A combined positioning method for track inspection based on semantic map

Abstract

The present invention relates to the technical field of track inspection and positioning, and specifically discloses a combined track inspection and positioning method based on a semantic map. First, a lightweight real-time semantic segmentation neural network based on Deeplabv3+ is used to segment track images to obtain semantic labels for each pixel. Subsequently, through image-point cloud projection technology, these semantic labels are mapped into the point cloud data, and combined with the LIO-SAM algorithm and the obtained semantic information, a global track semantic map is constructed to provide sleeper position information for subsequent combined positioning. The local motion model provides local motion constraints for the combined positioning method. Although its accuracy is limited and there are cumulative errors, it can output continuous and stable motion information within a short distance range. Finally, based on the constructed track semantic map, the local motion is constrained and adjusted to achieve the fusion positioning of multi-source information.

Description

Technical Field

[0001] The present invention relates to the technical field of track inspection and positioning, and particularly to a combined track inspection and positioning method based on a semantic map. Background Art

[0002] High-precision and reliable positioning technology is crucial for track inspection. Although traditional odometer and satellite positioning technologies are widely used, they both have obvious limitations. Odometer positioning has the problem of cumulative error, which will cause the performance of the positioning system to degrade rapidly, resulting in large errors and seriously affecting the positioning accuracy. On the other hand, in special environments such as under viaducts and tunnels, the satellite positioning ability will be greatly reduced, resulting in the loss of positioning information. With the continuous growth of track inspection requirements, the requirements for the reliability and accuracy of the positioning system are also increasing day by day. In this context, a positioning system relying solely on a single information source has been difficult to meet the needs of modern railway track operation and maintenance.

[0003] Existing track inspection equipment usually uses the method of continuous pushing and trigger sampling for positioning and track status data collection. However, there are significant cumulative errors in the positioning mileage of these devices at present, and the mileage error can even reach more than one meter. The existing solution is to add markers such as RFID in the track environment to reduce the cumulative error. However, this method is time-consuming and laborious, and cannot meet the positioning requirements of large-scale positioning and non-visualization. As an important part of the track structure, sleepers are discretely distributed at equal distances along the track length direction. Using this characteristic, it has become an important reference for track positioning. Therefore, the present invention visualizes the positioning information of global sleepers by constructing a semantic map, effectively constrains the local motion model, and obtains a more accurate positioning result. Summary of the Invention

[0004] To solve the problems existing in the prior art, the present invention provides a combined track inspection and positioning method based on a semantic map. The present invention uses a lightweight real-time semantic segmentation neural network based on Deeplabv3+ to segment track images to obtain semantic labels for each pixel. Subsequently, through image-point cloud projection technology, these semantic labels are mapped into the point cloud data, and combined with the LIO-SAM algorithm and the obtained semantic information, a global track semantic map is constructed to provide sleeper position information for subsequent combined positioning. The local motion model provides local motion constraints for the combined positioning method. Although its accuracy is limited and there are cumulative errors, it can output continuous and stable motion information within a short distance range. Finally, based on the constructed track semantic map, the local motion is constrained and adjusted to achieve the fusion positioning of multi-source information, solving the problems mentioned in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solution: a track inspection combined positioning method based on semantic map, comprising the following steps:

[0006] S1. Track image semantic segmentation: Use the dilated convolution of DeepLabV3+ to segment the track image and obtain the semantic label of each pixel;

[0007] S2, point cloud and image space calibration: obtain point cloud frame and camera frame data, synchronize the data, and then map the obtained semantic labels to the point cloud data to complete the point cloud and image space calibration;

[0008] S3. Construction of global semantic map of track: The LIO-SAM algorithm estimates the IMU data to determine the movement posture and direction of the device, and then generates IMU odometer data to correct the distortion of the point cloud, thus optimizing the 3D mapping process. After completing the point cloud and image space calibration, the LIO-SAM algorithm is combined with the acquired semantic information to construct a global semantic map of the track.

