A positioning method and system based on twin neural network and anti-outlier

Abstract

The application discloses a positioning method and system based on a twin neural network and anti-outlier, and belongs to the technical field of robot navigation. The method comprises the following steps: adopting adaptive Kalman filtering to process robot laser radar scanning data, so as to obtain processed robot laser radar scanning data; performing global feature extraction on the processed robot laser radar scanning data based on a twin neural network, so as to obtain first extracted features; constructing a simulated laser radar data set, performing feature extraction on the laser radar data set based on the twin neural network, so as to obtain second extracted features; comparing the first extracted features and the second extracted features to obtain an initial pose estimation of the robot; and completing pose tracking of the robot based on the initial pose estimation of the robot through a point-line ICP algorithm. The application can overcome the defect that the positioning accuracy of the existing positioning method is low.

Description

Technical Field

[0001] This invention relates to the field of robot navigation technology, and in particular to a localization method system based on Siamese neural networks and anti-outlier. Background Technology

[0002] Currently, the Adaptive Monte Carlo Localization (AMCL) algorithm is the most widely used and computationally efficient method for global relocalization. Based on an improved Monte Carlo Localization (MCL) algorithm, AMCL takes the robot's motion model and LiDAR observation model as inputs and uses a particle filter probability model to obtain the robot's global position in the navigation coordinate system. Notably, AMCL assumes that the odometry model accurately matches the environment.

[0003] Another commonly used localization algorithm, Point-to-Line Iterative ClosestPoint (PL-ICP), calculates the nearest point in ICP iterations by minimizing the distance from a point to a line. In PL-ICP, a point in one point cloud does not always match a point in another point cloud; instead, matching is based on the distance from the point to a line segment or line. The robot's pose is estimated using iterative equations by utilizing the pose difference between two consecutive laser scans.

[0004] However, in practical applications, odometry models may not accurately match the environment because odometry data can be affected by factors such as noise and drift, leading to discrepancies with the actual environment and reducing the accuracy and reliability of the adaptive Monte Carlo positioning (AMCL) algorithm. Furthermore, PL-ICP requires precise initial positioning and radar data during movement, and cannot be moved arbitrarily. Summary of the Invention

[0005] The purpose of this invention is to provide a localization method and system based on Siamese neural networks and anti-outlier features. The method uses a Siamese neural network to extract features from acquired robot LiDAR scanning data and a constructed simulated LiDAR dataset, compares these features to obtain an initial pose estimate for the robot, and inputs this estimate into a point-line ICP algorithm to achieve precise robot pose tracking. This invention is achieved through the following technical solutions.

[0006] In a first aspect, the present invention provides a localization method based on Siamese neural networks and anti-outlier values, comprising:

[0007] The robot's lidar scanning data is acquired, and the robot's lidar scanning data is processed using an adaptive Kalman filter to obtain the processed robot lidar scanning data.

[0008] Global feature extraction is performed on the processed robot lidar scanning data based on a Siamese neural network to obtain the first extracted features; a simulated lidar dataset is constructed, and feature extraction is performed on the lidar dataset based on a Siamese neural network to obtain the second extracted features;

[0009] The robot's initial pose is estimated by comparing the first and second extracted features.

[0010] The robot's pose tracking is achieved by using the point-line ICP algorithm based on the robot's initial pose estimation.

[0011] Optionally, adaptive Kalman filtering can be used to process the robot's lidar scanning data, including dynamically adjusting the measurement noise covariance and detecting outliers.

[0012] Optionally, the measurement noise covariance is dynamically adjusted using the following formula:

[0013] ,

[0014] In the formula, This is a new information sequence that reflects the deviation between the actual measurement and the predicted state. To set a minimum threshold of 0.1 to maintain filter stability, The time index represents the discrete time step. for The distance observation value at a given time, i.e., the actual measurement value of the lidar. for Always Predicting the state at any given time. for The measurement noise covariance at any given time represents the uncertainty of the measurement error and is dynamically adjusted to adapt to environmental changes. This formula is dynamically adjusted... Measurement noise covariance at time To suppress noise and improve filtering accuracy.

[0015] Optionally, the Distance observation at time It is obtained by calculation using the following formula:

[0016] ,

[0017] In the formula, for The distance measurement at any given moment, i.e., the robot's actual distance. To measure noise, a Gaussian distribution is used, representing sensor error;

[0018] The Distance measurement at time It is obtained by calculation using the following formula:

[0019] ,

[0020] In the formula, express Distance measurement at time, This is process noise.

