A 3D vision track contour detection system resistant to sunlight interference

Abstract

This invention relates to the field of rail transit inspection technology, and specifically discloses a 3D visual track contour detection system resistant to sunlight interference. The system includes a four-dimensional situational awareness module that collects light intensity, light direction, sunlight interference type, and track surface condition type; a three-factor coupling analysis module that calculates first to third coupling coefficients between sunlight incident angle and track surface reflection characteristics, sunlight incident angle and camera imaging posture, and track surface reflection characteristics and camera imaging posture; a multimodal imaging and hierarchical anti-interference module that determines the dominant interference factor and interference intensity level based on the coupling coefficients, matches and executes collaborative anti-interference strategies from a preset anti-interference strategy library, and obtains an anti-interference enhanced image; and a three-dimensional reconstruction and contour output module that performs multimodal fusion point cloud reconstruction and outputs the track contour detection result. This invention achieves high-precision detection of track contours under complex lighting conditions through an active anti-interference architecture of environmental perception, coupling analysis, and dynamic adaptation.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a 3D visual track contour detection system resistant to sunlight interference. Background Technology

[0002] With the rapid development of rail transit, the detection of geometric parameters of rail profiles is crucial for ensuring train operation safety. Existing rail profile detection systems mostly employ non-contact measurement methods based on structured light vision. This involves projecting structured light onto the rail surface using a laser, and then using a camera to capture the structured light image and reconstruct the rail's three-dimensional profile. However, in actual outdoor inspection environments, sunlight interference is a major factor affecting detection accuracy. Existing technologies, such as CN117246374A, employ a passive redundancy scheme with a symmetrical dual-camera layout. Two cameras are symmetrically installed around a line laser, ensuring that at least one camera can acquire valid data regardless of the direction of sunlight. This symmetrical redundancy structure addresses the basic sunlight interference problem to some extent, but it is essentially a passive fault-tolerance mechanism relying on hardware redundancy.

[0003] While the aforementioned passive redundancy scheme can cope with conventional sunlight interference scenarios, it has fundamental technical defects under complex lighting conditions: the scheme cannot actively sense changes in ambient light and differences in track surface conditions, nor can it dynamically adjust the imaging strategy according to different types and intensities of interference. As a result, in extreme scenarios such as strong backlight at noon, sunlight reflection after rain or snow, and sudden changes in light intensity at tunnel entrances and exits, both cameras may fail due to glare or reflected light interference. At the same time, differences in surface conditions such as rust on the rails, oil stains, and snow residue can significantly alter the sunlight reflection characteristics. The passive redundancy scheme cannot adaptively adjust to different surface conditions, resulting in large fluctuations in detection accuracy and making it difficult to meet the requirements for high-precision detection in all weather conditions.

[0004] Therefore, this invention proposes a 3D visual track contour detection system resistant to sunlight interference. Summary of the Invention

[0005] This invention provides a 3D visual track contour detection system resistant to sunlight interference. By constructing an active anti-interference architecture that integrates environmental perception, coupled analysis, and dynamic adaptation, it overcomes the fundamental defects of existing passive redundancy schemes that cannot perceive environmental changes and track surface conditions, and cannot dynamically optimize anti-interference strategies. This enables stable all-weather, high-precision track contour detection under extreme lighting and complex surface conditions.

[0006] This invention provides a 3D visual track contour detection system resistant to sunlight interference, comprising: The four-dimensional situational awareness module includes a light sensor array, a spectral analysis unit, and a track surface state recognition unit. The light sensor array is used to collect ambient light intensity and direction in real time. The spectral analysis unit is used to identify the type of sunlight interference based on the spectral distribution. The track surface state recognition unit is used to perform image recognition based on image data from a visible light image sensor, extract track surface texture features, and identify the type of track surface state. The three-factor coupling analysis module is used to receive data on light intensity, light direction, sunlight interference type, and track surface state type output by the four-dimensional situational awareness module. It calculates the first coupling coefficient between sunlight incident angle and track surface reflection characteristics, the second coupling coefficient between sunlight incident angle and camera imaging attitude, and the third coupling coefficient between track surface reflection characteristics and camera imaging attitude through a pre-trained coupling analysis model. The multimodal imaging and hierarchical anti-interference module includes a visible light binocular camera unit, a near-infrared laser module, an adjustable polarization filter unit, and an anti-interference strategy decision unit. The anti-interference strategy decision unit is used to determine the dominant interference factor and the interference intensity level based on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient output by the three-factor coupling analysis module. According to the dominant interference factor and the interference intensity level, it matches the corresponding collaborative anti-interference strategy from the preset anti-interference strategy library and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data. The 3D reconstruction and contour output module is used to perform multimodal fusion point cloud reconstruction based on anti-interference enhanced visible light image data and near-infrared image data, extract the 3D contour of the rail surface and output the track contour detection results.

[0007] Furthermore, the track surface condition recognition unit includes: The texture feature extraction subunit is used to extract the texture features of the track surface based on the image data of the visible light image sensor, and to calculate the texture roughness of the rusted state, the gloss feature value of the oil-covered state, and the edge sharpness feature value of the snow-residual state. The multi-dimensional quantization sub-unit is used to input the texture roughness of the rusted state, the gloss feature value of the oil-covered state, and the edge sharpness feature value of the snow-residual state into the pre-trained surface state quantization model to obtain the rusted state coefficient, the oil-covered state coefficient, and the snow-residual state coefficient. The surface state fusion subunit is used to calculate the confidence distribution of the fused track surface state based on the rust state coefficient, oil stain coverage state coefficient, and snow residue state coefficient using the evidence theory method. The surface state type with the highest confidence is output as the track surface state type data to the three-factor coupled analysis module.

[0008] Furthermore, the three-factor coupling analysis module includes: The historical data organization subunit is used to organize the lighting conditions, track surface conditions and imaging quality recorded in the historical detection data in a multi-dimensional manner according to the monitoring time point, monitoring section location and feature parameter type, to form a time series dataset of lighting conditions, a spatial distribution dataset of track surface conditions and a set of imaging quality feature parameters. The coupled graph neural network constructs a subunit, which is used to establish a three-node dynamic coupled graph structure based on the time series dataset of illumination conditions, the spatial distribution dataset of track surface state, and the set of imaging quality feature parameters. The graph neural network is used to establish a three-node dynamic coupled graph structure with the sunlight incident angle as the first node, the track surface reflection characteristics as the second node, and the camera imaging attitude as the third node. The connection weights in the dynamic coupled graph structure are dynamically updated according to the real-time monitoring data. The multi-scale coupled analytical subunit is used to learn the short-term fluctuation coupling relationship and long-term trend coupling relationship between the first node and the second node at different time scales through a dynamic coupled graph structure. The short-term fluctuation coupling relationship and the long-term trend coupling relationship are weighted and fused to obtain the first coupling coefficient. The dynamic coupled graph structure is used to learn the static geometric coupling relationship and dynamic motion coupling relationship between the first node and the third node at different spatial scales through a dynamic coupled graph structure. The static geometric coupling relationship and the dynamic motion coupling relationship are weighted and fused to obtain the second coupling coefficient. The dynamic coupled graph structure is used to learn the material property coupling relationship and imaging physics coupling relationship between the second node and the third node at different physical scales through a dynamic coupled graph structure. The material property coupling relationship and the imaging physics coupling relationship are weighted and fused to obtain the third coupling coefficient. The first coupling coefficient, the second coupling coefficient, and the third coupling coefficient are output to the anti-interference strategy decision unit.

[0009] Furthermore, the anti-interference strategy decision-making unit includes: The coupling coefficient time series analysis subunit is used to perform time series analysis on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient, and to calculate the mean, variance, rate of change, and fluctuation frequency of each coupling coefficient within a preset time window. The interference intensity dynamic grading subunit is used to dynamically divide the coupling intensity of the three factors into three levels: mild interference level, moderate interference level, and severe interference level based on the current value, variance, and rate of change of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient using fuzzy logic. The grading threshold is adaptively adjusted according to the statistical distribution of the coupling coefficient. The hierarchical trigger condition mapping subunit is used to establish the mapping relationship between mild interference level and single anti-interference measure trigger condition, moderate interference level and dual-channel collaborative anti-interference measure trigger condition, and severe interference level and three-channel collaborative anti-interference measure trigger condition. The mapping relationship is stored in the storage space accessible by the strategy matching process, and the interference intensity level is output to the strategy matching process of the anti-interference strategy decision unit.

[0010] Furthermore, the anti-interference strategy decision-making unit includes: The coupling coefficient contribution calculation subunit is used to calculate the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient to the overall interference intensity. The contribution is obtained by weighting the normalized value and weight coefficient of each coupling coefficient. The dominant factor dynamic identification subunit is used to compare the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient. When the contribution of the first coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the sunlight incident angle and the reflection characteristics of the track surface is determined to be the dominant interference factor. When the contribution of the second coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the sunlight incident angle and the camera imaging attitude is determined to be the dominant interference factor. When the contribution of the third coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the reflection characteristics of the track surface and the camera imaging attitude is determined to be the dominant interference factor. The confidence assessment subunit for dominant factors is used to calculate the confidence of identifying dominant interference factors based on the contribution distribution. When no single coupling coefficient has a contribution exceeding the preset dominant threshold, it is determined to be a multi-factor composite interference mode. The coupling relationships corresponding to all coupling coefficients whose contributions exceed the preset composite threshold are taken as composite interference dominant factors. The composite interference mode and composite interference dominant factors are output as interference dominant factors to the strategy matching process of the anti-interference strategy decision unit.

[0011] Furthermore, the preset anti-interference strategy library adopts a layered storage structure, including a basic strategy layer, a collaborative strategy layer, and an emergency strategy layer; The basic strategy layer stores individual adjustment strategies for the adjustable polarization filter unit, individual adjustment strategies for the exposure time of the visible light binocular camera unit, and individual adjustment strategies for the power of the near-infrared laser module. The collaborative strategy layer stores dual-channel collaborative strategies for adjusting the adjustable polarization filter unit and the exposure time of the visible light binocular camera unit, dual-channel collaborative strategies for adjusting the adjustable polarization filter unit and the power of the near-infrared laser module, and dual-channel collaborative strategies for adjusting the exposure time of the visible light binocular camera unit and the power of the near-infrared laser module. The emergency strategy layer stores a three-channel collaborative strategy: adjustable polarization filter unit adjustment, visible light binocular camera unit exposure time adjustment, and near-infrared laser module power adjustment. The collaborative strategy layer also stores multi-channel collaborative strategies corresponding to multi-factor composite interference modes. The multi-channel collaborative strategies corresponding to multi-factor composite interference modes are multi-channel collaborative strategies that simultaneously include the adjustment measures corresponding to all composite interference dominant factors. The anti-interference strategy decision unit is also used to receive the output interference intensity level, and determine the target strategy layer corresponding to the current interference intensity level based on the established mapping relationship between the mild interference level and the triggering conditions of the basic strategy layer, the mapping relationship between the moderate interference level and the triggering conditions of the cooperative strategy layer, and the mapping relationship between the severe interference level and the triggering conditions of the emergency strategy layer. At the same time, it receives the output interference dominant factor or composite interference mode and composite interference dominant factor. When the interference mode is a single interference mode, the anti-interference strategy decision unit matches the cooperative anti-interference strategy corresponding to the interference dominant factor from the target strategy layer according to the interference dominant factor and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data. When the interference mode is a multi-factor composite interference mode, the anti-interference strategy decision unit matches the multi-channel collaborative strategy corresponding to the multi-factor composite interference mode from the collaborative strategy layer according to the dominant factor of composite interference and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data.