[0009] S4. Track global semantic map optimization: The track global semantic map is optimized through statistical filtering algorithm and plane fitting algorithm RANSAC;

[0010] S5. Combined positioning of track inspection: On the optimized global semantic map of the track, the sleeper number is used as the global positioning reference to constrain the local motion error between the sleepers, thereby realizing the combined positioning of track inspection with multi-source information fusion.

[0011] Preferably, in step S1, the track image is segmented using the dilated convolution of DeepLabV3+, and the formula is expressed as follows:

[0012]

[0013] In the formula, y[i 1 ] is the i-th output feature map 1 elements, w[j] is the weight in the convolution kernel, and j represents the position in the convolution kernel. 1 +r·j] is the i-th 1 +r·j elements, r is the dilation rate, which indicates the jump length of the convolution kernel on the input signal. f is the size of the convolution kernel, which indicates the number of elements in the convolution kernel.

[0014] Preferably, in step S2, let the laser radar coordinate system be a point p l =[x l ,y l ,z l ] T , after the camera internal parameter K, rotation matrix Translation Matrix After the distortion parameters, the conversion to the pixel coordinate system is obtained Represents the pose transformation matrix between the lidar and the industrial camera. The specific conversion relationship is as follows:

[0015]

[0016] Through n 1 The distance difference between the pixel coordinates p” corresponding to the lidar points in the group and the calculated pixel coordinates The residual constraint equation constructed by the point pairs is used to solve the external parameters by the least squares iterative optimization method e 1 Is the corresponding residual, and the formula is expressed as follows:

[0017]

[0018] Preferably, in step S3, it is assumed that at time point i 2 The extracted edge features and plane features are And Then at time point i 2 The lidar frame is Then, based on the extracted feature points, a constraint relationship is constructed, and the key frame strategy is used for optimized registration and relative transformation calculation. Finally, the point cloud is registered into the global map, and a point cloud map containing the nearest fixed number of lidar scan data is created using a sliding window method; specifically, it includes the following:

[0019] By extracting the nearest n 2 Sub-key frames Then use the transformation associated with them Convert them into the world frame The converted sub-key frames are merged together to form a voxel map Where Is the edge voxel map, Is the plane voxel map;

[0020] Assume that the new frame scanned by the lidar is According to the motion matrix estimated by the IMU Will Convert from coordinate system B to coordinate system W to obtain Then in the edge voxel map And the plane voxel map Find the corresponding relationship between its edges or planes;

[0021] The distance between the feature and its corresponding edge or plane feature can be calculated by the following formula:

[0022]

[0023] Among them, k, u, v, and w are feature indices in the corresponding sets.

[0024] For the edge feature in it, and are two points that form the corresponding edge line in For the planar feature in it, and are three points that form the corresponding planar point in

[0025] Finally, the Gauss-Newton method is used to minimize the following residual model to solve for the optimal transformation, e 2 is the corresponding residual, as shown in the formula:

[0026]

[0027] Finally, we can obtain the relative transformation matrix and between That is, the lidar measurement factor of the following two postures, and the formula is as follows:

[0028]

[0029] Preferably, in step S4, it specifically includes the following:

[0030] S41. Set the initial point cloud data set S = {p 1 , p 2 , p 3 …, p m}, where each point For the point in S to the distance Dis from any point p r =(x r , y r , z r ), the following formula is used for calculation:

[0031]

[0032] Next, set a neighborhood a for each point, and calculate the average distance θ from this point to all points in its neighborhood. The formula is as follows:

[0033]

[0034] After calculating the average distance of all points, calculate the mean μ and standard deviation σ of these distances. The formulas are as follows:

[0035]

[0036] Then, set a threshold interval as [μ - λσ, μ + λσ], and check whether the average distance θ of the neighborhood a of each point falls within this threshold interval. If it does, retain the point; otherwise, regard it as an outlier and remove it. Finally, all the retained points form a new point cloud data set J = {p 1 , p 2 , p 3 …, p j} after statistical filtering;

[0037] S42. Randomly select a sampling subset N 1 from the filtered point cloud data, and assume that these data are all inliers, and then calculate the model parameters. Subsequently, use other points in the data set to verify the estimated model parameters, and judge whether each data point is an inlier through a set maximum threshold. After a certain number of iterations, select the model with the most inliers. The termination conditions of the iteration are as follows:

[0038]

[0039] In the formula, I is the maximum number of iterations; s is the proportion of inliers in the data, which is continuously updated with the iteration; P is the confidence condition.

