A machine vision-based gate inclination angle detection method

Abstract

The application discloses a gate inclination angle detection method based on machine vision and belongs to the technical field of machine vision detection. The method obtains a gate operation image, corrects the image by using an environment self-adaptive correction network, the network comprises an illumination self-adaptive equalization sub-module, a motion blur self-adaptive denoising sub-module and a dust blur restoration sub-module; inputs the corrected image into a target detection model with multi-scale feature enhancement, detects the detection frame coordinates of the pulleys at both ends of the gate, the model comprises a feature extraction network with a dynamic receptive field adjustment mechanism; performs jump identification and filter correction on the detection frame coordinates by using an intelligent self-adaptive filter module, and outputs stable coordinates; and finally, the gate inclination angle is calculated based on the stable coordinates according to the geometric principle. The application solves the technical problem of low gate inclination angle detection precision in a complex environment, realizes high-precision gate inclination angle detection through multi-level image correction, intelligent target detection and self-adaptive filter processing, and is suitable for real-time monitoring of the gate operation state in water conservancy projects.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, and specifically to a gate tilt angle detection method based on machine vision. Background Technology

[0002] Gate tilt angle detection is a key technology for monitoring the operation of important infrastructure such as water conservancy projects, lock systems, and hydropower stations, and is of great significance for ensuring the safe operation of equipment and preventing accidents. Traditional gate tilt angle detection mainly uses contact sensor solutions, including physical measurement devices such as tilt sensors and accelerometers. However, these contact detection solutions face many challenges in the complex environment of water conservancy projects, including mechanical vibration interference, high humidity corrosion, and equipment aging, leading to decreased measurement accuracy and increased maintenance costs.

[0003] With the rapid development of machine vision technology, non-contact detection methods based on image processing have gradually become a research hotspot. Existing machine vision detection technologies have been widely applied in the field of industrial inspection. For example, Chinese patent CN118918364B discloses a visual recognition method for missing bolts in pipe racks under robot-driven shooting conditions. This method uses a median filter and a DnCNN combined model to remove image noise and uses an improved Yolov8 model for target detection. Chinese patent CN111260615A discloses a UAV-based method for detecting apparent defects in bridges based on the fusion of laser and machine vision. It identifies defects in target images and calculates defect size parameters through a defect recognition neural network. Chinese patent CN114387262A describes a machine vision-based nut positioning detection method that uses the Retinex algorithm to enhance the image and reduce the impact of complex environments on recognition accuracy. Chinese patent CN120259210B discloses a machine vision-based method for identifying defects in wire ropes of port cranes. It uses a Kalman filter algorithm to predict motion states and constructs a fuzzy kernel function for image sharpening. Chinese patent CN120088583B relates to a machine vision-based FOD recognition system, which achieves target recognition in complex environments through multimodal image acquisition and dynamic environment adaptive preprocessing modules.

[0004] Despite advancements in machine vision inspection, significant technical limitations remain in the specific application of gate tilt angle detection. First, the complex environment of hydraulic engineering projects severely degrades image quality, due to factors such as strong lighting variations, dust contamination, and motion blur caused by mechanical vibrations. Existing image preprocessing techniques struggle to effectively address the combined effects of multiple degrading factors. Second, the scale of gate pulley targets varies greatly, and the fixed receptive field of traditional target detection models cannot adaptively match pulley targets of different sizes, resulting in insufficient detection accuracy. Third, coordinate jumps during detection significantly impact measurement stability; existing filtering algorithms cannot effectively distinguish between real motion changes and noise interference, hindering accurate coordinate tracking. Finally, existing solutions lack end-to-end collaborative optimization design; each technical component operates independently, preventing overall performance optimization and limiting further improvements in detection accuracy and real-time performance. Summary of the Invention