[0012] Furthermore, the tunable polarization filter unit includes: The solar reflected light polarization characteristic analysis subunit is used to calculate the polarization direction and degree of polarization of the sunlight reflected on the rail surface based on the illumination direction and spectral distribution collected by the four-dimensional situational awareness module and the Fresnel equation. The polarization direction dynamic tracking subunit is used to predict the polarization direction of the next moment based on the changing trajectory of the polarization direction of the sunlight reflected light using the Kalman filter algorithm, and adjust the polarization direction of the adjustable polarization filter unit in real time to make the polarization direction orthogonal to the predicted polarization direction of the reflected light. The polarization degree adaptive adjustment subunit is used to calculate the optimal polarization degree that maximizes the image signal-to-noise ratio under the current lighting conditions based on the light intensity collected by the four-dimensional situational awareness module, and adjust the polarization degree of the adjustable polarization filter unit in real time to match the transmittance with the imaging requirements. The multi-region independent adjustment subunit is used to divide the adjustable polarization filter unit into multiple independent adjustment regions. Based on the brightness distribution of different regions in the image acquired by the visible light binocular camera unit and the interference intensity level output by the anti-interference strategy decision unit, the polarization direction and polarization degree of each independent adjustment region are adjusted respectively.

[0013] Furthermore, the visible light binocular camera unit includes: The illumination-surface joint modeling subunit is used to establish an illumination-surface joint reflection model based on the illumination intensity and orbital surface state type collected by the four-dimensional situational awareness module, and to predict the image brightness distribution under different exposure times. The exposure time optimization solution subunit is used to minimize the mean square error between the predicted image brightness distribution and the ideal brightness distribution as the objective function, with the exposure time adjustment range as the constraint condition, and dynamically solves the optimal exposure time using the gradient descent algorithm. The exposure time of the visible light binocular camera unit is then adjusted according to the optimal exposure time. The gain dynamic adjustment subunit is used to perform frequency domain analysis on the real-time images acquired by the visible light binocular camera unit, calculate the main frequency and amplitude of the light intensity fluctuation, and dynamically adjust the gain value according to the light intensity fluctuation frequency to maximize the image signal-to-noise ratio within the light intensity fluctuation period. The binocular image consistency correction subunit is used to compare the images acquired by the left and right cameras in real time, calculate the brightness and contrast differences of the binocular images, and adjust the exposure time and gain of the left and right cameras differently based on the brightness and contrast differences to maintain the consistency of the binocular images. The gain dynamic adjustment subunit works in conjunction with the binocular image consistency correction subunit. First, the gain value is adjusted according to the frequency of light intensity fluctuations, and then differential fine-tuning is performed according to the differences in the binocular images.

[0014] Furthermore, the near-infrared laser module includes: The penetration depth prediction subunit is used to query a pre-established surface state-penetration depth mapping table based on the orbital surface state type output by the four-dimensional situational awareness module, and to predict the penetration depth of the near-infrared laser under the current surface state. The power dynamic optimization subunit is used to dynamically solve the optimal output power of the near-infrared laser module based on the estimated penetration depth and the interference intensity level output by the three-factor coupled analysis module, with the optimization objective of maximizing the weighted sum of penetration depth and imaging signal-to-noise ratio, and adjust the output power of the near-infrared laser module according to the optimal output power. The pulse frequency adaptive adjustment subunit is used to analyze the images acquired by the visible light binocular camera unit in real time, calculate the fluctuation frequency of ambient light, and dynamically adjust the pulse frequency of the near-infrared laser module according to the fluctuation frequency of ambient light, so that the pulse frequency is staggered with the fluctuation frequency of ambient light to form a difference frequency and suppress ambient light interference. The laser speckle suppression subunit is used to perform time-varying modulation of the laser output from the near-infrared laser module. It employs random phase modulation in the time dimension and multimode fiber coupling output in the spatial dimension to suppress speckle noise formed by the near-infrared laser on the rail surface. The pulse frequency adaptive adjustment subunit works in conjunction with the laser speckle suppression subunit. First, the pulse frequency is adjusted to suppress ambient light interference, and then the laser is subjected to time-varying control to suppress speckle noise.

[0015] Furthermore, the 3D reconstruction and contour output module includes: The multimodal feature point extraction subunit is used to extract the center point of visible light structured light stripes based on the anti-interference enhanced visible light image data, and to extract the center point of near-infrared structured light stripes based on the anti-interference enhanced near-infrared image data, so as to obtain visible light structured light spot clouds and near-infrared structured light spot clouds. The point cloud quality assessment subunit is used to assess the quality of visible light structured light spot clouds and near-infrared structured light spot clouds respectively, and calculates the point cloud density, point cloud signal-to-noise ratio, and point cloud edge sharpness as quality assessment indicators. The fusion weight dynamic calculation subunit is used to dynamically calculate the fusion weight of visible light point cloud and near-infrared point cloud based on the interference intensity level output by the anti-interference strategy decision unit and the quality evaluation index output by the point cloud quality evaluation subunit, using the fuzzy comprehensive evaluation method. The higher the interference intensity level, the greater the weight of near-infrared point cloud; the higher the signal-to-noise ratio of visible light point cloud, the greater the weight of visible light point cloud. The point cloud fusion and reconstruction sub-unit is used to perform weighted fusion of visible light structured light spot clouds and near-infrared structured light spot clouds based on dynamically calculated fusion weights to obtain fused point clouds. The fused point cloud is smoothed by the moving least squares method, and the three-dimensional mesh model of the rail surface is reconstructed by the triangulation algorithm. The contour geometry parameter extraction sub-unit is used to extract the rail cross-sectional contour line from the 3D mesh model of the rail surface, calculate the rail head width, rail web height, rail base width, and rail gauge angle geometry parameters, and output the extracted rail contour geometry parameters as the track contour detection result.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: This invention overcomes the fundamental shortcomings of existing passive redundancy anti-interference schemes, which rely solely on a symmetrical dual-camera layout and cannot actively sense changes in ambient light and differences in track surface conditions. This invention uses a four-dimensional situational awareness module to collect real-time data on light intensity, light direction, sunlight interference type, and track surface condition type. A three-factor coupling analysis module quantifies the coupling relationships between sunlight incident angle and track surface reflection characteristics, sunlight incident angle and camera imaging attitude, and track surface reflection characteristics and camera imaging attitude. A multi-modal imaging and hierarchical anti-interference module dynamically combines three anti-interference measures—polarization adjustment, exposure compensation, and near-infrared supplementary lighting—based on the dominant interference factors and interference intensity levels. Finally, a three-dimensional reconstruction and contour output module outputs high-precision track contour detection results. This invention achieves a fundamental leap from passive redundancy to active adaptation, enabling the system to dynamically select the optimal anti-interference strategy based on real-time perceived environmental information and track surface conditions. It can stably acquire high-quality images even under extreme lighting and complex surface conditions, ensuring that the track contour detection accuracy meets the all-weather detection requirements of high-speed rail.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall architecture diagram of a 3D visual track contour detection system resistant to sunlight interference according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the internal processing of the anti-interference strategy decision unit in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] refer to Figure 1 , Figure 2 This invention provides an embodiment of a 3D visual track contour detection system resistant to sunlight interference, comprising: a four-dimensional situational awareness module, including a light sensor array, a spectral analysis unit, and a track surface state recognition unit; the light sensor array is used to collect ambient light intensity and light direction in real time; the spectral analysis unit is used to identify the type of sunlight interference based on spectral distribution; the track surface state recognition unit is used to perform image recognition based on image data from a visible light image sensor, extract track surface texture features, and identify the type of track surface state; The three-factor coupling analysis module is used to receive data on light intensity, light direction, sunlight interference type, and track surface state type output by the four-dimensional situational awareness module. It calculates the first coupling coefficient between sunlight incident angle and track surface reflection characteristics, the second coupling coefficient between sunlight incident angle and camera imaging attitude, and the third coupling coefficient between track surface reflection characteristics and camera imaging attitude through a pre-trained coupling analysis model. The multimodal imaging and hierarchical anti-interference module includes a visible light binocular camera unit, a near-infrared laser module, an adjustable polarization filter unit, and an anti-interference strategy decision unit. The anti-interference strategy decision unit is used to determine the dominant interference factor and the interference intensity level based on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient output by the three-factor coupling analysis module. According to the dominant interference factor and the interference intensity level, it matches the corresponding collaborative anti-interference strategy from the preset anti-interference strategy library and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data. The 3D reconstruction and contour output module is used to perform multimodal fusion point cloud reconstruction based on anti-interference enhanced visible light image data and near-infrared image data, extract the 3D contour of the rail surface and output the track contour detection results.

[0022] In this embodiment, the light sensor array is a sensor group composed of multiple photoelectric detection units arranged in a specific geometric layout. Each detection unit can convert the received light signal into an electrical signal and obtain the light intensity value in the corresponding direction by measuring the intensity of the electrical signal.

[0023] In this embodiment, the ambient light intensity and direction are collected in real time by synchronously measuring the light signal through detection units at different positions in the light sensor array. Based on the difference in the intensity of the electrical signal output by each unit and combined with the geometric layout parameters of the sensor array, the main incident direction of the ambient light and its corresponding light intensity value at the current moment are calculated using the triangulation principle.

[0024] In this embodiment, spectral distribution refers to the energy distribution characteristics obtained after decomposing the collected ambient light according to different wavelengths. It is usually expressed as the relative proportion of light energy in different wavelength ranges, reflecting the composition of visible light, infrared light and other band components contained in the ambient light.

[0025] In this embodiment, identifying the type of sunlight interference based on spectral distribution involves comparing the collected ambient light spectrum with a pre-stored library of typical sunlight interference spectral features using a spectral analysis unit. When the energy ratio of a specific band in the ambient light spectrum matches typical spectral features such as direct sunlight, sunlight reflected from water, and sunlight reflected from snow, the specific type of sunlight interference is determined.

[0026] In this embodiment, sunlight interference type refers to the classification of interference patterns formed according to different sunlight incident methods and environmental media, including direct sunlight interference, sunlight reflected from water surface interference, sunlight reflected from snow, sunlight scattered by raindrops, and other specific types. The impact mechanism and degree of different types on the imaging system vary.

[0027] In this embodiment, image recognition based on image data from a visible light image sensor, extracting track surface texture features and identifying track surface state types are achieved by processing track images acquired by a visible light camera using a convolutional neural network. During the training phase, the network learns a large number of labeled sample images of track surface states such as rust, oil, and snow. During forward inference, it extracts feature maps such as texture roughness, gloss distribution, and edge sharpness from the input images, and finally outputs a classification result indicating whether the track surface state belongs to rust, oil, snow, or normal state.

[0028] In this embodiment, the image data of the visible light image sensor refers to the grayscale or color image of the track surface acquired by a CMOS or CCD image sensor that responds to the visible light band. The grayscale value or RGB component of each pixel in the image reflects the intensity of visible light reflection from the track surface.

[0029] In this embodiment, the track surface condition type refers to the classification of the surface condition of the rail surface due to different environmental and usage conditions, including rusted state, oil-covered state, snow-residual state, and clean and normal state. Different states have different light reflection characteristics and texture features.

[0030] In this embodiment, the pre-trained coupling analytical model is a regression model built on a multilayer perceptron neural network. During the training phase, the model uses a large amount of historical detection data as input samples. Each sample contains light intensity, light direction, sunlight interference type, track surface state type, and the corresponding correlation values ​​between sunlight incident angle and track surface reflection characteristics, sunlight incident angle and camera imaging attitude, and track surface reflection characteristics and camera imaging attitude as output labels. The network weights are optimized through backpropagation algorithm. After training, the model can calculate three coupling coefficients based on the input real-time environmental data.

