores the local region of interest window to its original size.

Images