[0040] Preferably, in step S5, it specifically includes the following:

[0041] S51. Use the KDE method and peak detection to complete the detection and counting of sleeper positions;

[0042] There are n 3 independent and identically distributed sample data which are taken from a continuous function f(y) of an unknown probability distribution, where y ∈ R. Let be the kernel density estimation function of f(y), and the calculation formula of

[0043]

[0044] is as follows: In the formula, h is the bandwidth, which is used to control the smooth shape of the kernel density function; K(·) is the kernel function, which takes the data value and bandwidth of each sample data as parameters; for different sample data, find their corresponding kernel functions respectively, and then perform linear superposition and averaging on the obtained n 3 kernel functions to obtain the final kernel density estimation function;

[0045] S52. After completing the detection and counting of the sleeper positions using the KDE method and peak detection, the position coordinates L of the track inspection data are expressed using four parameters, namely the starting point coordinate value X, the main number of sleepers N that have passed from the starting point to the current position of the track inspection device 2 , the sleeper spacing d', and the distance l' that the track inspection device moves between adjacent sleepers locally; based on these four parameters, the position coordinates L of the track inspection data are calculated by the following formula: L = X + N 2 ×d' + l'.

[0046] The beneficial effects of the present invention are as follows: The present invention visualizes the global sleeper positioning information by constructing a semantic map, uses the sleeper number as the global positioning reference, constrains the local motion errors between sleepers, and realizes the combined positioning of track inspection with multi-source information fusion, thereby obtaining a more robust and globally accurate positioning result. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic flow chart of the method steps of the present invention;

[0048] Figure 2 is a schematic diagram of the combined positioning of track inspection of the present invention;

[0049] Figure 3 is a schematic diagram of the semantic map constructed by the present invention;

[0050] Figure 4 is a schematic diagram of the detection result of the kernel density estimation KDE algorithm of the present invention;

[0051] Figure 5 is a schematic diagram of the error comparison result of the combined positioning and local positioning methods of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] The track inspection and positioning algorithm can be divided into local positioning, global positioning, and combined positioning that combines both. In the research of track inspection and positioning, the method of using a local motion model to locate the track inspection position is relatively common. The sensors commonly used in local positioning algorithms include odometers and inertial navigation systems. These sensors do not rely on external environmental information and have high accuracy in a short period of time. However, the cumulative errors generated by these sensors during the mileage positioning process are likely to cause problems where the measurement data of multiple cycles cannot be aligned. In contrast, the global positioning algorithm only depends on the position information of the sleepers, so it will not generate cumulative errors over time. However, the discontinuity of the sleepers makes it impossible to obtain accurate position information within the sleeper intervals. Combined positioning is a method that combines global and local positioning, fusing and complementing the advantages of the positions obtained by the two methods. The present invention uses the sleeper number as the global positioning reference to constrain the local motion errors between the sleepers, realizing the combined positioning of track inspection with multi-source information fusion, thereby obtaining a more robust and globally accurate positioning result.