[0031] In this embodiment, the data on light intensity, light direction, sunlight interference type, and track surface state type output by the four-dimensional situational awareness module are received. The first coupling coefficient between the sunlight incident angle and the track surface reflection characteristics, the second coupling coefficient between the sunlight incident angle and the camera imaging attitude, and the third coupling coefficient between the track surface reflection characteristics and the camera imaging attitude are calculated through a pre-trained coupled analytical model. This means that the four types of data collected by the four-dimensional situational awareness module are fed into the pre-trained multilayer perceptron neural network as input vectors. After the network undergoes multiple nonlinear transformations, it outputs three values ​​from the output layer. These three values ​​quantify the degree of mutual influence between the sunlight incident angle and the track surface reflection capability, the degree of matching between the sunlight incident angle and the camera imaging geometric attitude, and the degree of adaptation between the track surface reflection characteristics and the camera imaging parameters.

[0032] In this embodiment, the track surface reflection characteristics refer to the parameters of the rail surface's ability to reflect incident light, including the ratio of specular reflection components to diffuse reflection components, the functional relationship between reflectivity and incident angle, and optical properties such as the reflection differences of light of different wavelengths. These properties are affected by both the material and condition of the track surface.

[0033] In this embodiment, the near-infrared laser module is a light source device capable of emitting near-infrared laser light. It consists of a laser diode, a driving circuit, and a beam shaping optical system. The driving circuit controls the laser diode to output near-infrared laser light with a specific power and pulse frequency, and the beam shaping system adjusts the laser light into a linear structured light that is projected onto the surface of the rail.

[0034] In this embodiment, the dominant interference factor refers to the type of coupling relationship that has the greatest impact on imaging quality under the current lighting environment and track surface conditions. Specifically, it is the coupling relationship corresponding to the coefficient with the largest value among the three coupling coefficients. The strength of this coupling relationship dominates the degree of imaging interference in the current scene.

[0035] In this embodiment, the interference intensity level is a graded assessment of the current interference level based on the combined numerical value of the three coupling coefficients. The continuous coupling coefficient numerical space is divided into three discrete levels: mild interference, moderate interference, and severe interference using fuzzy logic. Each level corresponds to a different combination of anti-interference measures.

[0036] In this embodiment, the preset anti-interference strategy library is a structured data collection stored in non-volatile memory and organized using a hierarchical storage architecture. The basic strategy layer stores adjustment schemes for a single imaging unit, the collaborative strategy layer stores adjustment schemes for two imaging units working together, and the emergency strategy layer stores adjustment schemes for three imaging units working simultaneously. Each layer is indexed according to the dominant interference factor and the interference intensity level to facilitate rapid retrieval and matching.

[0037] In this embodiment, the collaborative anti-interference strategy refers to a set of specific control parameters retrieved from the preset anti-interference strategy library based on the dominant interference factor and the interference intensity level. These parameters include the polarization angle and polarization degree setting values ​​of the adjustable polarization filter unit, the exposure time and gain setting values ​​of the visible light binocular camera unit, and the output power and pulse frequency setting values ​​of the near-infrared laser module. These parameter combinations can simultaneously adjust multiple imaging units to work collaboratively to counteract the current interference.

[0038] In this embodiment, the enhanced visible light image data and near-infrared image data refer to the image data re-acquired by the visible light binocular camera unit and the near-infrared laser module after the multimodal imaging and hierarchical anti-interference module has executed the collaborative anti-interference strategy. These image data have a higher signal-to-noise ratio and clearer orbital contour information compared to the original images.

[0039] In this embodiment, multimodal fusion point cloud reconstruction based on anti-interference enhanced visible light image data and near-infrared image data, extracting the three-dimensional contour of the rail surface and outputting the track contour detection result refers to first extracting the center point of the structured light stripe from the visible light image to obtain the visible light point cloud, and extracting the center point of the structured light stripe from the near-infrared image to obtain the near-infrared point cloud. Then, the fusion weights of the two point clouds are dynamically adjusted according to the interference intensity level for weighted fusion. Then, the fused point cloud is connected into a three-dimensional mesh model using a triangulation algorithm. Finally, geometric parameters such as rail head width, rail web height, rail bottom width, and track gauge angle are calculated from the model as the detection result output.

[0040] Furthermore, the track surface state recognition unit includes: a texture feature extraction subunit, used to extract track surface texture features based on image data from a visible light image sensor, and calculate texture roughness for rusted state, gloss feature value for oil-covered state, and edge sharpness feature value for snow residue state; The multi-dimensional quantization sub-unit is used to input the texture roughness of the rusted state, the gloss feature value of the oil-covered state, and the edge sharpness feature value of the snow-residual state into the pre-trained surface state quantization model to obtain the rusted state coefficient, the oil-covered state coefficient, and the snow-residual state coefficient. The surface state fusion subunit is used to calculate the confidence distribution of the fused track surface state based on the rust state coefficient, oil stain coverage state coefficient, and snow residue state coefficient using the evidence theory method. The surface state type with the highest confidence is output as the track surface state type data to the three-factor coupled analysis module.

[0041] In this embodiment, the texture features of the track surface are extracted based on the image data of the visible light image sensor. The texture roughness of the rusted state, the gloss feature value of the oil-covered state, and the edge sharpness feature value of the snow-residual state are calculated by analyzing the visible light image using an image processing algorithm. The texture roughness of the rusted state is calculated by the gray-level co-occurrence matrix method to calculate the texture roughness of the local area of ​​the image. The gloss feature value of the oil-covered state is obtained by analyzing the area and intensity distribution of the highlight area of ​​the image. The edge sharpness feature value of the snow-residual state is obtained by calculating the statistical features of the edge gradient magnitude after extracting the image edge using the Sobel operator.

[0042] In this embodiment, the pre-trained surface state quantification model is a machine learning model built on support vector regression. During the training phase, the model uses a large number of labeled track surface image samples. Each sample contains manually labeled rust level scores, oil stain level scores, and snow accumulation level scores as output labels. The texture roughness, glossiness feature values, and edge sharpness feature values ​​extracted from the images are used as input features. Through training, the model learns the mapping relationship between feature values ​​and surface state scores.

[0043] In this embodiment, the texture roughness of the rusted state, the gloss of the oil-covered state, and the edge sharpness of the snow-residual state are input into a pre-trained surface state quantification model to obtain the rusted state coefficient, the oil-covered state coefficient, and the snow-residual state coefficient. This means that the feature vector composed of the three feature values ​​calculated from the current image is sent to the pre-trained support vector regression model. After kernel function mapping and regression calculation, the model outputs three values. These three values, respectively ranging from zero to one, quantify the severity of rust, oil coverage, and snow residue on the track surface.

[0044] In this embodiment, the rust state coefficient, oil stain coverage state coefficient, and snow residue state coefficient are three dimensionless values ​​in the range of zero to one. The closer the rust state coefficient is to one, the more severe the rust on the rail surface. The closer the oil stain coverage state coefficient is to one, the larger the oil stain coverage area or the thicker the oil layer. The closer the snow residue state coefficient is to one, the greater the amount of snow residue or the more complete the snow coverage.

[0045] In this embodiment, the confidence distribution of the fused track surface state is calculated using the evidence theory method based on the rust state coefficient, oil stain coverage state coefficient, and snow residue state coefficient. This means that the three coefficients are used as the basic probability allocation values ​​supporting the three states of rust, oil stain, and snow, respectively. By combining information from different evidence sources through the synthesis rules of evidence theory, a joint confidence distribution that considers the possibilities of the three surface states is obtained. This distribution includes the confidence values ​​of the four hypotheses: rust state, oil stain state, snow residue state, and normal state.

[0046] In this embodiment, outputting the surface state type with the highest confidence level as the track surface state type data to the three-factor coupling analysis module means finding the state hypothesis with the highest confidence value from the fused confidence distribution. If the rust state has the highest confidence level, the current track surface is determined to be in the rust state type; if the oil stain state has the highest confidence level, it is determined to be in the oil stain coverage state type; if the snow accumulation state has the highest confidence level, it is determined to be in the snow residue state type; if the normal state has the highest confidence level, it is determined to be in the clean and normal state type. The determination result is sent to the three-factor coupling analysis module in the form of a state type identifier as the basis for subsequent calculation and analysis.

[0047] Furthermore, the three-factor coupling analysis module includes: a historical data organization subunit, which is used to organize the lighting conditions, track surface conditions, and imaging quality recorded in the historical detection data in a multi-dimensional manner according to the monitoring time point, monitoring section location, and feature parameter type, to form a time series dataset of lighting conditions, a spatial distribution dataset of track surface conditions, and a set of imaging quality feature parameters; The coupled graph neural network constructs a subunit, which is used to establish a three-node dynamic coupled graph structure based on the time series dataset of illumination conditions, the spatial distribution dataset of track surface state, and the set of imaging quality feature parameters. The graph neural network is used to establish a three-node dynamic coupled graph structure with the sunlight incident angle as the first node, the track surface reflection characteristics as the second node, and the camera imaging attitude as the third node. The connection weights in the dynamic coupled graph structure are dynamically updated according to the real-time monitoring data. The multi-scale coupled analytical subunit is used to learn the short-term fluctuation coupling relationship and long-term trend coupling relationship between the first node and the second node at different time scales through a dynamic coupled graph structure. The short-term fluctuation coupling relationship and the long-term trend coupling relationship are weighted and fused to obtain the first coupling coefficient. The dynamic coupled graph structure is used to learn the static geometric coupling relationship and dynamic motion coupling relationship between the first node and the third node at different spatial scales through a dynamic coupled graph structure. The static geometric coupling relationship and the dynamic motion coupling relationship are weighted and fused to obtain the second coupling coefficient. The dynamic coupled graph structure is used to learn the material property coupling relationship and imaging physics coupling relationship between the second node and the third node at different physical scales through a dynamic coupled graph structure. The material property coupling relationship and the imaging physics coupling relationship are weighted and fused to obtain the third coupling coefficient. The first coupling coefficient, the second coupling coefficient, and the third coupling coefficient are output to the anti-interference strategy decision unit.

[0048] In this embodiment, the lighting conditions, track surface conditions, and imaging quality recorded in the historical detection data are organized in multiple dimensions according to the monitoring time point, monitoring section location, and feature parameter type to form a lighting condition time series dataset, a track surface condition spatial distribution dataset, and an imaging quality feature parameter set. This means that the historical records in the database are reorganized in three dimensions: the lighting condition data is arranged in chronological order to form a sequence of light intensity and light direction that changes over time; the track surface condition data is arranged according to the track section location to form a spatial distribution sequence along the longitudinal direction of the track; and the imaging quality data is classified according to the feature parameter type to form a feature vector set containing parameters such as signal-to-noise ratio, edge sharpness, and contrast.

[0049] In this embodiment, based on the time-series dataset of illumination conditions, the spatial distribution dataset of track surface states, and the set of imaging quality feature parameters, a graph neural network is used to establish a three-node dynamically coupled graph structure with the sunlight incident angle as the first node, the track surface reflection characteristics as the second node, and the camera imaging attitude as the third node. The connection weights in the dynamically coupled graph structure are dynamically updated according to real-time monitoring data. This means that the three datasets are mapped to three nodes in the graph structure, and each node contains the feature representation of the corresponding data. The initial values ​​of the connection edge weights between nodes are determined by the statistical correlation of historical data. During system operation, whenever new real-time monitoring data arrives, the graph neural network iteratively adjusts the weights of the three edges according to the degree of matching between the new data and the historical data, so that the graph structure can reflect the coupling strength under the current environment in real time.

[0050] In this embodiment, the short-term fluctuation coupling relationship and long-term trend coupling relationship between the first node and the second node are learned at different time scales through a dynamic coupling graph structure. The first coupling coefficient is obtained by weighted fusion of the short-term fluctuation coupling relationship and the long-term trend coupling relationship. This means that the graph neural network extracts the data fluctuation characteristics of the first node and the second node in the minute-level time window as the short-term fluctuation coupling relationship, and extracts the data change trend in the hour-level or day-level time window as the long-term trend coupling relationship. Then, the two coupling relationships are linearly weighted and combined according to the preset weight ratio to obtain a comprehensive value as the first coupling coefficient characterizing the degree of mutual influence between the sunlight incident angle and the reflection characteristics of the orbital surface.

