Railway track traffic scene segmentation processing method driven by prior knowledge

Abstract

The application discloses a railway track traffic scene segmentation processing method under prior knowledge driving, comprising the following steps: S1, point cloud data voxel filtering; S2, fitting plane filtering of the point cloud data screened in S1; S3, obtaining a new optimal fitting plane based on the point cloud data filtered in S2 and updating a reference frame; rotating the original point cloud to the reference frame; rotating the point cloud filtered in S2 to the reference frame to obtain reference point cloud; S4, setting a railway segmentation length; obtaining a segmentation mask for segmenting the original point cloud and the reference point cloud adjusted in pose obtained in S3, and constructing a corresponding scene index; adjusting the scene index of the original point cloud segmented and with a point number lower than the lower limit of the quartile to the index value of the adjacent scene; S5, local scene pose adaptive adjustment. The scene segmentation strategy can be applied to large-scale track traffic scene point cloud processing, and feature deviation is reduced through dynamic adaptive calibration.

Description

Technical Field

[0001] This invention relates to the field of point cloud processing, specifically to a method for segmented processing of railway track traffic scenarios driven by prior knowledge. Background Technology

[0002] In railway rail transit scenarios, mobile laser scanning (MLS) technology can effectively acquire large-scale 3D point cloud data, providing crucial support for the monitoring and maintenance of railway infrastructure. However, due to the massive scale and complex structure of point cloud data, traditional point cloud processing methods suffer from excessive computational resource consumption and low efficiency. When real-time and large-scale processing is required, conventional segmentation methods struggle to handle the dynamic changes and geometric diversity of railway scenarios. Furthermore, the curvature and undulations of railway tracks increase the complexity of processing methods, making it impossible for traditional methods to accurately capture the continuity and detailed features of the track, thus affecting the subsequent processing and analysis results.

[0003] The inconsistency between the railway scene reference coordinate system and the processing coordinate system mainly stems from two factors: the randomness of the pose during data acquisition and the change in the acquisition direction. The former often results from neglecting the impact of scene pose on the processing during acquisition, while the latter is closely related to the MLS (Multi-Site Observation System) route. Due to changes in the track's forward direction and unavoidable terrain variations during operation, the MLS will experience dynamic changes such as undulations during acquisition. Figure 1 As shown in (a) and (b), this spatial reference deviation will cause cumulative errors in subsequent processing steps.

[0004] In addition, MLS acquisition results exhibit spatial resolution gradient variation characteristics, from Figure 2 It can be seen that in the near-field area of ​​the equipment, the point cloud density can reach up to 2000 points / square meter, which can clearly show the track details. However, when the scanning distance extends to more than 50 meters, the acquisition density rapidly decreases to below 50 points / square meter, and the geometric information of distant ground features is difficult to retain completely. This distribution pattern causes the geometric boundary between rail transit facilities and the surrounding environment to be less distinguishable in distant areas, which is more obvious in complex terrain sections. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a segmented processing method for railway track traffic scenarios driven by prior knowledge. This method can achieve optimal accuracy and ensure a certain level of testing efficiency.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0007] A method for segmented processing of railway rail transit scenarios driven by prior knowledge includes the following steps:

[0008] S1. The original point cloud of the railway track traffic scene is divided by the spatial voxel mean filtering method, the discrete point cloud is transformed into several voxel units, and then the voxelized point cloud is filtered based on the density threshold filtering mechanism.

[0009] S2, perform covariance matrix decomposition on the point cloud data filtered in S1 using principal component analysis to obtain the optimal fitting plane; establish a reference frame based on the optimal fitting plane, and calculate the distance from the points in the point cloud filtered in S1 to the optimal fitting plane; use an adaptive pass-through filter to filter the point cloud obtained in S1 based on the distance from the points to the optimal fitting plane.

[0010] S3, For the point cloud data filtered in S2, use the same method as in S2 to obtain a new optimal fitting plane and update the reference frame; calculate the rotation angle between the original point cloud and the new optimal fitting plane, rotate the original point cloud into the updated reference frame; rotate the point cloud filtered in S2 into the updated reference frame to obtain the reference point cloud;

[0011] S4, set the railway segment length; obtain the segment mask; segment the original point cloud after attitude adjustment in S3 and the reference point cloud obtained in S3 according to the segment mask, and construct the corresponding scene index; perform further statistical analysis on the number of points in each segment of the original point cloud; if the number of points is lower than the lower quartile limit, adjust the corresponding scene index to the index value of the adjacent scene to realize the merging of very small-scale scenes;

[0012] S5 performs adaptive adjustment of the pose in local scenes.