[0054] The present invention provides a technical solution: a combined track inspection and positioning method based on a semantic map, as Figure 1 and Figure 2 shown, including the following steps:

[0055] S1. Semantic segmentation of track images: Use the dilated convolution of DeepLabV3+ to segment the track images to obtain the semantic labels of each pixel;

[0056] S2. Calibration of point cloud and image space: Obtain point cloud frame and camera frame data, synchronize the data, and then map the obtained semantic labels to the point cloud data to complete the calibration of the point cloud and image space;

[0057] S3. Construction of the global track semantic map: The LIO-SAM algorithm estimates the IMU data to determine the movement attitude and direction of the device, and then generates IMU odometer data for distortion correction of the point cloud, optimizing the three-dimensional mapping process; After completing the calibration of the point cloud and image space, combine the LIO-SAM algorithm with the obtained semantic information to construct the global track semantic map;

[0058] S4. Optimization of the global track semantic map: Optimize the global track semantic map through the statistical filtering algorithm and the plane fitting algorithm RANSAC;

[0059] S5. Combined track inspection and positioning: On the optimized global track semantic map, use the sleeper number as the global positioning reference to constrain the local motion errors between the sleepers, realizing the combined positioning of track inspection with multi-source information fusion.

[0060] Further, in step S1, the atrous convolution of DeepLabV3+ is used to segment the track image, and the formula is expressed as follows:

[0061]

[0062] In the formula, y[i 1 is the i-th element of the output feature map, x[j] is the weight in the convolution kernel, and j represents the position in the convolution kernel. x[i 1 1 +r·j] is the i 1 +r·j-th element in the input signal, r is the dilation rate, indicating the jump length of the convolution kernel on the input signal. f is the size of the convolution kernel, representing the number of elements in the convolution kernel.

[0063] Further, in step S2, an improvement is made based on the LIO-SAM algorithm. By introducing a semantic projection model, after obtaining the point cloud frame and camera frame data, the data is first synchronized, and then the semantic segmentation information obtained from the image is projected into the point cloud data.

[0064] Let a point p l =[x l ,y l ,z l in the lidar coordinate system. After passing through the camera internal parameter K, rotation matrix T translation matrix and distortion parameters, the pixel coordinate is obtained in the pixel coordinate system.

[0065]

[0066] Through the residual constraint equation constructed by the distance difference between the pixel coordinates p” corresponding to n 1 pairs of lidar points and the calculated pixel coordinates , the external parameter e 1 is solved by using the least squares iterative optimization method. The formula is expressed as follows:

[0067]

[0068] Further, in step S3, the LIO-SAM algorithm estimates the IMU data to determine the movement attitude and direction of the device, and then generates IMU odometry data for point cloud distortion correction, thereby optimizing the 3D mapping process.

[0069] Suppose at time point i​​2 The extracted edge features and plane features are and At time point i 2 the lidar frame is Then, based on the extracted feature points, a constraint relationship is constructed, and the key-frame strategy is used for optimized registration and relative transformation calculation. Finally, the point cloud is registered into the global map, and a point cloud map containing the most recent fixed number of lidar scan data is created using a sliding window method. Specifically, it includes the following:

[0070] By extracting the most recent n 2 sub-key frames Then, using the transformations associated with them transform them into the world frame The transformed sub-key frames are merged together to form a voxel map where is the edge voxel map, and

[0071] Let the new frame of the lidar scan be According to the motion matrix estimated by the IMU transform from coordinate system B to coordinate system W to obtain Then, in the edge voxel map and the plane voxel map find the corresponding relationships of its edges or planes;

[0072] The distance between the feature and its corresponding edge or plane feature can be calculated using the following formula:

[0073]

[0074] where k, u, v, and w are the feature indices in the corresponding sets.

[0075] For the edge feature in in terms of and are the two points that form the corresponding edge line in. For the plane feature in in terms of and are the three points that form the corresponding plane point in.

[0076] Finally, the Gauss-Newton method is used to minimize the following residual model to solve for the optimal transformation, e 2 is the corresponding residual, as shown in the formula:

[0077]

[0078] Finally, we can obtain the relative transformation matrix and between which is the lidar measurement factor for the following two poses.

[0079]

[0080] This method uses a sliding window to generate a point cloud map containing a fixed number of frames at the most recent moment. The LIO-SAM algorithm can provide robust state estimation and map construction by combining IMU pre-integration and lidar odometry, as Figure 3 shown, in the semantic map, the rail, sleeper, trajectory points obtained by the local positioning method, and the main points of the sleeper calculated are marked.