[0021] Optionally, the outlier detection includes global outlier detection and local outlier detection; the global outlier detection determines the data range based on a 99.9% confidence interval, and data points outside the data range are marked as outliers; the local outlier detection uses a sliding window to calculate the Z-score; wherein, the formula for calculating the Z-score is as follows:

[0022] ,

[0023] In the formula, This refers to a single distance measurement value in the robot's LiDAR scanning data. To simulate the mean of the lidar dataset, To simulate the standard deviation of the lidar dataset, Data points with scores exceeding 3 are considered outliers.

[0024] Optionally, the simulated LiDAR dataset is constructed using the Bresenham line drawing algorithm, which includes traversing grid occupancy by slope and updating pixels in the simulated LiDAR dataset using integer decision parameters; wherein, the slope is calculated using the following formula:

[0025] ,

[0026] In the formula, The slope , The absolute value of the slope determines the main traversal direction.

[0027] when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal; when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal;

[0028] The update of pixels in the simulated LiDAR dataset is performed using integer decision parameters, and the update expression is as follows:

[0029] ,

[0030] In the formula, The time index represents the discrete time step. for Integer decision parameters at time t. , , and These represent the starting and ending coordinates of a laser beam within the occupied raster map;

[0031] Integer decision parameters at time points It is obtained by calculation using the following formula:

[0032] .

[0033] Optionally, the initial pose estimate is calculated using the following formula:

[0034] ,

[0035] In the formula, To share the CNN encoder, For the final classification function, As the first extracted feature, For the second extracted feature, The initial pose estimation for the robot represents the rough position and orientation obtained based on feature comparison, providing a starting point for the point-line ICP algorithm.

[0036] Optionally, the training data setup for the Siamese neural network includes constructing training sample pairs, which include positive sample pairs and negative sample pairs, wherein the expression for the positive sample pair is as follows:

[0037] ,

[0038] In the formula, For positive sample pairs, i and j are sample indices. and For the second extracted features Different samples in and Representing samples respectively and Position coordinates;

[0039] The expression for the negative sample pair is as follows:

[0040] ,

[0041] In the formula, These are negative sample pairs.

[0042] The simulated LiDAR dataset contains samples from many locations. Balancing the training set (i.e., the simulated LiDAR dataset) ensures that each location contributes an equal number of positive and negative sample pairs. This guarantees that the Siamese neural network can effectively learn feature similarity and improve pose estimation accuracy.

[0043] Optionally, the loss function of the Siamese neural network is a binary cross-entropy loss function, which is calculated using the following formula:

[0044] ,

[0045] In the formula, The binary cross-entropy loss function measures the difference between predicted and actual values. This is the true label vector, containing the true labels (0 or 1) for all samples. This is the predicted probability vector, containing the predicted probability values ​​for all samples (between 0 and 1), where N is the total number of samples in the training batch. For the first The true label of each sample For the first The predicted probability of a sample.

[0046] Secondly, this invention introduces a localization system based on Siamese neural networks and outlier resistance, comprising:

[0047] The data processing module is used to acquire robot lidar scanning data, and to process the robot lidar scanning data using adaptive Kalman filtering to obtain processed robot lidar scanning data.

[0048] The feature extraction module is used to perform global feature extraction on the processed robot lidar scanning data based on a Siamese neural network to obtain a first extracted feature; construct a simulated lidar dataset, and perform feature extraction on the lidar dataset based on a Siamese neural network to obtain a second extracted feature;

[0049] The feature comparison module is used to compare the first extracted features and the second extracted features to obtain the robot's initial pose estimate;

[0050] The pose tracking module is used to complete the precise pose tracking of the robot based on the robot's initial pose estimation using the point-line ICP algorithm.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] This invention combines a Siamese neural network with the point-line ICP algorithm, achieving both rapid global relocalization and high-precision pose tracking, thus improving position recovery capabilities in robot kidnapping situations. A simulated LiDAR dataset is constructed using a Bresenham-accelerated LiDAR simulator, enabling high-precision, real-time omnidirectional synthetic scanning and generating a standard-format LiDAR dataset containing distance and angle information. Initial robot pose estimation is obtained through feature comparison, achieving rapid initial pose recovery and continuous accurate positioning. An adaptive Kalman filter framework with dynamic noise covariance estimation effectively suppresses outliers in highly dynamic environments while maintaining reliable scan matching performance. Attached Figure Description

[0053] Figure 1 The diagram shown is a schematic flowchart of a localization method based on Siamese neural network and anti-outlier value in one embodiment of the present invention;

[0054] Figure 2 The figure shown is a schematic diagram of a lidar scanning simulation based on the Bresenham line drawing algorithm in one embodiment of the present invention;

[0055] Figure 3 The diagram shown illustrates the experimental environment and effects of global relocation in one embodiment of the present invention.