[0013] The method for obtaining the segment mask in S4 above is as follows: set the railway segment length; determine the main direction based on the bounding box of the reference point cloud obtained in S3, and calculate the number of segments based on the set segment length; construct a moving window with the lowest point of the main direction as the starting point and the segment length as the moving step size, wherein the moving window only moves on the axis where the main direction is located; obtain the segment mask based on the constructed moving window.

[0014] The length of the aforementioned railway section is 50m.

[0015] The above S5 includes the following steps:

[0016] S51, The local reference point cloud is used to construct an affine transformation matrix to achieve preliminary alignment of the local pose;

[0017] S52, the segmented point cloud obtained in S51 is discretized into several voxels in three-dimensional space and point cloud mean filtering is performed to suppress outliers.

[0018] S53, the local reference point cloud is mapped to the XOY plane through projection transformation to achieve dimensionality reduction expression to the two-dimensional feature space. Then, two-dimensional spatial boundary constraint analysis is performed along the X-axis and Y-axis directions respectively, and prominent noise points are filtered out through a dynamic threshold mechanism.

[0019] S54, the same calculation method as in S2 is used to obtain a new optimal fitting plane and a corresponding reference system for the local reference point cloud after boundary analysis and fitting plane filtering; the original point cloud after segmentation in S4 is rotated to the corresponding reference system, and the local pose adjustment result of the segmented scene is output.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. The unified spatial reference coordinate system and scene segmentation strategy of the present invention are easy to process and can meet the challenges of large-scale rail transit scene point cloud processing;

[0022] 2. This invention can perform dynamic adaptive calibration, reduce feature offset caused by differences in the original reference coordinate system of the scene, and ensure that the prior knowledge learned by the model can be quickly applied to all scenes;

[0023] 3. In this invention, the railway spatial processing method reduces computational complexity by dividing the space domain; it establishes structured data units based on facility topology, which not only improves efficiency but also enhances the targeting of feature extraction by explicitly expressing spatial relationships. Attached Figure Description

[0024] Figure 1 The reasons for the unstable scene pose include (a) random pose and (b) undulating orbit.

[0025] Figure 2 A diagram illustrating the characteristics of near-density and far-density spacing in MLS railway rail transit scenarios;

[0026] Figure 3 This is a diagram showing the differences in axial features of the scene after manual correction.

[0027] Figure 4 A spatial relationship diagram between railway facilities in multiple scenarios;

[0028] Figure 5 This is a flowchart of the segmented processing method of the present invention;

[0029] Figure 6 The posture evaluation index provided by this invention is illustrated in the figure.

[0030] Figure 7 The angle of deviation between the processed scene and the true scene coordinate axes provided by this invention;

[0031] Figure 8The before-and-after comparison of the overall posture adjustment provided by this invention;

[0032] Figure 9 The final processed result provided by this invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0034] Example

[0035] A method for segmented processing of railway rail transit scenarios driven by prior knowledge includes the following steps:

[0036] S1, voxel filtering of point cloud data:

[0037] In rail transit scenarios, the dynamic changes in scanning distance cause significant density fluctuations and discrete noise in the original point cloud, a spatial heterogeneity particularly pronounced during MLS scanning. To address this, this invention first utilizes spatial voxel mean filtering to partition the original point cloud, transforming the discrete point cloud into several voxel units. Then, a density threshold filtering mechanism is used to filter the voxelized point cloud, suppressing noise interference in the point cloud data.

[0038] In practical applications, the voxel unit size is preset according to the scene density, key geometric details are retained and invalid data in low-density areas are removed to form a filtered point cloud. In this embodiment, the voxel unit used for overall attitude adjustment is a cube with a side length of 0.5m.

[0039] S2, fitted plane filtering, as follows:

[0040] First, the point cloud data obtained from S1 is decomposed into covariance matrix using principal component analysis (PCA) to obtain the optimal fitting plane.