[0081] Furthermore, in step S4, since the surface of the track slab in reality is not flat and there are a large number of noise data in the measurement data. In order to eliminate the noise introduced by factors such as equipment and environment, data filtering processing becomes a necessary step. The statistical filtering algorithm, which is a filtering technique based on statistics, performs excellently in removing outliers. The specific process is as follows:

[0082] S41. Let the initial point cloud data set S = {p 1 , p 2 , p 3 …, p m}, where each point For point in S to the distance Dis from any point p r = (x r , y r , z r ), it is calculated using the following formula:

[0083]

[0084] Next, set a neighborhood a for each point and calculate the average distance θ from this point to all points within its neighborhood. The formula is as follows:

[0085]

[0086] After calculating the average distance of all points, calculate the average value μ and standard deviation σ of these distances. The formula is as follows:

[0087]

[0088] Then, a threshold interval [μ - λσ, μ + λσ] is set, and it is checked whether the average distance θ of the neighborhood a of each point falls within this threshold interval. If so, the point is retained; otherwise, it is regarded as an outlier and removed. Finally, all the retained points form a new point cloud data set J = {p 1 , p 2 , p 3 …, p j} after statistical filtering;

[0089] S42. Randomly select a sampling subset N 1 from the filtered point cloud data, and assume that these data are all inliers, and then calculate the model parameters. Subsequently, use other points in the data set to verify the estimated model parameters, and judge whether each data point is an inlier through a set maximum threshold. After a certain number of iterations, select the model with the most inliers. The termination conditions for the iteration are as follows:

[0090]

[0091] In the formula, I is the maximum number of iterations; s is the proportion of inliers in the data and is updated continuously with the iteration; P is the confidence condition.

[0092] Furthermore, in step S5, since the intervals of the sleepers on the track slab are fixed and the geometric shapes have obvious changes, in the generated semantic map, the point cloud data will show obvious aggregation characteristics at the positions of the sleepers. Based on the change law that the point cloud density increases from small to large and then to small when passing through the sleepers during the track inspection process, the kernel density estimation (KDE) combined with the peak detection technology can be used to accurately determine the number of sleepers passed by the detection system. KDE is a non-parametric probability density estimation method, suitable for situations where the distribution of the observed data is unknown, such as Figure 4As shown, the curve represents the detection result of the Kernel Density Estimation (KDE) algorithm. By observing this curve, it can be found that at the positions where the sleepers appear, the curve shows a specific variation pattern. Specifically, whenever a sleeper appears, the curve first rises locally and then drops rapidly, and these two changes alternate. In order to capture the overall variation trend of the data within the window, the present invention uses the peak as the feature for identifying the sleepers, and the dots in the figure represent the positions of the sleepers determined by peak detection. In addition, the present invention also sets a sliding window to process the data within the window so as to more accurately obtain the position information of the sleepers. It can be clearly seen from the figure that the track inspection device has passed a distance of 8 sleepers. The true positioning distance is 5 meters, while the actual positioning distance is 5.01 meters, and the global positioning error is 0.20%. This indicates that the positioning error is effectively restricted within the range between the sleepers. It makes no assumptions about the data distribution, but models the probability density function based on the observed data itself. Specifically, it includes the following:

[0093] S51. Use the KDE method and peak detection to complete the detection and counting of the sleeper positions;

[0094] There are n 3 independent and identically distributed sample data which are taken from the continuous function f(y) of an unknown probability distribution, where y ∈ R. Let be the kernel density estimation function of f(y), and the calculation formula of

[0095]

[0096] is as follows: In the formula, h is the bandwidth, which is used to control the smooth form of the kernel density function; K(·) is the kernel function, which takes the data value and bandwidth of each sample data as parameters; for different sample data, the corresponding kernel functions are respectively obtained, and then the n 3 obtained kernel functions are linearly superimposed and averaged to obtain the final kernel density estimation function;

[0097] S52. After using the KDE method and peak detection to complete the detection and counting of the sleeper positions, the position coordinates L of the track inspection data are expressed by four parameters, namely the starting point coordinate value X, the main number N 2 of the sleepers passed from the starting point to the current position of the track inspection device, the sleeper spacing d', and the distance l' that the track inspection device moves between locally adjacent sleepers; based on these four parameters, the position coordinates L of the track inspection data are calculated by the following formula: L = X + N 2 × d' + l'.