[0005] To address the technical challenges of existing gate tilt angle detection technologies, such as poor environmental adaptability of contact-based detection schemes, severe image degradation in complex environments, insufficient adaptability to pulley targets across all scales, measurement result fluctuations caused by changes in detection coordinates, and the fragmentation of existing machine vision solutions, this paper proposes a machine vision-based gate tilt angle detection method. This method aims to achieve tilt angle measurement accuracy of ±0.1°, significantly enhanced environmental adaptability, significantly improved stability, outstanding adaptability across all scales, and excellent real-time performance.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: A gate tilt angle detection method based on machine vision is provided, comprising: acquiring a gate operation image; inputting the gate operation image into an environment adaptive correction network to perform degradation correction on the image and outputting a corrected image; the environment adaptive correction network includes at least an illumination adaptive equalization submodule for correcting image illumination distortion, a motion blur adaptive denoising submodule for removing image motion blur, and a dust blur restoration submodule for restoring image dust blur; inputting the corrected image into a multi-scale feature enhancement target detection model to detect and output the detection box coordinates of the pulleys at both ends of the gate; the target detection model includes at least a feature extraction network with a dynamic receptive field adjustment mechanism to adapt to pulley targets of different sizes; using an intelligent adaptive filtering module to perform jump recognition and filtering correction on the detection box coordinates and output stable coordinates; and calculating the gate tilt angle based on geometric principles according to the stable coordinates.

[0007] Preferably, the adaptive illumination equalization submodule decomposes the input image I(x,y) into illumination component L(x,y) and reflection component R(x,y) based on Retinex theory, dynamically calculates the illumination adjustment factor λL according to the local information of the illumination component, and corrects the illumination component to obtain the corrected illumination component. , where k is the dynamically adjusted kernel size; the motion blur adaptive denoising submodule estimates the motion blur kernel K based on image gradient information through a blur kernel estimation network, and performs deblurring using an iterative algorithm based on the alternating direction multiplier method; the dust blur restoration module identifies dust attachment areas in the image through a dust region semantic segmentation network, evaluates its blur level, and performs feature restoration using a corresponding adaptive restoration strategy based on the level.

[0008] Furthermore, the dynamic receptive field adjustment mechanism dynamically adjusts the dilation rate r of the dilated convolution in the feature extraction network based on the size information of the detected pulley target, so as to match the receptive field with the pulley size, where λs is the scale perception factor and rs is the ratio of the pulley target size to the image size.

[0009] Optionally, the target detection model further includes a small pulley feature enhancement unit, which is used to perform super-resolution reconstruction, detail enhancement and feature filtering on the feature map of the small pulley, so as to improve the detection accuracy of the small pulley.

[0010] Preferably, the object detection model is trained using a scenario-based training optimization strategy, the optimization strategy including the use of a scenario-adaptive hybrid loss function. in For classifying losses, To regress the loss, For target confidence loss, For scale adaptation loss, , , , Its corresponding weighting coefficient.

[0011] Furthermore, the intelligent adaptive filtering module identifies coordinate jumps in real time based on the time-series characteristics of the detection box coordinates and the physical space constraints of the gate movement, and distinguishes the types of jumps. When a jump is identified as no jump or a true coordinate change, Kalman filtering is used for coordinate tracking and prediction. When a jump is identified as a noisy jump, adaptive weighted moving average filtering is used to assign lower weights to the current frame coordinates where the jump occurs and higher weights to the historical stable frame coordinates in order to suppress the jump.

[0012] Optionally, after calculating the gate tilt angle, the method further includes an adaptive tilt angle error compensation submodule based on a preset error prediction model.

[0013] The calculated inclination angle The values ​​are uncorrected tilt measurements, where T is the ambient temperature, dist is the lens distortion coefficient, and a, b, and c are learnable compensation coefficients.