[0041] Then, a reference frame is established based on the optimal fitting plane, and the distance from the points in the point cloud obtained by S1 to the optimal fitting plane is calculated. Scattered points deviating from the optimal fitting plane generally correspond to environmental noise or non-track facilities. This operation can capture the geometric features of planar structures such as track surfaces and track beds.

[0042] The point cloud obtained in S1 is then filtered using an adaptive pass-through filter based on the distance from the point to the optimal fitting plane, including:

[0043] Sort the distances from the points to the best-fit plane in ascending order;

[0044] Calculate the lower quartile Q1 and the upper quartile Q3, then the interquartile range IQR = Q3 - Q1;

[0045] The thresholds for identifying outliers are as follows: the lower boundary (lower limit) is Q1 - 1.5 × IQR; the upper boundary (upper limit) is Q3 + 1.5 × IQR.

[0046] Outlier identification: The sorted data is compared with the lower and upper boundaries, and the data in between is retained as the filtered point cloud.

[0047] Compared to fixed threshold schemes, the above methods can retain key reference information and effectively remove background noise in complex terrain scenes.

[0048] S3, Overall Scene Posture Adjustment:

[0049] For the point cloud data after S2 filtering, a new optimal fitting plane is obtained and the reference reference system is updated using the same method as in S2; the rotation angle between the original point cloud and the new optimal fitting plane is calculated, and the original point cloud is rotated into the updated reference reference system (coordinate system) to complete the attitude adjustment of the original point cloud data; the point cloud after S2 filtering is rotated into the updated reference reference system (coordinate system) to obtain the reference point cloud.

[0050] Adaptive attitude adjustment can effectively eliminate errors caused by inconsistent scanning device attitude or terrain changes, making point cloud data more consistent with the geometry of the actual scene.

[0051] S4, Scene Segmentation, includes the following steps:

[0052] The length of each railway segment is set. In this embodiment, the segment length is set to 50m.

[0053] The main direction is determined based on the bounding box of the reference point cloud obtained in S3, and the number of segments is calculated based on the set segment length. The moving window is constructed with the lowest point of the main direction as the starting point and the segment length as the moving step size, wherein the moving window only moves on the axis where the main direction is located.

[0054] Finally, a segmentation mask is obtained based on the constructed moving window, which is used to segment the original point cloud after S3 pose adjustment to obtain a segmented original point cloud; the reference point cloud obtained in S3 is also segmented to obtain a local reference point cloud. Scene indices corresponding to the segmented reference point cloud and the original point cloud are also constructed. Furthermore, the number of points in each segmented original point cloud is further statistically analyzed. If the number of points is below the lower quartile limit, the corresponding scene index is adjusted to the index value of the adjacent scene to achieve the merging of very small-scale scenes.

[0055] S5, Local scene pose adaptive adjustment:

[0056] After the aforementioned processing, the bending and undulating phenomena in some scenes make it difficult for a unified spatial reference coordinate system to be effectively applied to the processing of different segmented results. Therefore, this invention designs a local scene attitude adaptive adjustment method, aiming to further adjust and optimize the scene according to the characteristics of the local railway section. Specifically, as follows:

[0057] S51 uses the local reference point cloud obtained in S4 to construct an affine transformation matrix to achieve preliminary alignment of the segmented point cloud poses, thereby ensuring the reliability of the geometric reference for subsequent processing.

[0058] S52 introduces spatial voxel mean filtering technology, which suppresses outliers by discretizing the segmented point cloud into several voxels after the initial local pose alignment is completed in S51 and applying point cloud mean filtering.

[0059] S53, Projection Transformation, Boundary Analysis.

[0060] The local reference point cloud is mapped to the XOY plane through projection transformation to achieve dimensionality reduction representation in a two-dimensional feature space. This processing strategy significantly simplifies the complexity of subsequent spatial analysis while focusing on extracting key geometric characteristics in each coordinate dimension. Specifically, the appropriate discarding of vertical direction information during projection can specifically enhance the analytical accuracy of core features such as track gauge and alignment, thereby better highlighting the local track extension direction information.

[0061] Based on the above projection transformation, two-dimensional spatial boundary constraint analysis is performed along the X-axis and Y-axis directions respectively, and prominent noise is filtered out through a dynamic threshold mechanism.