[0098] As Figure 5 shown, Figure 5The error comparison results between the combined positioning and local positioning methods in the indoor experiment are shown. In the figure, the x-axis represents the number of sleepers passed, and the y-axis represents the error value of the positioning method. By observing the chart, it can be seen that before passing the first sleeper, since the combined positioning method completely relies on the local positioning data, the errors generated by the two methods are the same. However, as the number of sleepers passed by the track inspection equipment increases, the error of the local positioning method gradually increases due to the cumulative error. In contrast, the error of the combined positioning method fluctuates stably within the range of -0.02 to 0.02 meters. As shown in Table 1:

[0099] Table 1 Error statistical results of track inspection positioning

[0100] Evaluation Index Combined Positioning Method Local Positioning Method MAX 0.032 0.072 RMSE 0.011 0.030

[0101] Table 1 lists the error statistical results of track inspection positioning. According to the analysis results, the combined positioning method shows stable maximum error and root mean square error throughout the process, and its accuracy has been significantly improved compared with the local positioning method.

[0102] The present invention uses the sleeper number as the global positioning reference to constrain the local motion error between sleepers, realizing the combined positioning of track inspection with multi-source information fusion, thereby obtaining a more robust and globally accurate positioning result.

[0103] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A combined positioning method for track inspection based on semantic maps, characterized in that: The steps include: S1. Track image semantic segmentation: Use the dilated convolution of DeepLabV3+ to segment the track image and obtain the semantic label of each pixel; S2, point cloud and image space calibration: obtain point cloud frame and camera frame data, synchronize the data, and then map the obtained semantic labels to the point cloud data to complete the point cloud and image space calibration; S3. Construction of global semantic map of track: The LIO-SAM algorithm estimates the IMU data to determine the movement posture and direction of the device, and then generates IMU odometer data to correct the distortion of the point cloud, thus optimizing the 3D mapping process. After completing the point cloud and image space calibration, the LIO-SAM algorithm is combined with the acquired semantic information to construct a global semantic map of the track. S4. Track global semantic map optimization: The track global semantic map is optimized through statistical filtering algorithm and plane fitting algorithm RANSAC; S5. Track inspection combined positioning: On the optimized track global semantic map, the sleeper number is used as the global positioning reference to constrain the local motion error between sleepers, and the track inspection combined positioning with multi-source information fusion is realized; specifically, it includes the following: S51, using KDE method and peak detection to complete sleeper position detection and counting; Suppose there are n3 independent and identically distributed sample data They are taken from the continuous function f(y) of the unknown probability distribution, where y∈R, let is the kernel density estimation function of f(y), The calculation formula is as follows: Where h is the bandwidth, which is used to control the smoothing shape of the kernel density function; K(·) is the kernel function, which takes the data value and bandwidth of each sample data as parameters; for different sample data, the corresponding kernel function is calculated respectively, and then the obtained n3 kernel functions are linearly superimposed and averaged to obtain the final kernel density estimation function; S52. After the sleeper position detection and counting are completed using the KDE method and peak detection, the position coordinate L of the track inspection data is expressed using four parameters, namely the starting point coordinate value X, the number of sleeper main points N2 that have been passed from the starting point to the current position of the track inspection equipment, and the sleeper spacing d. ′ , the distance l that the track inspection equipment moves between local adjacent sleepers ′ Based on these four parameters, the position coordinate L of the track inspection data is calculated by the following formula: L = X + N2 × d ′ +l ′ .