[0014] The beneficial effects of this invention are that non-contact measurement has significant advantages: it does not require direct contact with the gate, avoiding the defects of contact sensors that are susceptible to environmental corrosion and vibration interference; it is flexible in installation and deployment, does not affect the normal operation of the gate, and improves the safety and applicability of the detection. Extremely adaptable to the environment: Through adaptive illumination distortion correction, de-blurring fusion and attention enhancement design, it effectively copes with harsh environmental interference such as direct sunlight, backlight, dust, and vibration. The average accuracy (mAP) of pulley detection is increased to over 98.5%, and the false negative rate is reduced to below 1%. High detection accuracy and stability: The multi-scale feature enhancement network ensures accurate detection of full-size pulleys, and the improved filtering algorithm reduces the standard deviation of the center point coordinates to within 0.5 pixels. The tilt angle measurement error is controlled within ±0.1°, effectively suppressing result fluctuations and truly reflecting the gate's operating posture. Its engineering application value is outstanding: it can be widely adapted to the monitoring needs of different types of gates in water conservancy, shipping, municipal and other fields, providing high-precision data support for intelligent operation and maintenance and fault early warning of gates, and helping to realize the digital and intelligent management of engineering facilities. Attached Figure Description

[0015] Figure 1 This is a diagram of the overall system architecture of this solution; Figure 2 This is a diagram of the environment adaptive correction network structure for this scheme; Figure 3 This is a schematic diagram of the multi-scale feature enhancement target detection model of this scheme; Figure 4 This is a schematic diagram of the intelligent adaptive filtering module in this solution. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] Example 1 like Figure 1 As shown, this invention provides a gate tilt angle detection method based on machine vision. This method achieves accurate detection of gate tilt angle in complex environments through an end-to-end environment adaptive correction network, a multi-scale feature-enhanced target detection model, an intelligent adaptive filtering module, and geometric principle calculation.

[0018] S1: Acquire gate operation images: Real-time images of the gate in operation are acquired using a high-resolution industrial camera. The image resolution is 1920×1080 pixels, and the acquisition frame rate is 30fps. The complex environment of the water conservancy site is taken into account during the acquisition process, including factors such as drastic changes in lighting, mechanical vibration, and dust interference.

[0019] S2: Environmental Adaptive Correction Processing: The acquired gate operation image is input into the environmental adaptive correction network for degradation correction. This network includes three core sub-modules: such as Figure 2 As shown, the illumination adaptive equalization submodule decomposes the input image I(x,y) into illumination component L(x,y) and reflection component R(x,y) based on Retinex theory. First, it calculates the local brightness mean μL, brightness variance σL, and contrast C, where the formulas are as follows:

[0020] Where H and W are the image height and width, and GaussianBlur is the Gaussian blur (kernel size k is determined by...). Dynamic adjustment The larger k is, the larger it is. For the corrected illumination components; Reflection component enhancement and restoration: The adaptive contrast enhancement (CLAHE) algorithm is applied to the reflection component R(x,y), combined with the edge protection mechanism, to enhance the pulley edge contour features while suppressing noise amplification, and finally obtain the corrected image.

[0021] The motion blur adaptive denoising submodule estimates the blur kernel K of the motion blur based on image gradient information through a blur kernel estimation network. The gradient information is calculated as follows: The fuzzy kernel is estimated using a lightweight U-Net CNN network: An iterative algorithm based on the alternating direction multiplier method is used for deblurring. The iterative formula is as follows: , where * represents the convolution operation, TV(I) is the total variational regularization term used to suppress image smoothing distortion, and λ is a regularization parameter dynamically adjusted according to the blur kernel size. Iteration stops when the structural similarity SSIM of two adjacent deblurred images is ≥ 0.98.

[0022] The dust-blurred complex atomic module first identifies the dust adhesion area in the image through a lightweight dust area semantic segmentation network based on MobileNetV4, with a segmentation accuracy rate of over 97.5%. The blur degree level is evaluated based on the gray variance σd of the dust area: slightly blurred (σd ≤ 0.8), moderately blurred (0.8 < σd ≤ 1.5), and severely blurred (σd > 1.5). Corresponding adaptive restoration strategies are adopted for different levels: adaptive Gaussian filtering restoration for slight blur, restoration algorithm based on sparse representation for moderate blur, and GAN generative restoration for severe blur.

[0023] S3: Object detection with multi-scale feature enhancement: Input the corrected image into the object detection model with multi-scale feature enhancement, and detect and output the detection box coordinates of the pulleys at both ends of the gate.