[0062] S54, perform the same fitting plane filtering operation as in S2 on the local reference point cloud obtained from S53 after bounding analysis. Use the same calculation method as in S2 to obtain a new optimal fitting plane and corresponding reference frame for the local reference point cloud after bounding analysis and fitting plane filtering. Rotate the original point cloud after segmentation in S4 to the corresponding reference frame, and output the local pose adjustment result of the segmented scene.

[0063] To verify the technical effect of this invention, using a manually calibrated pose scene as a benchmark, and by leveraging the angular differences in the 3D scene before and after processing along each axis of the coordinate system, three key parameters, α, β, and γ, were defined. These parameters represent the degree of offset from the X, Y, and Z axes, respectively. Figure 6 As shown, by analyzing the numerical distribution of the three axial parameters, the spatial pose error at different processing stages can be clearly identified.

[0064] A systematic quantitative and qualitative evaluation of the optimization accuracy and applicability of this method in the attitude adjustment process is conducted. Details are as follows:

[0065] Quantitative analysis:

[0066] The processed scene coordinate axis deflection statistics are presented as follows: Figure 7 In the figure, the three attitude deviation angle error values ​​corresponding to each scene are the deviation values ​​from the X, Y, and Z axes, respectively. As can be seen from the figure, after segmented scene processing, the average attitude deviation angles of the scene relative to the X, Y, and Z axes are 0.16°, 2.13°, and 2.11°, respectively, which are within the acceptable range for attitude adjustment error. This indicates that the method of the present invention successfully reduces the attitude error caused by random attitude settings or the turning of the scanning vehicle on the XOY plane.

[0067] Qualitative analysis:

[0068] This invention selected five scenarios to demonstrate the before-and-after effects of overall posture adjustment, such as... Figure 8 As shown. The original railway scene posture data often exhibits a degree of randomness, making it difficult to align with the actual scene, thus affecting the utilization of prior knowledge and the accuracy of subsequent tasks. However, the three-view diagram after overall posture adjustment reveals that the adjusted coordinate system basically meets expectations: the X-axis is associated with the track extension direction, the Y-axis represents the railway cross-section information, and the Z-axis is perpendicular to the track plane.

[0069] Furthermore, this embodiment demonstrates the processing effects of different scenarios before and after local pose adjustment, specifically as follows: Figure 9 As shown, the railway scene will change to some extent with the change of direction. Even so, the method of the present invention can still effectively achieve adaptive adjustment of the local scene posture. This result shows that the method of the present invention has strong robustness in the processing process and can effectively overcome the challenges caused by environmental differences in the scene and maintain the accuracy of posture adjustment.

[0070] The impact of voxel size on processing performance and efficiency:

[0071] Voxel filters were used in both the overall attitude adjustment process and the local attitude adjustment process. The former aimed to reduce the impact of a sudden increase in local frame point cloud density caused by abrupt pauses in the process, while the latter aimed to remove noise using voxel filters to ensure the robustness of subsequent attitude adjustments. Therefore, this embodiment balances the processing performance and efficiency of the method by adjusting the voxel size parameters used in the voxel filters during the above processes. The processing results are shown in Table 1.

[0072] Table 1. Results of voxel size testing

[0073]

[0074] The results show that when the voxel filter size in the method decreases, although the average pose error angle decreases slightly, the method's execution time, especially the overall pose adjustment process, increases significantly. Simultaneously, when the voxel size is too small, the effectiveness of the voxel filter weakens considerably, causing some scene point clouds to affect the final adjustment result. Conversely, when the voxel size increases, although the method's execution time further decreases, the larger voxel size means that a large number of point clouds are not filtered out, thus affecting the pose adaptive adjustment process and causing a significant increase in the average pose error angle. Furthermore, when the voxel filter size increases during the local pose adjustment process, a small number of point clouds are also not effectively filtered out, resulting in a slight increase in the final local pose error angle.

[0075] In summary, based on the experimental results, the voxel filter size in the overall attitude adjustment process is set to 0.5m in this embodiment, and the voxel filter size in the local attitude adjustment process is set to 1.0m.

[0076] The effect of different segment lengths on the performance and efficiency of the method of the present invention:

[0077] This embodiment tests and evaluates the performance of the method under different segment lengths, and the test results are shown in Table 2.