[0051] In this embodiment, the static geometric coupling relationship and dynamic motion coupling relationship between the first node and the third node are learned at different spatial scales through a dynamic coupling graph structure. The second coupling coefficient is obtained by weighted fusion of the static geometric coupling relationship and the dynamic motion coupling relationship. This means that the graph neural network extracts the data correspondence between the first node and the third node at a fixed monitoring section as the static geometric coupling relationship, and extracts the data change correspondence when monitoring along the longitudinal movement of the track as the dynamic motion coupling relationship. Then, the two coupling relationships are linearly weighted and combined according to a preset weight ratio to obtain a comprehensive value as the second coupling coefficient characterizing the degree of matching between the sunlight incident angle and the camera imaging posture.

[0052] In this embodiment, the material property coupling relationship and imaging physics coupling relationship between the second node and the third node are learned at different physical scales through a dynamic coupling graph structure. The third coupling coefficient is obtained by weighted fusion of the material property coupling relationship and the imaging physics coupling relationship. The graph neural network extracts the correspondence between the track surface material features in the second node and the camera imaging parameters in the third node as the material property coupling relationship, and extracts the correspondence between the surface reflection characteristics in the second node and the image quality indicators in the third node as the imaging physics coupling relationship. Then, the two coupling relationships are linearly weighted and combined according to a preset weight ratio to obtain a comprehensive value as the third coupling coefficient that characterizes the degree of adaptation between the track surface reflection characteristics and the camera imaging attitude.

[0053] Furthermore, the anti-interference strategy decision unit includes: a coupling coefficient time series analysis subunit, which is used to perform time series analysis on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient, and calculate the mean, variance, rate of change, and fluctuation frequency of each coupling coefficient within a preset time window; The interference intensity dynamic grading subunit is used to dynamically divide the coupling intensity of the three factors into three levels: mild interference level, moderate interference level, and severe interference level based on the current value, variance, and rate of change of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient using fuzzy logic. The grading threshold is adaptively adjusted according to the statistical distribution of the coupling coefficient. The hierarchical trigger condition mapping subunit is used to establish the mapping relationship between mild interference level and single anti-interference measure trigger condition, moderate interference level and dual-channel collaborative anti-interference measure trigger condition, and severe interference level and three-channel collaborative anti-interference measure trigger condition. The mapping relationship is stored in the storage space accessible by the strategy matching process, and the interference intensity level is output to the strategy matching process of the anti-interference strategy decision unit.

[0054] In this embodiment, time series analysis is performed on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient. The mean, variance, rate of change, and fluctuation frequency of each coupling coefficient within a preset time window are calculated by taking the numerical sequence of each coupling coefficient continuously collected over a period of time as the analysis object. The mean reflects the average level within the time period, the variance reflects the dispersion of the values, the rate of change is calculated by dividing the difference between the current value and the previous value by the sampling interval, and the fluctuation frequency is obtained by counting the number of times the value crosses the mean within a unit of time.

[0055] In this embodiment, the three-factor coupling strength refers to the overall degree of interference of the current environment on the imaging system, which is reflected by the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient. This strength is determined by the comprehensive value of the three coupling coefficients. The larger the value, the stronger the mutual influence between sunlight, orbital surface, and camera attitude, and the more serious the interference to imaging.

[0056] In this embodiment, based on the current values, variances, and rates of change of the first, second, and third coupling coefficients, a fuzzy logic method is used to dynamically classify the coupling strength of the three factors into three levels: mild interference, moderate interference, and severe interference. The adaptive adjustment of the classification threshold with the statistical distribution of the coupling coefficients means that the current values, variances, and rates of change of the three coupling coefficients are used as input variables of the fuzzy logic system. The input values ​​are converted into fuzzy sets through a predefined membership function. The interference level is output by reasoning based on the fuzzy rule base. At the same time, the system periodically counts the distribution characteristics of historical coupling coefficients and dynamically updates the threshold parameters in the fuzzy rules according to the statistical results, so that the classification standard can adapt to the typical interference level changes in different seasons and time periods.

[0057] In this embodiment, the grading threshold refers to the specific numerical boundary used in the fuzzy logic method to divide different levels of interference, including the threshold that distinguishes between mild and moderate interference, and the threshold that distinguishes between moderate and severe interference. These thresholds are not fixed, but are dynamically adjusted according to historical statistical distribution.

[0058] In this embodiment, the trigger condition for a single anti-interference measure refers to the set of control rules activated by the system when the interference intensity level is determined to be mild interference. This set of rules stipulates that only the adjustment function of one imaging unit is activated. Which unit is activated is determined by the dominant interference factor. For example, when the dominant interference factor is the coupling of the sunlight incident angle and the camera imaging attitude, the trigger condition stipulates that the adjustable polarization filter unit is activated for individual adjustment.

[0059] In this embodiment, establishing a mapping relationship between the mild interference level and the triggering condition of a single anti-interference measure means associating and storing the status identifier of the mild interference level with the corresponding triggering rule of the single anti-interference measure within the system. When the system determines that the current interference level is mild interference, it can quickly retrieve the specific triggering rule that should be executed through this mapping relationship.

[0060] In this embodiment, the triggering condition for the dual-channel collaborative anti-interference measure refers to the set of control rules activated by the system when the interference intensity level is determined to be moderate interference. This set of rules stipulates that the adjustment functions of two imaging units are activated simultaneously and that the two work together. The specific two units activated are determined by the dominant interference factor. For example, when the dominant interference factor is the coupling of the sunlight incident angle and the reflection characteristics of the orbital surface, the triggering condition stipulates that the adjustable polarization filter unit and the visible light binocular camera unit's exposure time adjustment are activated simultaneously and that the two work together.

[0061] In this embodiment, the mapping relationship between the moderate interference level and the triggering conditions of the dual-channel collaborative anti-interference measures refers to the system's internal association and storage of the status identifier of the moderate interference level with the corresponding triggering rules of the dual-channel collaborative anti-interference measures. When the system determines that the current interference level is moderate interference, it can quickly retrieve the specific triggering rules that should be executed through this mapping relationship.

[0062] In this embodiment, the triggering condition for the three-channel collaborative anti-interference measure refers to the set of control rules activated by the system when the interference intensity level is determined to be severe interference. This set of rules stipulates that the adjustment functions of the three imaging units are activated simultaneously and that the three work together. Specifically, all three units are activated, including the adjustment of the adjustable polarization filter unit, the exposure time adjustment of the visible light binocular camera unit, and the power adjustment of the near-infrared laser module, which are executed simultaneously and cooperate with each other.

[0063] In this embodiment, the mapping relationship between the severe interference level and the triggering conditions of the three-channel collaborative anti-interference measures refers to the system's internal association and storage of the severe interference level status identifier with the corresponding three-channel collaborative anti-interference measure triggering rules. When the system determines that the current interference level is severe interference, it can quickly retrieve the specific triggering rules that should be executed through this mapping relationship.

[0064] In this embodiment, storing the mapping relationships in the storage space accessible to the strategy matching process means that the mapping relationships between mild interference levels and basic strategy layer triggering conditions, moderate interference levels and cooperative strategy layer triggering conditions, and severe interference levels and emergency strategy layer triggering conditions are stored in the system's shared memory area in the form of lookup tables. This storage space can be read and accessed in real time by the strategy matching process of the anti-interference strategy decision unit. Outputting the interference intensity level to the strategy matching process of the anti-interference strategy decision unit means that after the dynamic interference intensity grading subunit completes the grading calculation, it sends the current mild, moderate, or severe interference level identifier as dynamic data to the strategy matching process. After receiving the interference intensity level, the strategy matching process determines whether the target strategy layer corresponding to the current interference intensity level is the basic strategy layer, cooperative strategy layer, or emergency strategy layer by accessing the pre-established mapping relationship lookup table in the storage space.

[0065] Furthermore, the anti-interference strategy decision unit includes: a coupling coefficient contribution calculation subunit, which is used to calculate the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient to the overall interference intensity. The contribution is obtained by weighting the normalized value and weight coefficient of each coupling coefficient. The dominant factor dynamic identification subunit is used to compare the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient. When the contribution of the first coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the sunlight incident angle and the reflection characteristics of the track surface is determined to be the dominant interference factor. When the contribution of the second coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the sunlight incident angle and the camera imaging attitude is determined to be the dominant interference factor. When the contribution of the third coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the reflection characteristics of the track surface and the camera imaging attitude is determined to be the dominant interference factor. The confidence assessment subunit for dominant factors is used to calculate the confidence of identifying dominant interference factors based on the contribution distribution. When no single coupling coefficient has a contribution exceeding the preset dominant threshold, it is determined to be a multi-factor composite interference mode. The coupling relationships corresponding to all coupling coefficients whose contributions exceed the preset composite threshold are taken as composite interference dominant factors. The composite interference mode and composite interference dominant factors are output as interference dominant factors to the strategy matching process of the anti-interference strategy decision unit.

[0066] In this embodiment, the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient to the overall interference intensity is calculated. The contribution is obtained by weighting the normalized value and weight coefficient of each coupling coefficient. First, the three coupling coefficients are normalized by dividing the current value of each coefficient by its historical maximum value to obtain a normalized value in the range of zero to one. Then, a weight coefficient is assigned according to the importance of each coupling coefficient in affecting image quality. The sum of the three weight coefficients is one. The specific values ​​of the weight coefficients are obtained through experimental calibration. For example, the weight of the first coupling coefficient is 0.4, the weight of the second coupling coefficient is 0.3, and the weight of the third coupling coefficient is 0.3. Finally, the normalized value of each coupling coefficient is multiplied by its corresponding weight coefficient and summed to obtain the contribution value of each coupling coefficient to the overall interference intensity. The larger the contribution, the more significant the role of the coupling relationship in the current interference.

[0067] In this embodiment, the preset dominant threshold is a pre-set numerical parameter with a value range between 0.5 and 0.8. The specific value is determined through experimental calibration based on the track detection accuracy requirements and the field environment. It is used to determine whether the contribution of a certain coupling coefficient is sufficient to become the dominant interference factor. When the contribution value of a certain coupling coefficient exceeds the threshold, the coupling relationship corresponding to the coefficient is determined to be the dominant factor of the current interference scenario. For example, when the contribution of the first coupling coefficient is 0.6 and exceeds the preset dominant threshold of 0.5, the coupling between the sunlight incident angle and the reflection characteristics of the track surface is determined to be the dominant interference factor.

[0068] In this embodiment, calculating the confidence level for identifying the dominant interference factor based on the contribution distribution involves constructing a probability distribution from the contribution values ​​of the three coupling coefficients, calculating the difference between the largest and second largest contribution, and using the ratio of this difference to the sum of the contributions as the confidence index. The larger the difference, the clearer the dominant factor and the higher the confidence level. Conversely, when multiple contribution values ​​are similar, the confidence level is low, indicating that there may be multi-factor compound interference. For example, when the three contribution values ​​are 0.5, 0.4, and 0.1, the difference between the largest and second largest contribution is 0.1, the sum of the contributions is 1, and the confidence level is 0.1, which is at a low level.

[0069] In this embodiment, the preset composite threshold is a pre-set numerical parameter with a value ranging from 0.2 to 0.4. The specific value is determined through experimental calibration based on the sensitivity requirements for judging multi-factor composite interference. It is used to determine which coupling relationships should be included in the dominant factor of composite interference under the multi-factor composite interference mode. When the contribution value of a certain coupling coefficient exceeds the threshold, the coupling relationship corresponding to the coefficient is regarded as a component of composite interference. For example, when the preset composite threshold is 0.25, the coupling coefficient with a contribution of 0.3 is included in the dominant factor of composite interference, while the coupling coefficient with a contribution of 0.2 is not included.