[0024] As Figure 3 shown, the object detection model includes a feature extraction network with a dynamic receptive field adjustment mechanism for adapting to pulley targets of different sizes. The dynamic receptive field adjustment mechanism quickly locates the pulley target through a pulley scale perception unit, and calculates the ratio rs = s / S of the pixel area s of the circumscribed rectangle of the pulley to the image area S. Based on the detected pulley target size information, the dilation rate r of the dilated convolution in the feature extraction network is dynamically adjusted, and the decision formula is: r = round(λs·f(rs)), where f(rs) determines the base dilation rate according to rs: 5 when rs ≤ 0.05, 4 when 0.05 < rs ≤ 0.1, 3 when 0.1 < rs ≤ 0.2, 2 when 0.2 < rs ≤ 0.3, and 1 when rs > 0.3. λs is the scale perception factor, λs = 1.2 for small-sized pulleys, λs = 1.0 for medium-sized pulleys, and λs = 0.8 for large-sized pulleys.

[0025] The object detection model also includes a small-sized pulley feature enhancement unit, which uses a lightweight ESPCN network to perform super-resolution reconstruction on the feature map of the small-sized pulley, doubling the resolution. The key detail features such as the pulley edge contour and hub holes are enhanced through an improved sharpening convolution kernel, and at the same time, L1 regularization is used to suppress noise amplification. The core features of the small-sized pulley are screened out through feature saliency detection, suppressing background noise interference, and improving the detection accuracy of the small-sized pulley by over 25%.

[0026] The object detection model is trained through a scene-adapted training optimization strategy, using a scene-adaptive hybrid loss function: Wherein, Lcls is the classification loss using an improved Focal loss, which introduces a scene weight factor to assign higher weights to pulley samples in small-sized pulleys and dusty / blurred scenes; Lreg is the regression loss using a weighted fusion of CIoU loss and GIOU loss, combined with circular IoU constraints to improve the localization accuracy of pulley detection boxes; Lobj is the target confidence loss using an improved BCEWithLogitsLoss, which introduces spatial attention weights to strengthen the confidence response of the pulley region; and Lscale is the scale adaptation loss, which constrains the consistency of feature extraction for pulleys of different scales. The weight coefficients were determined through Bayesian optimization: Wcls=0.3, Wreg=0.4, Wobj=0.2, Wscale=0.1.

[0027] S4: Intelligent Adaptive Filtering Processing: The intelligent adaptive filtering module performs jump recognition and filtering correction on the coordinates of the detection box, and outputs stable coordinates.

[0028] like Figure 4 As shown, this module identifies coordinate jumps and distinguishes jump types in real time based on the time-series characteristics of the detection box coordinates and the physical space constraints of the gate movement. First, it calculates the coordinates of the center point of the pulley detection box in the current frame. The average of the coordinates of the previous 3 frames The Euclidean distance d is given by the formula: Based on the preset gate operating speed v ≤ 0.5 pixels / frame, the coordinate change rate vt = d / Δt is calculated. When vt > v and d > 0.3 pixels, it is determined to be a coordinate jump. The jump type is distinguished by pulley feature similarity matching: if the SSIM of the pulley features between the current frame and the previous frame is ≥ 0.95, it is a true coordinate change; otherwise, it is a noise jump.

[0029] When a no-jump or true coordinate change is identified, Kalman filtering is used for coordinate tracking and prediction. The state prediction equation of Kalman filtering is: Where Pt is the covariance matrix, H is the observation matrix, R is the observation noise variance, A is the state transition matrix, and Q is the process noise variance.

[0030] When a noise jump is identified, an adaptive weighted moving average filter is used. The coordinates of the current frame where the jump occurs are given a lower weight (0.1), and the coordinates of historical stable frames are given a higher weight (0.3). The weighted average coordinate is calculated as: x̂t=∑(i=0 to 3)wixt-i, where ∑wi=1, to suppress the impact of the jump.

[0031] S5: Inclination Angle Calculation and Error Compensation: The gate inclination angle is calculated based on stable coordinates and geometric principles. Using the stable coordinates of the pulleys at both ends of the gate, the angle between the gate and the horizontal plane is calculated through geometric relationships.