[0056] Figure 4 The figure shown is a schematic diagram comparing the experimental trajectory results of global relocation in one embodiment of the present invention;

[0057] Figure 5 The diagram shown is a comparison of continuous positioning trajectory results in a dynamic laboratory environment according to one embodiment of the present invention.

[0058] Figure 6 The diagram shown is a comparison of continuous positioning trajectory results in a dynamic laboratory environment according to one embodiment of the present invention. Detailed Implementation

[0059] The following description, in conjunction with the accompanying drawings and specific embodiments, provides further details. In this description, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0060] Example 1

[0061] This embodiment provides a localization method based on Siamese neural networks and anti-outlier values, including:

[0062] The robot's lidar scanning data is acquired, and the robot's lidar scanning data is processed using an adaptive Kalman filter to obtain the processed robot lidar scanning data.

[0063] Global feature extraction is performed on the processed robot lidar scanning data based on a Siamese neural network to obtain the first extracted features; a simulated lidar dataset is constructed, and feature extraction is performed on the lidar dataset based on a Siamese neural network to obtain the second extracted features;

[0064] The robot's initial pose is estimated by comparing the first and second extracted features.

[0065] The robot's pose tracking is achieved by using the point-line ICP algorithm based on the robot's initial pose estimation.

[0066] Example 2

[0067] Based on Example 1, this example describes the specific implementation process of a localization method based on Siamese neural networks and anti-outlier, such as... Figure 1 As shown, it specifically includes the following:

[0068] I. Data Processing

[0069] In one specific embodiment of the present invention, adaptive Kalman filtering is used to process robot lidar scanning data, including dynamically adjusting the measurement noise covariance and outlier detection.

[0070] The measurement noise covariance is dynamically adjusted using the following formula:

[0071] ,

[0072] In the formula, This is a new information sequence that reflects the deviation between the actual measurement and the predicted state. To set a minimum threshold of 0.1 to maintain filter stability, The time index represents the discrete time step. for The distance observation value at a given time, i.e., the actual measurement value of the lidar. for Always Predicting the state at any given time. for The measurement noise covariance at any given time represents the uncertainty of the measurement error and is dynamically adjusted to adapt to environmental changes. This formula is dynamically adjusted... Measurement noise covariance at time To suppress noise and improve filtering accuracy.

[0073] Distance observation at time It is obtained by calculation using the following formula:

[0074] ,

[0075] In the formula, for The distance measurement at any given moment, i.e., the robot's actual distance. To measure noise, a Gaussian distribution is used, representing sensor error;

[0076] Distance measurement at time It is obtained by calculation using the following formula:

[0077] ,

[0078] In the formula, express Distance measurement at time, This is process noise.

[0079] Outlier detection includes global outlier detection and local outlier detection. Global outlier detection establishes a reasonable data range based on a 99.9% confidence interval, and data points outside this range are marked as outliers. Local outlier detection uses a sliding window to calculate local statistical data. In this embodiment, the window size is 5, and the judgment is based on the Z-score. The formula for calculating the Z-score is as follows:

[0080] ,

[0081] In the formula, This refers to a single distance measurement value in the robot's LiDAR scanning data. To simulate the mean of the lidar dataset, To simulate the standard deviation of the lidar dataset, Data points with scores exceeding 3 are considered outliers, or abnormal values.

[0082] In one specific embodiment of this invention, two strategies are implemented for detected outliers: infinite values ​​(inf) are replaced with a specified maximum range (18m), while other outliers are replaced with the average of the last 5 valid values. This method successfully identifies and processes a large number of outliers while preserving the original geometric features.

[0083] The simulated LiDAR dataset was constructed using the Bresenham line drawing algorithm. The Bresenham algorithm includes traversing grid occupancy based on slope and updating pixels in the simulated LiDAR dataset using integer decision parameters. The slope is calculated using the following formula:

[0084] ,

[0085] In the formula, The slope , The absolute value of the slope determines the main traversal direction.