[0070] In this embodiment, taking the coupling relationships corresponding to all coupling coefficients whose contribution exceeds the preset composite threshold as the dominant factor of composite interference means that, under the multi-factor composite interference mode, the contribution values ​​of the three coupling coefficients are traversed, and all contributions greater than the preset composite threshold are selected. The coupling relationships between the sunlight incident angle and the reflection characteristics of the orbital surface, the coupling relationship between the sunlight incident angle and the camera imaging attitude, or the coupling relationship between the reflection characteristics of the orbital surface and the camera imaging attitude corresponding to these contributions are taken together as the dominant factor of composite interference. For example, when the contribution of the first coupling coefficient is 0.4, the contribution of the second coupling coefficient is 0.3, the contribution of the third coupling coefficient is 0.1, and the preset composite threshold is 0.25, the coupling relationships corresponding to the first coupling coefficient and the second coupling coefficient together constitute the dominant factor of composite interference, which is used for subsequent matching of multi-channel collaborative strategies that simultaneously include polarization adjustment and exposure adjustment measures.

[0071] Furthermore, the preset anti-interference strategy library adopts a layered storage structure, including a basic strategy layer, a collaborative strategy layer, and an emergency strategy layer; The basic strategy layer stores individual adjustment strategies for the adjustable polarization filter unit, individual adjustment strategies for the exposure time of the visible light binocular camera unit, and individual adjustment strategies for the power of the near-infrared laser module. The collaborative strategy layer stores dual-channel collaborative strategies for adjusting the adjustable polarization filter unit and the exposure time of the visible light binocular camera unit, dual-channel collaborative strategies for adjusting the adjustable polarization filter unit and the power of the near-infrared laser module, and dual-channel collaborative strategies for adjusting the exposure time of the visible light binocular camera unit and the power of the near-infrared laser module. The emergency strategy layer stores a three-channel collaborative strategy: adjustable polarization filter unit adjustment, visible light binocular camera unit exposure time adjustment, and near-infrared laser module power adjustment. The collaborative strategy layer also stores multi-channel collaborative strategies corresponding to multi-factor composite interference modes. The multi-channel collaborative strategies corresponding to multi-factor composite interference modes are multi-channel collaborative strategies that simultaneously include the adjustment measures corresponding to all composite interference dominant factors. The anti-interference strategy decision unit is also used to receive the output interference intensity level, and determine the target strategy layer corresponding to the current interference intensity level based on the established mapping relationship between the mild interference level and the triggering conditions of the basic strategy layer, the mapping relationship between the moderate interference level and the triggering conditions of the cooperative strategy layer, and the mapping relationship between the severe interference level and the triggering conditions of the emergency strategy layer. At the same time, it receives the output interference dominant factor or composite interference mode and composite interference dominant factor. When the interference mode is a single interference mode, the anti-interference strategy decision unit matches the cooperative anti-interference strategy corresponding to the interference dominant factor from the target strategy layer according to the interference dominant factor and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data. When the interference mode is a multi-factor composite interference mode, the anti-interference strategy decision unit matches the multi-channel collaborative strategy corresponding to the multi-factor composite interference mode from the collaborative strategy layer according to the dominant factor of composite interference and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data.

[0072] In this embodiment, the adjustable polarization filter unit adjustment strategy refers to activating only the adjustment function of the adjustable polarization filter unit, adjusting the polarization direction and degree of polarization of the polarizer according to the current illumination direction and intensity, without changing the exposure time of the visible light binocular camera unit and the output power of the near-infrared laser module. This strategy is suitable for scenarios with mild interference and where the dominant interference factor is related to the camera imaging posture.

[0073] In this embodiment, the strategy of individually adjusting the exposure time of the visible light binocular camera unit refers to activating only the exposure time adjustment function of the visible light binocular camera unit, optimizing the exposure time of the left and right cameras according to the current light intensity and the type of track surface condition, without changing the polarization parameters of the adjustable polarization filter unit and the output power of the near-infrared laser module. This strategy is suitable for scenarios with mild interference and where the dominant interference factor is related to the reflection characteristics of the track surface.

[0074] In this embodiment, the near-infrared laser module power adjustment strategy refers to activating only the power adjustment function of the near-infrared laser module, adjusting the laser output power according to the current track surface condition type and interference intensity, without changing the polarization parameters of the adjustable polarization filter unit and the exposure time of the visible light binocular camera unit. This strategy is suitable for scenarios with mild interference and where the dominant interference factor is related to the near-infrared imaging requirements.

[0075] In this embodiment, the dual-channel collaborative strategy of adjusting the adjustable polarization filter unit and adjusting the exposure time of the visible light binocular camera unit refers to simultaneously activating the adjustment functions of the adjustable polarization filter unit and the visible light binocular camera unit, so that the polarization direction, polarization degree and exposure time are coordinated with each other. For example, the polarization direction is adjusted according to the illumination direction while the exposure time is optimized according to the illumination intensity, so that the two adjustment measures work together to achieve the best imaging effect. This strategy is suitable for scenarios with moderate interference and where the dominant interference factors involve the camera imaging posture and the reflection characteristics of the track surface.

[0076] In this embodiment, the dual-channel coordinated strategy of adjusting the adjustable polarization filter unit and adjusting the near-infrared laser module power refers to simultaneously activating the adjustment functions of the adjustable polarization filter unit and the near-infrared laser module, so that the polarization parameters and laser power work together. For example, while adjusting the polarization direction to suppress reflected light, the near-infrared laser power is increased to enhance the structured light signal, so that the two adjustment measures work together. This strategy is suitable for scenarios with moderate interference where the dominant interference factors involve camera imaging posture and near-infrared imaging requirements.

[0077] In this embodiment, the dual-channel coordinated strategy of adjusting the exposure time of the visible light binocular camera unit and the power of the near-infrared laser module refers to simultaneously activating the adjustment functions of the visible light binocular camera unit and the near-infrared laser module, so that the exposure time and the laser power are coordinated with each other. For example, while increasing the exposure time to improve the brightness of the visible light image, the near-infrared laser power is appropriately reduced to avoid overexposure, so that the two adjustment measures work together. This strategy is suitable for scenarios with moderate interference and where the dominant interference factors involve the reflection characteristics of the track surface and the requirements of near-infrared imaging.

[0078] In this embodiment, the three-channel coordinated strategy of adjusting the adjustable polarization filter unit, adjusting the exposure time of the visible light binocular camera unit, and adjusting the power of the near-infrared laser module refers to simultaneously activating the adjustment functions of all three imaging units, so that the polarization parameters, exposure time, and laser power work together. For example, the polarization direction is adjusted according to the illumination direction to suppress reflected light, the exposure time is optimized according to the illumination intensity to improve the image signal-to-noise ratio, and the near-infrared laser power is adjusted according to the track surface condition to enhance the structured light signal. The three adjustment measures work together to achieve the best imaging effect. This strategy is suitable for heavily interfered scenarios.

[0079] In this embodiment, the interference intensity level is received, and based on the established mapping relationships between mild interference level and basic strategy layer triggering conditions, moderate interference level and cooperative strategy layer triggering conditions, and severe interference level and emergency strategy layer triggering conditions, the target strategy layer corresponding to the current interference intensity level is determined. The strategy matching process of the anti-interference strategy decision unit receives the interference intensity level identifier from weight 4, and then accesses the pre-established mapping relationship lookup table in the storage space. Based on the correspondence recorded in the lookup table between mild interference and basic strategy layer, moderate interference and cooperative strategy layer, and severe interference and emergency strategy layer, it is determined whether the current interference intensity level should activate the basic strategy layer, cooperative strategy layer, or emergency strategy layer as the target strategy layer for this strategy matching.

[0080] In this embodiment, when the interference mode is a single interference mode, the anti-interference strategy decision unit matches and executes the cooperative anti-interference strategy corresponding to the dominant interference factor from the target strategy layer according to the dominant interference factor. To obtain the enhanced visible light image data and near-infrared image data, the strategy matching process first determines that the dominant interference factor received from weight 5 is of a single type. Then, according to the specific type of the dominant factor, it retrieves the corresponding strategy item from the strategy set stored in the target strategy layer. For example, when the dominant factor is the coupling of the sunlight incident angle and the camera imaging attitude, it matches the adjustable polarization filter unit's individual adjustment strategy or the dual-channel strategy containing the adjustment measure from the target strategy layer. The retrieved strategy parameters are sent to the corresponding execution unit to complete the adjustment. Finally, the adjusted image is acquired as the enhancement result.

[0081] In this embodiment, when the interference mode is a multi-factor composite interference mode, the anti-interference strategy decision unit matches the multi-channel collaborative strategy corresponding to the multi-factor composite interference mode from the collaborative strategy layer according to the dominant factor of composite interference and executes it. The obtained anti-interference enhanced visible light image data and near-infrared image data refer to the strategy matching process determining that the interference mode received from weight 5 is a composite type, obtaining all coupling relationship types contained in the dominant factor of composite interference, retrieving multi-channel collaborative strategy items that simultaneously contain the adjustment measures corresponding to these coupling relationships from the collaborative strategy layer, sending the parameters of the strategy to the corresponding execution unit to complete the adjustment, and finally acquiring the adjusted image as the enhancement result.

[0082] Furthermore, the adjustable polarization filtering unit includes: a sunlight reflection polarization characteristic analysis subunit, which is used to calculate the polarization direction and degree of polarization of sunlight reflected on the rail surface based on the illumination direction and spectral distribution collected by the four-dimensional situational awareness module and the Fresnel equation; The polarization direction dynamic tracking subunit is used to predict the polarization direction of the next moment based on the changing trajectory of the polarization direction of the sunlight reflected light using the Kalman filter algorithm, and adjust the polarization direction of the adjustable polarization filter unit in real time to make the polarization direction orthogonal to the predicted polarization direction of the reflected light. The polarization degree adaptive adjustment subunit is used to calculate the optimal polarization degree that maximizes the image signal-to-noise ratio under the current lighting conditions based on the light intensity collected by the four-dimensional situational awareness module, and adjust the polarization degree of the adjustable polarization filter unit in real time to match the transmittance with the imaging requirements. The multi-region independent adjustment subunit is used to divide the adjustable polarization filter unit into multiple independent adjustment regions. Based on the brightness distribution of different regions in the image acquired by the visible light binocular camera unit and the interference intensity level output by the anti-interference strategy decision unit, the polarization direction and polarization degree of each independent adjustment region are adjusted respectively.

[0083] In this embodiment, the polarization direction and degree of polarization of the sunlight reflected from the rail surface are calculated based on the illumination direction and spectral distribution collected by the four-dimensional situational awareness module using the Fresnel equation. This means taking the illumination direction as the incident angle input, using the light intensity of different wavelengths in the spectral distribution as weighting factors, and calculating the ratio of polarized to unpolarized components in the reflected light using the Fresnel formula based on the refractive index parameters of the rail surface material. This yields the main polarization direction angle and the proportion of polarized components in the reflected light.

[0084] In this embodiment, the trajectory of the change in the polarization direction of reflected sunlight refers to a series of polarization direction angle values ​​of reflected light recorded over a continuous period of time, arranged in chronological order to form a sequence of data. This sequence reflects the pattern and trend of the change in the polarization direction of reflected light over time due to the movement of the sun's position or the movement of the detection vehicle.

[0085] In this embodiment, predicting the polarization direction of the next moment using the Kalman filter algorithm based on the changing trajectory of the polarization direction of reflected sunlight means inputting the historical polarization direction sequence as the observation value into the Kalman filter. The filter estimates the polarization direction state of the next moment through the state prediction equation, and then uses the latest measured value to correct the prediction error through the observation update equation. The optimal estimate of the polarization direction of the next moment is obtained by iterative recursion.