[0032] The calculated tilt angle is compensated by an adaptive tilt angle error compensation submodule. This submodule, based on a preset error prediction model: Δθ = a·T + b·dist + c, compensates for the calculated tilt angle, obtaining the corrected tilt angle θcorr = θmeas - Δθ. Here, θmeas is the uncorrected tilt angle measurement, T is the ambient temperature, dist is the lens distortion coefficient, and a, b, and c are learnable compensation coefficients. The compensation coefficients are obtained through real-time calibration using a standard tilt angle calibration plate, and a time-series smoothing constraint is introduced to ensure the stability of the compensated tilt angle result.

[0033] Through the above processing steps, this method can accurately detect the gate inclination angle in complex hydraulic environments, with the inclination angle measurement error controlled within ±0.1°, meeting the high-precision monitoring requirements of hydraulic engineering projects. This method effectively solves the technical problems of low detection accuracy and poor stability of traditional detection methods in harsh environments such as drastic changes in lighting, vibration ambiguity, and dust interference.

[0034] Example 2 A machine vision-based gate tilt angle detection system includes an image acquisition unit, an environment adaptive correction unit, a target detection unit, an intelligent adaptive filtering unit, and a tilt angle calculation unit.

[0035] The image acquisition unit employs a high-resolution industrial camera to acquire image data in real time during the gate's operation. The camera is configured with a resolution of 2048×1536 pixels and a frame rate of 30fps, ensuring clear images of the pulleys during gate movement. The image acquisition unit transmits the acquired raw image data to the subsequent processing unit via an Ethernet interface.

[0036] The environmental adaptive correction unit receives the original image data and performs degradation correction processing on the image to address the complex environmental conditions at the water conservancy site. This unit comprises three core components: an illumination adaptive equalization subunit, a motion blur adaptive denoising subunit, and a dust blur complex atom unit. The illumination adaptive equalization subunit separates the illumination and reflection components based on Retinex theory. It dynamically adjusts the illumination correction factor by calculating illumination evaluation indicators such as local brightness mean, brightness variance, and contrast. When the local brightness mean is less than 30, it is considered a low-light environment, and the correction factor is set to 1.8; when the local brightness mean is greater than 220, it is considered an overexposed environment, and the correction factor is set to 0.6; under normal illumination conditions, the correction factor is 1.0. The motion blur adaptive denoising subunit dynamically estimates the size and orientation of the image blur kernel using a convolutional neural network, performs iterative deblurring using an alternating direction multiplier method, and combines non-local mean filtering to suppress noise. The dust blurring complexation unit uses a lightweight segmentation network based on MobileNetV4 to identify dust-attached regions. Based on the gray-level variance of the dust region, the blurring level is divided into three levels: mild, moderate, and severe. Adaptive Gaussian filtering, sparse representation restoration algorithm, and GAN generative restoration are then applied to process these levels, respectively. The environment adaptive correction unit outputs a high-quality image after degradation correction, providing clear input data for subsequent object detection.

[0037] The target detection unit is equipped with a multi-scale feature-enhanced target detection model to accurately detect the positions of the pulleys at both ends of the gate from the corrected image. The model's feature extraction network has a dynamic receptive field adjustment mechanism, which can adjust the convolution dilation rate in real time based on the pulley size information in the input image. The system first quickly locates the pulley target through the initial detection branch, calculates the ratio of the pulley's bounding rectangle pixel area to the total image area, and constructs a scale-aware factor by combining prior information about the gate model. For small pulleys (area ≤ 50 pixels), the dilation rate is set to 4-5 to expand the receptive field; for large pulleys (area > 120 pixels), the dilation rate is set to 1-2 to focus on local details. The feature extraction network employs a bidirectional cross-scale feature fusion mechanism, achieving deep fusion of features at different scales through top-down and bottom-up feature transfer. During the fusion process, a channel-spatial dual attention module is embedded to dynamically allocate the contribution of features at different scales.