[0086] when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal; when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal;

[0087] The update of pixels in the simulated LiDAR dataset is performed using integer decision parameters. The update expression is as follows:

[0088] ,

[0089] In the formula, for Integer decision parameters at time t. , , and These represent the starting and ending coordinates of a laser beam within the grid map.

[0090] Integer decision parameters at time points It is obtained by calculation using the following formula:

[0091] .

[0092] Integer decision parameters The symbol determines the position of the next traversal point, ensuring that the ray path optimally approximates the robot's actual trajectory.

[0093] By iteratively executing at different locations and directions, a standard-format simulated LiDAR dataset containing distance and angle information is generated, accurately reflecting obstacle distribution and suitable for neural network training.

[0094] A schematic diagram of a lidar scanning simulation based on Bresenham's line drawing algorithm is shown below. Figure 2 As shown. Figure 2 This demonstrates the process of simulating a LiDAR dataset on an occupancy grid map, including the implementation of ray path tracing and obstacle detection. The Bresenham line drawing algorithm first generates multiple uniformly distributed locations on the occupancy grid map, such as... Figure 2As shown in the blue circle on the left, a 360° ray is generated based on each position. Selecting one of the blue circles on the left and zooming in will produce the following effect: Figure 2 As shown on the right, the blue line segment represents the simulated lidar ray. Figure 2 In the three blue squares, the leftmost blue square represents the next position the ray can reach by traversing to the rightmost two diagonal squares if the ray is currently traversing to this blue square.

[0095] Detecting grid along ray path ( Figure 2 The gray grid is occupied. When the first occupied grid (obstacle) is encountered, the Euclidean distance is calculated and recorded as the measurement value; if there is no obstacle, the maximum range is returned; if it exceeds the map boundary, the mark is invalid.

[0096] II. Feature Extraction and Training

[0097] In one specific embodiment of the present invention, the initial pose estimation is calculated using the following formula:

[0098] ,

[0099] In the formula, To share the CNN encoder, For the final classification function, As the first extracted feature, For the second extracted feature, The initial pose estimation for the robot represents the rough position and orientation obtained based on feature comparison, providing a starting point for the point-line ICP algorithm.

[0100] In one specific embodiment of the present invention, the training data setup for the Siamese neural network includes constructing training sample pairs, which include positive sample pairs and negative sample pairs. The expression for a positive sample pair is as follows:

[0101] ,

[0102] In the formula, For positive sample pairs, i and j are sample indices. and For the second extracted features Different samples in and Representing samples respectively and Position coordinates;

[0103] The expression for negative sample pairs is as follows:

[0104] ,

[0105] In the formula, These are negative sample pairs.

[0106] The simulated LiDAR dataset contains samples from many locations. Balancing the training set (i.e., the simulated LiDAR dataset) ensures that each location contributes an equal number of positive and negative sample pairs. This guarantees that the Siamese neural network can effectively learn feature similarity and improve pose estimation accuracy.

[0107] In one specific embodiment of the present invention, the loss function of the Siamese neural network is a binary cross-entropy loss function, which is calculated using the following formula:

[0108] ,

[0109] In the formula, The binary cross-entropy loss function measures the difference between predicted and actual values. This is the true label vector, containing the true labels (0 or 1) for all samples. This is the predicted probability vector, containing the predicted probability values ​​for all samples (between 0 and 1), where N is the total number of samples in the training batch. For the first The true label of each sample For the first The predicted probability of a sample.

[0110] III. Experiment

[0111] This invention utilizes a custom-designed multi-sensor fusion robot platform for experimentation. This platform integrates an ACS506 single-drive controller with hub motors to achieve precise motion control. The multi-sensor system includes a 2D LiDAR with a 360° field of view and a 9-axis inertial measurement unit (IMU) for motion estimation. All sensor data processing and platform control are performed by an x86 industrial computer running a Robot Operating System (ROS).

[0112] In one specific embodiment of the present invention, two datasets are constructed for comprehensive evaluation. Dataset 1: an initialization dataset for global relocalization; Dataset 2: a dataset for validating pose tracking in dynamic environments. The following... Figure 3 The robot's original trajectory was generated using the point-line ICP algorithm.

[0113] 3.1 Global Relocation Experiment

[0114] like Figure 3 The image shows a schematic diagram of the global relocation experimental environment and its effects. Figure 3In the diagram, the orange line represents the robot's original trajectory when it was not externally abducted. The blue dot X on the orange line represents the robot's current position. The robot was then abducted from point X to one of 10 different test locations, AJ, selected in the laboratory environment. Experiments proved that the robot successfully found the corresponding location in the database at each of these 10 points, as shown by the red squares in the diagram.