[0086] In this embodiment, adjusting the polarization direction of the adjustable polarization filter unit in real time so that the polarization direction is orthogonal to the predicted polarization direction of the reflected light means adding 90 degrees to the angle value of the polarization direction of the reflected light predicted by the Kalman filter at the next moment as the target polarization direction. The polarizer is rotated by the drive mechanism so that the transmission axis of the polarizer is aligned with the target direction. At this time, the transmittance of the polarizer to the reflected light is the lowest, which can suppress the interference of the reflected light to the greatest extent.

[0087] In this embodiment, the optimal polarization degree that maximizes the image signal-to-noise ratio under the current lighting conditions is calculated based on the light intensity collected by the four-dimensional situational awareness module. This means substituting the light intensity value into a pre-calibrated polarization degree and image signal-to-noise ratio relationship function. This function is obtained by experimentally determining the polarization degree value that maximizes the ratio of the image sensor output signal to noise under different light intensities, and obtaining the optimal polarization degree corresponding to the current lighting conditions through interpolation or table lookup.

[0088] In this embodiment, adjusting the polarization degree of the adjustable polarization filter unit in real time to match the transmittance with the imaging requirements means adjusting the relative angle of the two polarization layers in the polarizer or changing the driving voltage of the liquid crystal polarization element according to the calculated optimal polarization degree value, so that the transmittance of the polarizer changes from zero to one to the transmittance level corresponding to the target polarization degree, thereby suppressing reflected light while ensuring sufficient light intake.

[0089] In this embodiment, transmittance refers to the percentage ratio of the remaining light intensity after the light passes through the polarization filter unit to the incident light intensity. The higher the transmittance, the more light passes through, the brighter the image, but the weaker the effect of suppressing reflected light. The lower the transmittance, the less light passes through, the darker the image, but the stronger the effect of suppressing reflected light.

[0090] In this embodiment, dividing the adjustable polarization filter unit into multiple independent adjustment regions means dividing the effective light-transmitting surface of the polarizer or liquid crystal polarization element into two-dimensional grids or annular sectors. Each segmented region is equipped with an independent driving electrode or adjustment mechanism, so that different regions can be set with different polarization directions and polarization degree parameters.

[0091] In this embodiment, adjusting the polarization direction and degree of polarization of each independent adjustment region based on the brightness distribution of different regions in the image acquired by the visible light binocular camera unit and the interference intensity level output by the anti-interference strategy decision unit means performing brightness statistics on the image acquired by the camera in different regions, calculating the average gray value of each image region as brightness distribution data, and determining the degree to which reflected light needs to be suppressed in each region in combination with the current interference intensity level. Regions with high brightness correspond to regions with strong reflected light, and need to be set with a polarization direction that is closer to orthogonal and a higher degree of polarization. Regions with low brightness are set with an appropriate polarization direction and a lower degree of polarization to ensure image uniformity.

[0092] In this embodiment, the polarization direction and polarization degree of the independently adjustable region refer to the polarization parameters independently set by each segmented polarization adjustment unit. The polarization direction is the orientation angle of the light transmission axis of the region, and the polarization degree is the ratio of the transmittance of the polarizer of the region to the transmittance to natural light. The two together determine the modulation effect of the region on the incident light.

[0093] Furthermore, the visible light binocular camera unit includes: an illumination-surface joint modeling subunit, which is used to establish an illumination-surface joint reflection model based on the illumination intensity and orbital surface state type collected by the four-dimensional situational awareness module, and predict the image brightness distribution under different exposure times; The exposure time optimization solution subunit is used to minimize the mean square error between the predicted image brightness distribution and the ideal brightness distribution as the objective function, with the exposure time adjustment range as the constraint condition, and dynamically solves the optimal exposure time using the gradient descent algorithm. The exposure time of the visible light binocular camera unit is then adjusted according to the optimal exposure time. The gain dynamic adjustment subunit is used to perform frequency domain analysis on the real-time images acquired by the visible light binocular camera unit, calculate the main frequency and amplitude of the light intensity fluctuation, and dynamically adjust the gain value according to the light intensity fluctuation frequency to maximize the image signal-to-noise ratio within the light intensity fluctuation period. The binocular image consistency correction subunit is used to compare the images acquired by the left and right cameras in real time, calculate the brightness and contrast differences of the binocular images, and adjust the exposure time and gain of the left and right cameras differently based on the brightness and contrast differences to maintain the consistency of the binocular images. The gain dynamic adjustment subunit works in conjunction with the binocular image consistency correction subunit. First, the gain value is adjusted according to the frequency of light intensity fluctuations, and then differential fine-tuning is performed according to the differences in the binocular images.

[0094] In this embodiment, based on the light intensity and track surface condition type collected by the four-dimensional situational awareness module, a light-surface joint reflection model is established to predict the image brightness distribution under different exposure times. This involves taking the light intensity as the incident light energy input, taking the reflectivity parameter corresponding to the track surface condition type as the surface attribute input, establishing an energy transfer function describing the energy transfer of incident light after reflection from the track surface into the camera, multiplying this function by the charge accumulation time corresponding to different exposure times, calculating the theoretical gray value of each pixel position of the image sensor, and forming the predicted brightness distribution image.

[0095] In this embodiment, the illumination-surface joint reflection model is a mathematical model built based on the physical imaging principle. The model includes illumination intensity input parameters, diffuse reflection coefficient and specular reflection coefficient corresponding to the track surface state type, transmittance parameters of the camera optical system, and photoelectric conversion efficiency parameters of the image sensor. The parameters are connected in series through multiplication operations to output the expected grayscale value of each pixel of the image sensor under a given exposure time.

[0096] In this embodiment, minimizing the mean square error between the predicted image brightness distribution and the ideal brightness distribution is the objective function. With the exposure time adjustment range as the constraint, the gradient descent algorithm is used to dynamically solve for the optimal exposure time. This means taking the ideal brightness distribution as the standard reference image, calculating the sum of the squares of the gray-level differences of each pixel between the predicted brightness distribution and the ideal brightness distribution, and then taking the average as the error function. The minimum and maximum allowable values ​​of the exposure time are used as the constraint boundaries. Starting from the current exposure time, the algorithm iteratively searches in the opposite direction of the gradient of the error function until it finds the exposure time value that minimizes the error function.

[0097] In this embodiment, adjusting the exposure time of the visible light binocular camera unit according to the optimal exposure time means converting the optimal exposure time value obtained by the gradient descent algorithm into a parameter format that the camera control register can recognize, and writing it into the exposure time control registers of the left and right cameras through the serial communication interface, so that the camera can use the new exposure time parameter in subsequent frame acquisition.

[0098] In this embodiment, frequency domain analysis is performed on the real-time images acquired by the visible light binocular camera unit. The calculation of the main frequency and amplitude of the light intensity fluctuation refers to extracting the data of the gray value of each pixel changing over time in the continuously acquired image sequence, converting the time domain signal into a frequency domain signal through fast Fourier transform, finding the frequency component with the largest amplitude in the spectrum as the main frequency, and the amplitude of this frequency component as the amplitude of the light intensity fluctuation.

[0099] In this embodiment, the gain value is dynamically adjusted according to the light intensity fluctuation frequency. Maintaining the maximum image signal-to-noise ratio within the light intensity fluctuation cycle means using the main frequency obtained from frequency domain analysis as the basis for gain adjustment. When the fluctuation frequency is high, the gain value is appropriately reduced to avoid amplifier saturation distortion. When the fluctuation frequency is low, the gain value is appropriately increased to enhance the weak light signal, so that the gain value and the fluctuation frequency form an inverse relationship, and the signal-to-noise ratio is kept in the optimal range within a complete fluctuation cycle.

[0100] In this embodiment, the images captured by the left and right cameras are compared in real time. The calculation of the brightness difference and contrast difference of the binocular images refers to performing pixel-level registration of the left and right images, calculating the difference in gray values ​​of the corresponding registered pixels and averaging the absolute values ​​to obtain the brightness difference, and calculating the local gray-level variance of the left and right images respectively and then calculating the difference to obtain the contrast difference.

[0101] In this embodiment, the exposure time and gain of the left and right cameras are adjusted differently based on the differences in brightness and contrast. Maintaining the consistency of the binocular images means that when the brightness difference exceeds a preset threshold, the exposure time of the left or right camera is adjusted according to the positive or negative direction of the difference, so that the exposure time of the darker side is increased or the exposure time of the brighter side is decreased. When the contrast difference exceeds a preset threshold, the gain value of the left or right camera is adjusted according to the magnitude of the difference, so that the gain of the side with lower contrast is appropriately increased, so that the binocular images reach a level of near consistency in brightness and contrast.

[0102] In this embodiment, the gain dynamic adjustment subunit and the binocular image consistency correction subunit work together. First, the gain value is adjusted according to the frequency of light intensity fluctuations, and then differential fine-tuning is performed according to the differences in binocular images. This means that the system sets a two-level adjustment process. In the first level, the gain dynamic adjustment subunit sets the same reference gain value for the left and right cameras according to the frequency domain analysis results. In the second level, the binocular image consistency correction subunit fine-tunes the left and right cameras respectively based on the reference gain and the differences in brightness and contrast of the binocular images, so that the final output binocular image adapts to light fluctuations and maintains binocular consistency.

[0103] Furthermore, the near-infrared laser module includes: a penetration depth estimation subunit, which is used to query a pre-established surface state-penetration depth mapping table based on the orbital surface state type output by the four-dimensional situational awareness module, and estimate the penetration depth of the near-infrared laser under the current surface state; The power dynamic optimization subunit is used to dynamically solve the optimal output power of the near-infrared laser module based on the estimated penetration depth and the interference intensity level output by the three-factor coupled analysis module, with the optimization objective of maximizing the weighted sum of penetration depth and imaging signal-to-noise ratio, and adjust the output power of the near-infrared laser module according to the optimal output power. The pulse frequency adaptive adjustment subunit is used to analyze the images acquired by the visible light binocular camera unit in real time, calculate the fluctuation frequency of ambient light, and dynamically adjust the pulse frequency of the near-infrared laser module according to the fluctuation frequency of ambient light, so that the pulse frequency is staggered with the fluctuation frequency of ambient light to form a difference frequency and suppress ambient light interference. The laser speckle suppression subunit is used to perform time-varying modulation of the laser output from the near-infrared laser module. It employs random phase modulation in the time dimension and multimode fiber coupling output in the spatial dimension to suppress speckle noise formed by the near-infrared laser on the rail surface. The pulse frequency adaptive adjustment subunit works in conjunction with the laser speckle suppression subunit. First, the pulse frequency is adjusted to suppress ambient light interference, and then the laser is subjected to time-varying control to suppress speckle noise.

[0104] In this embodiment, the pre-established surface state-penetration depth mapping table is a two-dimensional lookup table stored in non-volatile memory. The row index of the table corresponds to the track surface state type, including rusted state, oil-covered state, snow-residual state, and clean and normal state. The column index of the table corresponds to the typical wavelength of near-infrared laser. Each cell in the table stores the laser penetration depth value obtained by experimental calibration under the corresponding surface state and wavelength conditions.

[0105] In this embodiment, based on the track surface state type output by the four-dimensional situational awareness module, a pre-established surface state-penetration depth mapping table is queried. The estimated penetration depth of the near-infrared laser under the current surface state is achieved by using the state type output by the track surface state identification unit as the query keyword, locating the row corresponding to the state type in the surface state-penetration depth mapping table, reading the cell value in that row corresponding to the current near-infrared laser wavelength, and using this value as the estimated depth at which the near-infrared laser can penetrate the surface covering layer and reach the rail substrate under the current track surface state.