[0038] The object detection model also includes a small-size pulley feature enhancement unit, specifically designed for small pulleys with a pixel area ≤ 50 pixels. This unit uses a lightweight ESPCN network to double the resolution of the small-size pulley feature map, enhances key details such as pulley edge contours and hub holes through improved sharpening convolution kernels, and employs L1 regularization to suppress noise amplification. The feature enhancement unit filters out the core features of the small-size pulley through feature saliency detection, effectively suppressing background noise interference. The object detection unit ultimately outputs the precise detection box coordinates of the pulleys at both ends of the gate.

[0039] The intelligent adaptive filtering unit receives the detection box coordinates output by the target detection unit and performs jump identification and filtering correction on the coordinate data. Based on the time-series characteristics of the detection box coordinates and the physical spatial constraints of the gate's movement, this unit analyzes the coordinate change patterns in real time. The system calculates the Euclidean distance between the center point coordinates of the pulley detection box in the current frame and the average coordinates of the previous three frames, and combines this with a preset gate running speed constraint (≤0.5 pixels / frame) to determine whether a coordinate jump has occurred. When the coordinate change rate exceeds a threshold and the distance is greater than 0.3 pixels, the system further distinguishes between genuine coordinate changes and noise jumps through pulley feature similarity matching.

[0040] The intelligent adaptive filtering unit employs different filtering strategies based on the jump identification results. When a jump is identified as non-jump or a change in true coordinates, the system uses Kalman filtering for coordinate tracking and prediction, dynamically adjusting the process noise variance and observation noise variance, and adaptively optimizing the filtering gain based on the residual between the predicted and observed values. When a noisy jump is identified, the system uses adaptive weighted moving average filtering, assigning low weights (0.1-0.3) to the jump frame coordinates and high weights (0.7-0.9) to the normal coordinates of the first three frames, obtaining a stable coordinate output through weighted averaging. The intelligent adaptive filtering unit outputs stable coordinate data after filtering correction.

[0041] The tilt angle calculation unit receives stable coordinates output from the intelligent adaptive filtering unit and calculates the gate tilt angle based on geometric principles. The system constructs a straight line equation using the stable coordinates of the pulleys at both ends of the gate, and obtains the gate tilt angle value by calculating the angle between this line and the horizontal baseline. An adaptive error compensation mechanism is introduced during the tilt angle calculation process. Based on the on-site temperature data collected by the temperature sensor and the lens distortion coefficient obtained from camera calibration, an error prediction model is constructed to compensate for system errors in real time. The compensated tilt angle measurement result has a high accuracy of ±0.1°, fully meeting the monitoring requirements of water conservancy projects.

[0042] The entire system achieves accurate and stable detection of gate tilt angle through the coordinated work of image acquisition, environmental adaptive correction, target detection, intelligent filtering, and tilt angle calculation, effectively solving the technical challenges of tilt angle monitoring in complex hydraulic environments.

[0043] The above description is merely a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations that are directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A gate tilt angle detection method based on machine vision, characterized in that, include: Acquire images of the gate in operation; The gate operation image is input into an environment adaptive correction network to perform degradation correction on the image and output the corrected image. The environment adaptive correction network includes at least an illumination adaptive equalization submodule for correcting image illumination distortion, a motion blur adaptive denoising submodule for removing image motion blur, and a dust blur restoration submodule for restoring image dust blur. The corrected image is input into a multi-scale feature enhancement target detection model to detect and output the detection box coordinates of the pulleys at both ends of the gate. The target detection model includes at least a feature extraction network with a dynamic receptive field adjustment mechanism to adapt to pulley targets of different sizes. An intelligent adaptive filtering module is used to identify and filter the coordinates of the detection box to output stable coordinates. The gate tilt angle is calculated based on the stable coordinates and geometric principles.