[0115] Experimental results are as follows Figure 4 As shown, the horizontal axis represents time (s) and the vertical axis represents the error distance (m), demonstrating the recovery capabilities of the two methods when the robot is abducted to the same location. Figure 4 In the figure, amcl represents the trajectory generated by the adaptive Monte Carlo localization AMCL algorithm, and our represents the trajectory generated by the method of this invention. As can be seen from the figure, the method of this invention, through the matching of the Siamese neural network, allows the robot to quickly obtain its position on the map at the 3rd second with very small error; while the traditional Monte Carlo localization AMCL method requires continuous movement while performing algorithm matching and localization, taking approximately 15 seconds in total. Therefore, this demonstrates the superior performance of the method of this invention in terms of recovery speed.

[0116] 3.2 Continuous positioning in dynamic environments

[0117] Continuous positioning experiments in two dynamic environments, the office and the corridor, such as Figures 5-6 As shown, the results indicate that, compared with the point-line PL-ICP and adaptive Monte Carlo positioning AMCL methods, the method provided by this invention exhibits superior stability and positioning accuracy in dynamic environments.

[0118] Figure 5 The image shows a comparison of the results of a continuous positioning trajectory experiment in a dynamic office environment. Figure 5 In the image, the trajectory selection positions (1) and (2) generated by the three methods are magnified. In the magnified local effect image, the trajectory generated by the adaptive Monte Carlo positioning AMCL method shows a sharp turn at position (2), indicating a sudden loss of position. At position (1), the point-line PL-ICP also shows a sudden loss of position. Therefore, Figure 5 This demonstrates that the positioning method of the present invention is smoother and does not suffer from the problems of sudden loss of position and position jitter that occur in the other two methods.

[0119] Figure 5 and Figure 6 In this context, pose_amcl, pose_pl-icp, and pose_our represent the trajectories generated by the Adaptive Monte Carlo Localization AMCL method, the point-line ICP method, and the method of this invention, respectively.

[0120] Figure 6The image shows a comparison of the continuous positioning trajectory experimental results in a dynamic environment of a long corridor in the experimental building. To verify the universality of the superior continuous positioning performance of this invention, the same test was conducted in a long corridor environment of the experimental building. Figure 6 The same trajectory records were used to enlarge the selected positions (3) and (4) of the trajectories generated by the three methods. In the enlarged local effect diagram, it can be seen that the positioning method provided by the present invention is still smooth and close to reality, without sudden loss of position or large jitter problems.

[0121] Example 3

[0122] This embodiment introduces a localization system based on Siamese neural networks and outlier resistance, including:

[0123] The data processing module is used to acquire robot lidar scanning data, and to process the robot lidar scanning data using adaptive Kalman filtering to obtain processed robot lidar scanning data.

[0124] The feature extraction module is used to perform global feature extraction on the processed robot lidar scanning data based on a Siamese neural network to obtain a first extracted feature; construct a simulated lidar dataset, and perform feature extraction on the lidar dataset based on a Siamese neural network to obtain a second extracted feature;

[0125] The feature comparison module is used to compare the first extracted features and the second extracted features to obtain the robot's initial pose estimate;

[0126] The pose tracking module is used to complete the precise pose tracking of the robot based on the robot's initial pose estimation using the point-line ICP algorithm.

[0127] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A localization method based on Siamese neural networks and anti-outlier, characterized in that, include: The robot's lidar scanning data is acquired, and the robot's lidar scanning data is processed using an adaptive Kalman filter to obtain the processed robot lidar scanning data. Global feature extraction is performed on the processed robot lidar scanning data based on a Siamese neural network to obtain the first extracted features; a simulated lidar dataset is constructed, and feature extraction is performed on the lidar dataset based on a Siamese neural network to obtain the second extracted features; The robot's initial pose is estimated by comparing the first and second extracted features. The robot's pose tracking is achieved by using the point-line ICP algorithm based on the robot's initial pose estimation. The simulated LiDAR dataset is constructed using the Bresenham line drawing algorithm, which includes traversing grid occupancy based on slope and updating pixels in the simulated LiDAR dataset using integer decision parameters. The slope is calculated using the following formula: ， In the formula, The slope , , when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal; when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal; The update of pixels in the simulated LiDAR dataset is performed using integer decision parameters, and the update expression is as follows: ， In the formula, The time index represents the discrete time step. for Integer decision parameters at time t. , , and These represent the starting and ending coordinates of a laser beam within the occupied raster map; Integer decision parameters at time points It is obtained by calculation using the following formula: 。 2. The localization method based on Siamese neural network and anti-outlier as described in claim 1, characterized in that, The robot's lidar scanning data is processed using adaptive Kalman filtering, including dynamically adjusting the measurement noise covariance and detecting outliers.