2. The track inspection combined positioning method based on semantic map according to claim 1 is characterized in that: In step S1, the track image is segmented using the dilated convolution of DeepLabV3+, and the formula is as follows: Where y[i1] is the i1th element of the output feature map, w[j] is the weight in the convolution kernel, and j represents the position in the convolution kernel. x[i1+r·j] is the i1+r·jth element in the input signal, r is the hole rate, which represents the jump length of the convolution kernel on the input signal. f is the size of the convolution kernel, which represents the number of elements in the convolution kernel.

3. The track inspection combined positioning method based on semantic map according to claim 1 is characterized in that: Suppose a point p in the laser radar coordinate system l =[x l ,y l ,z l ] T , after the camera internal parameter K, rotation matrix Translation Matrix After the distortion parameters are obtained, it is converted to the pixel coordinate system to obtain Represents the pose transformation matrix of the laser radar and industrial camera. The specific conversion relationship is: The pixel coordinates p″ corresponding to the n1 pairs of laser radar points are compared with the calculated pixel coordinates The residual constraint equation is constructed by the distance difference between the point pairs, and the external parameter is solved by the method of least squares iterative optimization. e1 is the corresponding residual, and the formula is expressed as follows:

4. The track inspection combined positioning method based on semantic map according to claim 1 is characterized in that: In step S3, the edge features and plane features extracted at time point i2 are assumed to be and Then the lidar frame at time point i2 is Then, based on the extracted feature points, constraint relationships are constructed, and keyframe strategies are used to optimize registration and relative transformation calculations. Finally, the point cloud is registered to the global map, and a sliding window method is used to create a point cloud map containing the most recent fixed number of LiDAR scan data. The details are as follows: By extracting the most recent n2 sub-keyframes Then use the transforms associated with them Convert it to world frame The converted sub-keyframes are merged together to form a voxel map in is the edge voxel map, is a planar voxel map; Let the new frame scanned by LiDAR be Motion matrix estimated by IMU Will Convert from coordinate system B to coordinate system W to obtain Then in the edge voxel map and planar voxel maps Find the corresponding relationship between its edges or planes; The distance between a feature and its corresponding edge or plane feature is calculated using the following formula: Among them, k, u, v, w are feature indices in the corresponding set; for The edge features in In terms of and yes The two points that form the corresponding edge line in Plane features in In terms of and yes The three points that constitute the corresponding surface point; Use the Gauss-Newton method to minimize the following residual model to solve the optimal transformation, where e2 is the corresponding residual, as shown in the formula: Finally, we get the status node and The relative transformation matrix between That is, the lidar measurement factor for the following two postures is shown in the following formula:

5. The track inspection combined positioning method based on semantic map according to claim 1 is characterized in that: In step S4, the specific steps include: S41, let the initial point cloud data set S = {p1, p2, p3…, p m }, where each point For the midpoint of S To any point p r =(x r ,y r ,z r ) is calculated using the following formula: Next, set a neighborhood a for each point and calculate the average distance θ from the point to all points in its neighborhood. The formula is as follows: After calculating the average distance of all points, calculate the mean μ and standard deviation σ of these distances, the formula is as follows: Then, a threshold interval is set as [μ-λσ,μ+λσ], and the average distance θ of each point neighborhood a is checked to see if it falls within this threshold interval. If so, the point is retained; otherwise, it is considered an outlier and removed; finally, all the retained points form a new point cloud dataset J = {p1, p2, p3…, p j }; S42, randomly select a sampling subset N1 from the filtered point cloud data, and assume that these data are all inliers, and then calculate the model parameters; then, use other points in the data set to verify the estimated model parameters, and use a set maximum threshold to determine whether each data point is an inlier; after a certain number of iterations, select the model containing the most inliers, and the termination condition of the iteration is as follows: Where I is the maximum number of iterations; s is the proportion of internal points in the data, which is then continuously updated with iterations; P is the confidence condition.

Images