[0106] In this embodiment, the penetration depth of the near-infrared laser under the current surface condition refers to the maximum depth to which the near-infrared laser beam can penetrate the rust layer, oil layer, or snow layer to reach the surface of the rail substrate after entering the rail surface from the air. The larger the depth value, the stronger the laser penetration ability, and the more effectively it can form clear structured light stripes on the surface of the rail substrate.

[0107] In this embodiment, based on the estimated penetration depth and the interference intensity level output by the three-factor coupled analysis module, the optimal output power of the near-infrared laser module is dynamically solved with the weighted sum of penetration depth and imaging signal-to-noise ratio as the optimization objective. This involves taking the estimated penetration depth and interference intensity level as input parameters, establishing a proportional function relationship where the penetration depth increases with the increase of output power, and establishing a function relationship where the imaging signal-to-noise ratio first increases and then saturates with the increase of output power. The two functions are multiplied by weighting coefficients and then added together to obtain the comprehensive objective function. The output power value that maximizes the comprehensive objective function is then searched within the minimum and maximum power ranges allowed by the laser module.

[0108] In this embodiment, adjusting the output power of the near-infrared laser module according to the optimal output power means converting the obtained optimal output power value into the control parameter of the laser diode drive current, adjusting the reference voltage of the drive circuit through the digital-to-analog converter, changing the injection current of the laser diode, so that the actual output optical power of the near-infrared laser module reaches the target value.

[0109] In this embodiment, the images acquired by the visible light binocular camera unit are analyzed in real time. The calculation of the ambient light fluctuation frequency refers to extracting the average gray value of the region of interest from the continuously acquired image sequence, taking the sequence of gray value changes over time as the time domain signal of ambient light intensity, performing a fast Fourier transform on the signal to obtain the spectrum, and finding the frequency component with the largest amplitude in the spectrum as the ambient light fluctuation frequency.

[0110] In this embodiment, the pulse frequency of the near-infrared laser module is dynamically adjusted according to the ambient light fluctuation frequency to create a difference frequency between the pulse frequency and the ambient light fluctuation frequency. Suppressing ambient light interference means setting the pulse operating frequency of the laser module to differ from the ambient light fluctuation frequency by a certain value. For example, the laser pulse frequency is set to the ambient light fluctuation frequency plus or minus a fixed offset, so that the laser signal and the ambient light signal are separated in the frequency domain, which facilitates the subsequent image processing algorithm to extract the laser light stripe signal through frequency domain filtering.

[0111] In this embodiment, the laser output from the near-infrared laser module is time-controlled. Random phase modulation is used in the time dimension, and multimode fiber coupling is used in the spatial dimension. Suppressing speckle noise formed by near-infrared laser on the rail surface refers to modulating the phase of the driving current of the laser diode through a random signal generator in the time dimension, so that the phase of the laser pulse changes randomly between pulses, thus destroying the coherence of the laser. In the spatial dimension, the laser is coupled into a multimode fiber, and the optical path difference caused by the different transmission path lengths of different modes in the multimode fiber is used to further reduce the coherence of the laser. The two modulation methods work together to reduce the contrast of the speckle pattern formed by the laser on the rough rail surface.

[0112] In this embodiment, the pulse frequency adaptive adjustment subunit and the laser speckle suppression subunit work together. First, the pulse frequency is adjusted to suppress ambient light interference. Then, the laser is subjected to time-controlled regulation to suppress speckle noise. This means that the system sets a two-stage processing flow. In the first stage, the pulse frequency adaptive adjustment subunit sets the laser pulse frequency according to the ambient light fluctuation frequency, so that the laser signal is separated from the ambient light in the frequency domain. In the second stage, the laser speckle suppression subunit performs random phase modulation and multimode fiber coupling on the laser with the set pulse frequency, further optimizing the laser characteristics in the time and spatial domains. Finally, the laser projected onto the rail surface can resist ambient light interference and has low speckle noise.

[0113] Furthermore, the 3D reconstruction and contour output module includes: a multimodal feature point extraction subunit, used to extract the center point of visible light structured light stripes based on the anti-interference enhanced visible light image data, and to extract the center point of near-infrared structured light stripes based on the anti-interference enhanced near-infrared image data, so as to obtain visible light structured light spot clouds and near-infrared structured light spot clouds; The point cloud quality assessment subunit is used to assess the quality of visible light structured light spot clouds and near-infrared structured light spot clouds respectively, and calculates the point cloud density, point cloud signal-to-noise ratio, and point cloud edge sharpness as quality assessment indicators. The fusion weight dynamic calculation subunit is used to dynamically calculate the fusion weight of visible light point cloud and near-infrared point cloud based on the interference intensity level output by the anti-interference strategy decision unit and the quality evaluation index output by the point cloud quality evaluation subunit, using the fuzzy comprehensive evaluation method. The higher the interference intensity level, the greater the weight of near-infrared point cloud; the higher the signal-to-noise ratio of visible light point cloud, the greater the weight of visible light point cloud. The point cloud fusion and reconstruction sub-unit is used to perform weighted fusion of visible light structured light spot clouds and near-infrared structured light spot clouds based on dynamically calculated fusion weights to obtain fused point clouds. The fused point cloud is smoothed by the moving least squares method, and the three-dimensional mesh model of the rail surface is reconstructed by the triangulation algorithm. The contour geometry parameter extraction sub-unit is used to extract the rail cross-sectional contour line from the 3D mesh model of the rail surface, calculate the rail head width, rail web height, rail base width, and rail gauge angle geometry parameters, and output the extracted rail contour geometry parameters as the track contour detection result.

[0114] In this embodiment, the center point of the visible light structured light stripe is extracted based on the anti-interference enhanced visible light image data, and the center point of the near-infrared structured light stripe is extracted based on the anti-interference enhanced near-infrared image data. The visible light structured light spot cloud and the near-infrared structured light spot cloud are obtained by scanning the enhanced visible light image line by line to find the position of the maximum gray value of the laser light stripe, calculating the center coordinates of the light stripe with sub-pixel accuracy through the gray-level centroid method, and converting the image coordinates into three-dimensional coordinates in the camera coordinate system to obtain the visible light point cloud data. The enhanced near-infrared image is processed in the same way to obtain the near-infrared point cloud data.

[0115] In this embodiment, the quality of visible light structured light spot clouds and near-infrared structured light spot clouds is evaluated separately. The point cloud density, point cloud signal-to-noise ratio, and point cloud edge sharpness are calculated as quality evaluation indicators. The point density indicator is the number of points per unit area. The signal-to-noise ratio indicator is the ratio of the mean gray value of the signal points in the point cloud to the standard deviation of the background noise. The edge sharpness indicator is the gray gradient amplitude of the point cloud at the edge of the rail. The three indicators together reflect the data quality of the point cloud.

[0116] In this embodiment, based on the interference intensity level output by the anti-interference strategy decision unit and the quality assessment index output by the point cloud quality assessment subunit, the fusion weight of the visible light point cloud and the near-infrared point cloud is dynamically calculated using a fuzzy comprehensive evaluation method. The higher the interference intensity level, the greater the weight of the near-infrared point cloud; the higher the signal-to-noise ratio of the visible light point cloud, the greater the weight of the visible light point cloud. This means that the interference intensity level and the three quality assessment indices are used as input variables of the fuzzy system, converted into fuzzy sets through a membership function, and inference is performed based on a preset fuzzy rule base. When the interference intensity level is high, the weight coefficient of the near-infrared point cloud is increased; when the signal-to-noise ratio of the visible light point cloud is high, the weight coefficient of the visible light point cloud is increased. After defuzzification, the fusion weight values ​​of the two point clouds are obtained.

[0117] In this embodiment, based on dynamically calculated fusion weights, visible light structured light spot clouds and near-infrared structured light spot clouds are weighted and fused to obtain a fused point cloud. The fused point cloud is then smoothed using the moving least squares method. The three-dimensional mesh model of the rail surface is reconstructed using the triangulation algorithm, which means that the coordinates of corresponding spatial positions in the visible light point cloud and the near-infrared point cloud are weighted and averaged to obtain the fused point coordinates. The moving least squares method is used to fit the local surface of all fused points and the point positions are adjusted to remove noise. The Delaunay triangulation algorithm is used to connect adjacent points to form a triangular mesh, thus constructing a three-dimensional mesh model of the rail surface.

[0118] In this embodiment, the rail cross-sectional contour line is extracted from the three-dimensional mesh model of the rail surface, and the geometric parameters of rail head width, rail web height, rail base width, and rail gauge angle are calculated. The extracted rail contour geometric parameters are output as the track contour detection result. This means that a slice is cut on the three-dimensional mesh model along the direction perpendicular to the rail length to obtain the cross-sectional contour line, the contour line is geometrically identified, the horizontal distance between the left and right points of the top of the rail head is calculated as the rail head width, the vertical height of the narrowest point of the rail web is calculated as the rail web height, the horizontal distance between the left and right points of the rail base is calculated as the rail base width, and the angle at the connection between the side of the rail head and the rail web is calculated as the rail gauge angle. These geometric parameters are output in numerical form as the final detection result.

[0119] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A 3D visual track contour detection system resistant to sunlight interference, characterized in that, include: The four-dimensional situational awareness module includes a light sensor array, a spectral analysis unit, and a track surface state recognition unit; the light sensor array is used to collect ambient light intensity and direction in real time; the spectral analysis unit is used to identify the type of sunlight interference based on the spectral distribution. The track surface condition recognition unit is used to perform image recognition based on image data from a visible light image sensor, extract track surface texture features, and identify the track surface condition type. The three-factor coupling analysis module is used to receive data on light intensity, light direction, sunlight interference type, and track surface state type output by the four-dimensional situational awareness module. It calculates the first coupling coefficient between sunlight incident angle and track surface reflection characteristics, the second coupling coefficient between sunlight incident angle and camera imaging attitude, and the third coupling coefficient between track surface reflection characteristics and camera imaging attitude through a pre-trained coupling analysis model. The multimodal imaging and hierarchical anti-interference module includes a visible light binocular camera unit, a near-infrared laser module, an adjustable polarization filter unit, and an anti-interference strategy decision unit. The anti-interference strategy decision unit is used to determine the dominant interference factor and the interference intensity level based on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient output by the three-factor coupling analysis module. According to the dominant interference factor and the interference intensity level, it matches the corresponding collaborative anti-interference strategy from the preset anti-interference strategy library and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data. The 3D reconstruction and contour output module is used to perform multimodal fusion point cloud reconstruction based on anti-interference enhanced visible light image data and near-infrared image data, extract the 3D contour of the rail surface and output the track contour detection results.

2. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The track surface condition recognition unit includes: The texture feature extraction subunit is used to extract the texture features of the track surface based on the image data of the visible light image sensor, and to calculate the texture roughness of the rusted state, the gloss feature value of the oil-covered state, and the edge sharpness feature value of the snow-residual state. The multi-dimensional quantization sub-unit is used to input the texture roughness of the rusted state, the gloss feature value of the oil-covered state, and the edge sharpness feature value of the snow-residual state into the pre-trained surface state quantization model to obtain the rusted state coefficient, the oil-covered state coefficient, and the snow-residual state coefficient. The surface state fusion subunit is used to calculate the confidence distribution of the fused track surface state based on the rust state coefficient, oil stain coverage state coefficient, and snow residue state coefficient using the evidence theory method. The surface state type with the highest confidence is output as the track surface state type data to the three-factor coupled analysis module.

3. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The three-factor coupling analysis module includes: The historical data organization subunit is used to organize the lighting conditions, track surface conditions and imaging quality recorded in the historical detection data in a multi-dimensional manner according to the monitoring time point, monitoring section location and feature parameter type, to form a time series dataset of lighting conditions, a spatial distribution dataset of track surface conditions and a set of imaging quality feature parameters. The coupled graph neural network constructs a subunit, which is used to establish a three-node dynamic coupled graph structure based on the time series dataset of illumination conditions, the spatial distribution dataset of track surface state, and the set of imaging quality feature parameters. The graph neural network is used to establish a three-node dynamic coupled graph structure with the sunlight incident angle as the first node, the track surface reflection characteristics as the second node, and the camera imaging attitude as the third node. The connection weights in the dynamic coupled graph structure are dynamically updated according to the real-time monitoring data. The multi-scale coupled analytical subunit is used to learn the short-term fluctuation coupling relationship and long-term trend coupling relationship between the first node and the second node at different time scales through a dynamic coupled graph structure. The short-term fluctuation coupling relationship and the long-term trend coupling relationship are weighted and fused to obtain the first coupling coefficient. The dynamic coupled graph structure is used to learn the static geometric coupling relationship and dynamic motion coupling relationship between the first node and the third node at different spatial scales through a dynamic coupled graph structure. The static geometric coupling relationship and the dynamic motion coupling relationship are weighted and fused to obtain the second coupling coefficient. The dynamic coupled graph structure is used to learn the material property coupling relationship and imaging physics coupling relationship between the second node and the third node at different physical scales through a dynamic coupled graph structure. The material property coupling relationship and the imaging physics coupling relationship are weighted and fused to obtain the third coupling coefficient. The first coupling coefficient, the second coupling coefficient, and the third coupling coefficient are output to the anti-interference strategy decision unit.

4. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The anti-interference strategy decision unit includes: The coupling coefficient time series analysis subunit is used to perform time series analysis on the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient, and to calculate the mean, variance, rate of change, and fluctuation frequency of each coupling coefficient within a preset time window. The interference intensity dynamic grading subunit is used to dynamically divide the coupling intensity of the three factors into three levels: mild interference level, moderate interference level, and severe interference level based on the current value, variance, and rate of change of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient using fuzzy logic. The grading threshold is adaptively adjusted according to the statistical distribution of the coupling coefficient. The hierarchical trigger condition mapping subunit is used to establish the mapping relationship between mild interference level and single anti-interference measure trigger condition, moderate interference level and dual-channel collaborative anti-interference measure trigger condition, and severe interference level and three-channel collaborative anti-interference measure trigger condition. The mapping relationship is stored in the storage space accessible by the strategy matching process, and the interference intensity level is output to the strategy matching process of the anti-interference strategy decision unit.

5. The 3D visual track contour detection system resistant to sunlight interference according to claim 4, characterized in that, The anti-interference strategy decision unit includes: The coupling coefficient contribution calculation subunit is used to calculate the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient to the overall interference intensity. The contribution is obtained by weighting the normalized value and weight coefficient of each coupling coefficient. The dominant factor dynamic identification subunit is used to compare the contribution of the first coupling coefficient, the second coupling coefficient, and the third coupling coefficient. When the contribution of the first coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the sunlight incident angle and the reflection characteristics of the track surface is determined to be the dominant interference factor. When the contribution of the second coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the sunlight incident angle and the camera imaging attitude is determined to be the dominant interference factor. When the contribution of the third coupling coefficient is the largest and exceeds the preset dominant threshold, the coupling between the reflection characteristics of the track surface and the camera imaging attitude is determined to be the dominant interference factor. The confidence assessment subunit for dominant factors is used to calculate the confidence of identifying dominant interference factors based on the contribution distribution. When no single coupling coefficient has a contribution exceeding the preset dominant threshold, it is determined to be a multi-factor composite interference mode. The coupling relationships corresponding to all coupling coefficients whose contributions exceed the preset composite threshold are taken as composite interference dominant factors. The composite interference mode and composite interference dominant factors are output as interference dominant factors to the strategy matching process of the anti-interference strategy decision unit.

6. The 3D visual track contour detection system resistant to sunlight interference according to claim 5, characterized in that, The preset anti-interference strategy library adopts a layered storage structure, including a basic strategy layer, a collaborative strategy layer, and an emergency strategy layer; The basic strategy layer stores individual adjustment strategies for the adjustable polarization filter unit, individual adjustment strategies for the exposure time of the visible light binocular camera unit, and individual adjustment strategies for the power of the near-infrared laser module. The collaborative strategy layer stores dual-channel collaborative strategies for adjusting the adjustable polarization filter unit and the exposure time of the visible light binocular camera unit, dual-channel collaborative strategies for adjusting the adjustable polarization filter unit and the power of the near-infrared laser module, and dual-channel collaborative strategies for adjusting the exposure time of the visible light binocular camera unit and the power of the near-infrared laser module. The emergency strategy layer stores a three-channel collaborative strategy: adjustable polarization filter unit adjustment, visible light binocular camera unit exposure time adjustment, and near-infrared laser module power adjustment. The collaborative strategy layer also stores multi-channel collaborative strategies corresponding to multi-factor composite interference modes. The multi-channel collaborative strategies corresponding to multi-factor composite interference modes are multi-channel collaborative strategies that simultaneously include the adjustment measures corresponding to all composite interference dominant factors. The anti-interference strategy decision unit is also used to receive the output interference intensity level, and determine the target strategy layer corresponding to the current interference intensity level based on the established mapping relationship between the mild interference level and the triggering conditions of the basic strategy layer, the mapping relationship between the moderate interference level and the triggering conditions of the cooperative strategy layer, and the mapping relationship between the severe interference level and the triggering conditions of the emergency strategy layer. At the same time, it receives the output interference dominant factor or composite interference mode and composite interference dominant factor. When the interference mode is a single interference mode, the anti-interference strategy decision unit matches the cooperative anti-interference strategy corresponding to the interference dominant factor from the target strategy layer according to the interference dominant factor and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data. When the interference mode is a multi-factor composite interference mode, the anti-interference strategy decision unit matches the multi-channel collaborative strategy corresponding to the multi-factor composite interference mode from the collaborative strategy layer according to the dominant factor of composite interference and executes it to obtain the anti-interference enhanced visible light image data and near-infrared image data.

7. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The adjustable polarization filter unit includes: The solar reflected light polarization characteristic analysis subunit is used to calculate the polarization direction and degree of polarization of the sunlight reflected on the rail surface based on the illumination direction and spectral distribution collected by the four-dimensional situational awareness module and the Fresnel equation. The polarization direction dynamic tracking subunit is used to predict the polarization direction of the next moment based on the changing trajectory of the polarization direction of the sunlight reflected light using the Kalman filter algorithm, and adjust the polarization direction of the adjustable polarization filter unit in real time to make the polarization direction orthogonal to the predicted polarization direction of the reflected light. The polarization degree adaptive adjustment subunit is used to calculate the optimal polarization degree that maximizes the image signal-to-noise ratio under the current lighting conditions based on the light intensity collected by the four-dimensional situational awareness module, and adjust the polarization degree of the adjustable polarization filter unit in real time to match the transmittance with the imaging requirements. The multi-region independent adjustment subunit is used to divide the adjustable polarization filter unit into multiple independent adjustment regions. Based on the brightness distribution of different regions in the image acquired by the visible light binocular camera unit and the interference intensity level output by the anti-interference strategy decision unit, the polarization direction and polarization degree of each independent adjustment region are adjusted respectively.

8. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The visible light binocular camera unit includes: The illumination-surface joint modeling subunit is used to establish an illumination-surface joint reflection model based on the illumination intensity and orbital surface state type collected by the four-dimensional situational awareness module, and to predict the image brightness distribution under different exposure times. The exposure time optimization solution subunit is used to minimize the mean square error between the predicted image brightness distribution and the ideal brightness distribution as the objective function, with the exposure time adjustment range as the constraint condition, and dynamically solves the optimal exposure time using the gradient descent algorithm. The exposure time of the visible light binocular camera unit is then adjusted according to the optimal exposure time. The gain dynamic adjustment subunit is used to perform frequency domain analysis on the real-time images acquired by the visible light binocular camera unit, calculate the main frequency and amplitude of the light intensity fluctuation, and dynamically adjust the gain value according to the light intensity fluctuation frequency to maximize the image signal-to-noise ratio within the light intensity fluctuation period. The binocular image consistency correction subunit is used to compare the images acquired by the left and right cameras in real time, calculate the brightness and contrast differences of the binocular images, and adjust the exposure time and gain of the left and right cameras differently based on the brightness and contrast differences to maintain the consistency of the binocular images. The gain dynamic adjustment subunit works in conjunction with the binocular image consistency correction subunit. First, the gain value is adjusted according to the frequency of light intensity fluctuations, and then differential fine-tuning is performed according to the differences in the binocular images.

9. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The near-infrared laser module includes: The penetration depth prediction subunit is used to query a pre-established surface state-penetration depth mapping table based on the orbital surface state type output by the four-dimensional situational awareness module, and to predict the penetration depth of the near-infrared laser under the current surface state. The power dynamic optimization subunit is used to dynamically solve the optimal output power of the near-infrared laser module based on the estimated penetration depth and the interference intensity level output by the three-factor coupled analysis module, with the optimization objective of maximizing the weighted sum of penetration depth and imaging signal-to-noise ratio, and adjust the output power of the near-infrared laser module according to the optimal output power. The pulse frequency adaptive adjustment subunit is used to analyze the images acquired by the visible light binocular camera unit in real time, calculate the fluctuation frequency of ambient light, and dynamically adjust the pulse frequency of the near-infrared laser module according to the fluctuation frequency of ambient light, so that the pulse frequency is staggered with the fluctuation frequency of ambient light to form a difference frequency and suppress ambient light interference. The laser speckle suppression subunit is used to perform time-varying modulation of the laser output from the near-infrared laser module. It employs random phase modulation in the time dimension and multimode fiber coupling output in the spatial dimension to suppress speckle noise formed by the near-infrared laser on the rail surface. The pulse frequency adaptive adjustment subunit works in conjunction with the laser speckle suppression subunit. First, the pulse frequency is adjusted to suppress ambient light interference, and then the laser is time-controlled to suppress speckle noise.

10. The 3D visual track contour detection system resistant to sunlight interference according to claim 1, characterized in that, The 3D reconstruction and contour output module includes: The multimodal feature point extraction subunit is used to extract the center point of visible light structured light stripes based on the anti-interference enhanced visible light image data, and to extract the center point of near-infrared structured light stripes based on the anti-interference enhanced near-infrared image data, so as to obtain visible light structured light spot clouds and near-infrared structured light spot clouds. The point cloud quality assessment subunit is used to assess the quality of visible light structured light spot clouds and near-infrared structured light spot clouds respectively, and calculates the point cloud density, point cloud signal-to-noise ratio, and point cloud edge sharpness as quality assessment indicators. The fusion weight dynamic calculation subunit is used to dynamically calculate the fusion weight of visible light point cloud and near-infrared point cloud based on the interference intensity level output by the anti-interference strategy decision unit and the quality evaluation index output by the point cloud quality evaluation subunit, using the fuzzy comprehensive evaluation method. The higher the interference intensity level, the greater the weight of near-infrared point cloud; the higher the signal-to-noise ratio of visible light point cloud, the greater the weight of visible light point cloud. The point cloud fusion and reconstruction sub-unit is used to perform weighted fusion of visible light structured light spot clouds and near-infrared structured light spot clouds based on dynamically calculated fusion weights to obtain fused point clouds. The fused point cloud is smoothed by the moving least squares method, and the three-dimensional mesh model of the rail surface is reconstructed by the triangulation algorithm. The contour geometry parameter extraction sub-unit is used to extract the rail cross-sectional contour line from the 3D mesh model of the rail surface, calculate the rail head width, rail web height, rail base width, and rail gauge angle geometry parameters, and output the extracted rail contour geometry parameters as the track contour detection result.

Images