2. The method according to claim 1, characterized in that, The adaptive illumination equalization submodule is specifically used to: decompose the input image I(x,y) into illumination component L(x,y) and reflection component R(x,y) based on Retinex theory; and dynamically calculate the illumination adjustment factor based on the local information of the illumination components. The illumination components are then corrected to obtain the corrected illumination components: Where k is the dynamically adjusted kernel size; after enhancing the reflection component R(x,y), based on and the image after reconstruction and correction of the enhanced reflection component ; The motion blur adaptive denoising submodule is specifically used for: estimating the blur kernel K of motion blur based on image gradient information using a blur kernel estimation network; and performing deblurring using an iterative algorithm based on the alternating direction multiplier method, wherein the iterative formula is: In the formula This represents the image after deblurring in the (k+1)th round, where * denotes the convolution operation. λ is the total variational regularization term, and λ is the regularization parameter. The dust blur restoration module is specifically used to: identify dust-attached areas in an image through a dust region semantic segmentation network, evaluate the blur level, and perform feature restoration using a corresponding adaptive restoration strategy based on the level.

3. The method according to claim 1, characterized in that, The dynamic receptive field adjustment mechanism includes: dynamically adjusting the dilation rate *r* of the dilated convolution in the feature extraction network based on the detected pulley size information, so that the receptive field matches the pulley size; the decision formula is as follows: in For scale-sensing factors This is the ratio of the pulley target size to the image size.

4. The method according to claim 1, characterized in that, The target detection model also includes a small pulley feature enhancement unit, which is used to perform super-resolution reconstruction, detail enhancement and feature filtering on the feature map of the small pulley to improve the detection accuracy of the small pulley.

5. The method according to claim 1, characterized in that, The object detection model is trained using a scenario-based training optimization strategy, which includes using a scenario-adaptive hybrid loss function. The loss function is: in For classifying losses, To regress the loss, For target confidence loss, For scale adaptation loss, , , , Its corresponding weighting coefficient.

6. The method according to claim 1, characterized in that, The intelligent adaptive filtering module is specifically used for: identifying coordinate jumps in real time based on the time-series characteristics of the detection box coordinates and the physical space constraints of the gate movement, and distinguishing the types of the jumps; when identified as no jump or a true coordinate change, Kalman filtering is used for coordinate tracking and prediction, and the calculation method of its filter gain is as follows: in Let H be the covariance matrix, H be the observation matrix, and R be the observation noise variance. When a noise jump is identified, an adaptive weighted moving average filter is used to assign a lower weight to the current frame coordinates where the jump occurs and a higher weight to the historical stable frame coordinates in order to suppress the jump.

7. The method according to claim 1, characterized in that, After calculating the gate tilt angle, the method further includes: using a tilt angle error adaptive compensation submodule based on a preset error prediction model. The calculated inclination angle Compensation is performed to obtain the corrected camber angle. in The values ​​are uncorrected tilt measurements, where T is the ambient temperature, dist is the lens distortion coefficient, and a, b, and c are learnable compensation coefficients.

8. A gate tilt angle detection system based on machine vision, characterized in that, include: The image acquisition unit is used to acquire images of the gate in operation; An environment adaptive correction unit is used to perform degradation correction on the image and output the corrected image. The environment adaptive correction unit includes at least an illumination adaptive equalization subunit, a motion blur adaptive denoising subunit, and a dust blur complex atom unit. The target detection unit is equipped with a multi-scale feature-enhanced target detection model, which is used to detect and output the detection box coordinates of the pulleys at both ends of the gate from the corrected image. The target detection model includes at least a feature extraction network with a dynamic receptive field adjustment mechanism. The intelligent adaptive filtering unit is used to identify and filter the coordinate jumps of the detection box and output stable coordinates. The tilt angle calculation unit is used to calculate the gate tilt angle based on the stable coordinates and geometric principles.

9. The system according to claim 8, characterized in that, The target detection model also includes a small pulley feature enhancement unit, which is used to perform super-resolution reconstruction, detail enhancement and feature selection on the feature map of the small pulley.

10. The system according to claim 8, characterized in that, The intelligent adaptive filtering unit is specifically used to: identify coordinate jumps and distinguish their types in real time based on the time series characteristics of the detection box coordinates and the physical space constraints of the gate motion; When the jump is identified as a non-jump or a true coordinate change, Kalman filtering is used for coordinate tracking and prediction; when the jump is identified as a noisy jump, adaptive weighted moving average filtering is used to assign low weights to the coordinates of the jump frame in order to suppress the jump.

Images