3. The localization method based on Siamese neural network and anti-outlier as described in claim 2, characterized in that, The measurement noise covariance is dynamically adjusted using the following formula: ， In the formula, For the new information sequence, To set a minimum threshold of 0.1 to maintain filter stability, The time index represents the discrete time step. for Distance observation at time, for Always Predicting the state at any given time. for Measurement noise covariance at time.

4. The localization method based on Siamese neural network and anti-outlier as described in claim 3, characterized in that, The Distance observation at time It is obtained by calculation using the following formula: ， In the formula, for Distance measurement at time, To measure noise, a Gaussian distribution is used, representing sensor error; The Distance measurement at time It is obtained by calculation using the following formula: ， In the formula, express Distance measurement at time, This is process noise.

5. The localization method based on Siamese neural network and anti-outlier as described in claim 2, characterized in that, The outlier detection includes global outlier detection and local outlier detection. Global outlier detection determines the data range based on a set confidence interval, and data points exceeding this range are marked as outliers. Local outlier detection uses a sliding window to calculate the Z-score. The formula for calculating the Z-score is as follows: ， In the formula, This refers to a single distance measurement value in the robot's LiDAR scanning data. To simulate the mean of the lidar dataset, To simulate the standard deviation of the lidar dataset, Data points whose scores exceed the set value are considered outliers.

6. The localization method based on Siamese neural network and anti-outlier as described in claim 1, characterized in that, The initial pose estimate is obtained by calculating the following formula: ， In the formula, To share the CNN encoder, For the final classification function, As the first extracted feature, For the second extracted feature, This is used for initial pose estimation of the robot.

7. The localization method based on Siamese neural network and anti-outlier as described in claim 6, characterized in that, The training data setup for the Siamese neural network includes constructing training sample pairs, which consist of positive sample pairs and negative sample pairs. The expression for the positive sample pair is as follows: ， In the formula, For positive sample pairs, and The sample number. and For the second extracted features Different samples in and Representing samples respectively and Position coordinates; The expression for the negative sample pair is as follows: ， In the formula, These are negative sample pairs.

8. The localization method based on Siamese neural network and anti-outlier as described in claim 1, characterized in that, The loss function of the Siamese neural network is the binary cross-entropy loss function, which is calculated using the following formula: ， In the formula, Binary cross-entropy loss function, For the actual label vector, Let N be the predicted probability vector, and N be the total number of samples in the training batch. For the first The true label of each sample For the first The predicted probability of a sample.

9. A localization system based on Siamese neural networks and anti-outlier values, characterized in that, include: The data processing module is used to acquire robot lidar scanning data, and to process the robot lidar scanning data using adaptive Kalman filtering to obtain processed robot lidar scanning data. The feature extraction module is used to perform global feature extraction on the processed robot lidar scanning data based on a Siamese neural network to obtain a first extracted feature; construct a simulated lidar dataset, and perform feature extraction on the lidar dataset based on a Siamese neural network to obtain a second extracted feature; The feature comparison module is used to compare the first extracted features and the second extracted features to obtain the robot's initial pose estimate; The pose tracking module is used to complete the precise pose tracking of the robot based on the robot's initial pose estimation using the point-line ICP algorithm. The simulated LiDAR dataset is constructed using the Bresenham line drawing algorithm, which includes traversing grid occupancy based on slope and updating pixels in the simulated LiDAR dataset using integer decision parameters. The slope is calculated using the following formula: ， In the formula, The slope , , when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal; when At that time, the Bresenham line drawing algorithm draws lines along the grid map. Axis traversal; The update of pixels in the simulated LiDAR dataset is performed using integer decision parameters, and the update expression is as follows: ， In the formula, The time index represents the discrete time step. for Integer decision parameters at time t. , , and These represent the starting and ending coordinates of a laser beam within the occupied raster map; Integer decision parameters at time points It is obtained by calculation using the following formula: 。

Images