[0078] Table 2 Test results for different segment lengths

[0079]

[0080] It can be seen that when the segment length is 10m, the method struggles to distinguish between the track-side direction and the track section direction, resulting in a sharp increase in the average deviation angle difference along the X and Y axes, reaching a maximum of 89.96. When the segment length is 90m, although the attitude adjustment effect is good in most scenarios, a small number of scenarios exhibit significant deviations in the Y and Z axis angles due to the influence of ground features on both sides of the track. Furthermore, as the set segment length increases, the processing time also increases, because the attitude of local scenes has the most profound impact on the method's processing efficiency, and large-scale local scenes corresponding to longer set segment lengths significantly exacerbate this inefficiency. In addition, shorter set segment lengths result in a scene count ratio of 14.61 before and after processing, significantly affecting the efficiency of subsequent method processing, and the model requires stronger robustness to adapt to a large number of segmented local scenes.

[0081] In summary, this embodiment takes into account the processing performance and efficiency of the method, as well as the impact of the processing length on the subsequent model, and selects 50m as the final segmented processing length.

[0082] The above is a detailed description of one embodiment of the present invention. The above embodiment cannot be used to limit the scope of protection of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for segmented processing of railway track traffic scenarios driven by prior knowledge, characterized in that, Includes the following steps: S1. The original point cloud of the railway track traffic scene is divided by the spatial voxel mean filtering method, the discrete point cloud is transformed into several voxel units, and then the voxelized point cloud is filtered based on the density threshold filtering mechanism. S2, perform covariance matrix decomposition on the point cloud data filtered in S1 using principal component analysis to obtain the optimal fitting plane; establish a reference frame based on the optimal fitting plane, and calculate the distance from the points in the point cloud filtered in S1 to the optimal fitting plane; use an adaptive pass-through filter to filter the point cloud obtained in S1 based on the distance from the points to the optimal fitting plane. S3. For the point cloud data filtered by S2, use the same method as in S2 to obtain a new optimal fitting plane and update the benchmark reference system. Calculate the rotation angle between the original point cloud and the new best-fit plane, and rotate the original point cloud to the updated reference frame; Rotate the S2-filtered point cloud to the updated reference frame to obtain the reference point cloud; S4, set the railway segment length; obtain the segment mask; segment the original point cloud after attitude adjustment in S3 and the reference point cloud obtained in S3 according to the segment mask, and construct the corresponding scene index; perform further statistical analysis on the number of points in each segment of the original point cloud; if the number of points is lower than the lower quartile limit, adjust the corresponding scene index to the index value of the adjacent scene to realize the merging of very small-scale scenes; S5 performs adaptive adjustment of the pose in local scenes.

2. The method for segmented processing of railway rail transit scenarios according to claim 1, characterized in that, The method for obtaining the segment mask in S4 is as follows: set the railway segment length; determine the main direction based on the bounding box of the reference point cloud obtained in S3, and calculate the number of segments based on the set segment length; construct a moving window with the lowest point of the main direction as the starting point and the segment length as the moving step size, wherein the moving window only moves on the axis where the main direction is located; obtain the segment mask based on the constructed moving window.

3. The method for segmented processing of railway rail transit scenarios according to claim 2, characterized in that, The railway section is 50m long.

4. The method for segmented processing of railway rail transit scenarios according to claim 1, characterized in that, S5 includes the following steps: S51, The local reference point cloud is used to construct an affine transformation matrix to achieve preliminary alignment of the local pose; S52, the segmented point cloud obtained in S51 is discretized into several voxels in three-dimensional space and point cloud mean filtering is performed to suppress outliers. S53, the local reference point cloud is mapped to the XOY plane through projection transformation to achieve dimensionality reduction expression to the two-dimensional feature space. Then, two-dimensional spatial boundary constraint analysis is performed along the X-axis and Y-axis directions respectively, and prominent noise points are filtered out through a dynamic threshold mechanism. S54: After boundary analysis and fitting plane filtering, the local reference point cloud is processed using the same calculation method as in S2 to obtain a new optimal fitting plane and the corresponding reference system; the original point cloud after segmentation described in S4 is rotated to the corresponding reference system, and the local pose adjustment result of the segmented scene is output.

Images