A laser inertial slam method and system based on transformer

Abstract

This invention relates to the field of autonomous vehicle positioning technology in underground mine tunnels, specifically a Transformer-based laser inertial SLAM method and system. The system includes an array IMU with a spatial pyramid layout and a lidar. The method includes: acquiring array IMU data and point clouds; recursively calculating three-dimensional angular acceleration based on lever arm constraints; using channel structure to perceive noise-adaptive parameters in the Transformer output process; mapping the point cloud to a distance image based on spherical projection, and generating three levels of masks—static background, potential dynamics, and uncertainty—through a lightweight semantic segmentation network; performing two-level filtering with semantic prior guidance and local geometric consistency constraints to output a cleaned point cloud; injecting adaptive parameters into an error state Kalman filter, fusing the cleaned point cloud to perform registration and measurement updates; outputting high-frequency poses and constructing a global map that removes dynamic features. This invention achieves robust positioning and high-fidelity mapping in highly dynamic and degraded underground scenarios.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle positioning technology in underground mine tunnels, specifically a laser inertial SLAM method and system based on Transformer. Background Technology

[0002] Coal mine underground roadways are characterized by their narrow, elongated spaces, complete lack of sunlight, and absolute shielding from GPS signals. Mobile robots operating autonomously in such environments must rely on onboard sensors to achieve real-time positioning and mapping without external absolute positioning signals. Due to the highly repetitive geometry of the roadway sidewalls, ground, and roof, and the lack of significant geometric changes along the roadway's direction, lidar point clouds exhibit inherent degradation along the roadway's orientation. Pure laser odometry is highly susceptible to rapid divergence in long, straight roadways due to feature matching ambiguity.

[0003] To overcome the aforementioned degradation issues, the current mainstream technical approach is to tightly couple and fuse an Inertial Measurement Unit (IMU) with a lidar system. However, in actual underground coal mine operations, mobile vehicles face continuous high-frequency vibrations and impacts on rugged terrain. The angular velocity and acceleration measurements from a single IMU are mixed with significant vibration noise, and the IMU's own zero bias drifts slowly with temperature and time, making it difficult to maintain stable inertial constraints during long-duration operations. More critically, a single IMU cannot directly measure rigid body angular acceleration, and the rate of change of angular acceleration during acceleration, deceleration, and turning directly affects the accuracy of attitude prediction. Conventional laser-inertial SLAM methods assume a fixed process noise covariance, which cannot adaptively adjust to changes in the vehicle's motion state and vibration intensity. In the high-dynamic conditions underground, this easily leads to distortion of the filtered prediction confidence level, resulting in a continuous accumulation of attitude estimation errors.

[0004] In dynamic mapping, moving objects such as mine cars and inspection personnel in coal mine roadways generate a large number of dynamic features in laser point clouds. If these features are directly used for mapping, they will cause ghosting and trailing in the map. Existing semantic segmentation-assisted dynamic removal schemes (such as CN202210586859.X) use binary classification semantic masks to distinguish between dynamic and static features. However, underground point clouds suffer from insufficient semantic classification confidence due to long-distance sparseness and severe occlusion. The binary classification mechanism cannot handle such "uncertain" areas and lacks geometric correction methods for semantic misjudgments. It is difficult to balance recall and accuracy in dynamic removal.

[0005] Therefore, in view of the above situation, there is an urgent need to develop a laser inertial SLAM method and system based on Transformer to overcome the shortcomings in current practical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a Transformer-based laser inertial SLAM method and system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A Transformer-based laser inertial SLAM method includes the following steps:

[0009] S1. Acquire angular velocity and linear acceleration data collected by an array of inertial measurement units arranged in a spatial pyramid, as well as three-dimensional point cloud data collected by lidar.

[0010] S2. Based on the spatial configuration constraints of the array inertial measurement unit and the kinematic principle of the lever arm, the acceleration consistency constraint residual is constructed by utilizing the measurement redundancy between the main inertial measurement unit and the auxiliary inertial measurement unit. The three-dimensional angular acceleration vector of the rigid body of the carrier is calculated in real time by recursively estimating the optimal value of the residual.

[0011] S3. The angular velocity and linear acceleration data of the array inertial measurement unit are coupled with the three-dimensional angular acceleration vector in a time sequence. The multi-head self-attention encoder network with channel structure awareness is used to perform time sequence correlation modeling on the joint feature sequence within the preset sliding window, and the adaptive adjustment parameter set of the process noise covariance matrix is ​​output.

[0012] S4. Based on spherical projection geometric transformation, the three-dimensional point cloud is mapped to a two-dimensional distance image space, and a lightweight semantic segmentation network is used to perform pixel-by-pixel classification to generate a three-level semantic segmentation mask containing static background class, potential dynamic class and uncertain class.

[0013] S5. Construct a two-level filtering mechanism with semantic prior guidance and local geometric consistency constraints. Use the three-level semantic segmentation mask to perform the first-level semantic gating filtering. Perform the second-level geometric verification on the points to be verified after the first-level filtering based on the local flatness response value and the adaptive dynamic discrimination threshold. Output a cleaned static point cloud set.

[0014] S6. Map the adaptive adjustment parameter set to a process noise covariance matrix and inject it into the prior prediction stage of the error state Kalman filter, and use the purified static point cloud to perform point cloud registration and measurement update.

[0015] S7. Based on the fusion estimation results of the error state Kalman filter, output the six-degree-of-freedom high-frequency pose of the carrier and construct a global point cloud map after removing dynamic features.

[0016] As a further aspect of the present invention: in step S2, the acceleration of the auxiliary inertial measurement unit and the acceleration of the main inertial measurement unit satisfy the following lever arm relationship:

[0017] ;

[0018] in, Let be the acceleration of the i-th auxiliary IMU (i=1,2,3). Main IMU accelerometer Main IMU angular velocity, r i For spatial location, This represents the angular acceleration obtained by constructing an acceleration constraint residual function and recursively estimating the optimal value.

[0019] As a further aspect of the present invention: step S3 specifically includes:

[0020] S31. Stack the six-dimensional measurement vectors of the four inertial measurement units at sampling time k vertically according to a fixed spatial topology from the main inertial measurement unit to the auxiliary inertial measurement unit to construct a joint measurement vector of the twenty-four-dimensional array inertial measurement units. The three-dimensional angular acceleration vector and the twenty-four-dimensional joint measurement vector are concatenated longitudinally to construct a twenty-seven-dimensional joint feature vector at time k. ;

[0021] S32. Perform time-series expansion on the 27-dimensional joint feature vector with a preset sliding window length N to construct an N×27-dimensional time-series input matrix;

[0022] S33. Apply timing position encoding to the timing input matrix to obtain an encoded input sequence matrix carrying timing position information;

[0023] S34. The encoded input sequence matrix is ​​fed into a multi-layer channel structure-aware multi-head self-attention encoder network. The self-attention calculation in the encoder network introduces a channel structure-aware bias matrix. The bias matrix is ​​divided into several structural sub-blocks according to the block composition rules of the 27-dimensional joint feature vector. The initial value of each sub-block is physically guided and initialized based on the Euclidean distance between the spatial position vectors of the four inertial measurement units, and end-to-end learning and updating are performed during the training phase. After being encoded layer by layer by the encoder network, a 12-dimensional process noise adaptive parameter vector is output, and the process noise covariance diagonal matrix is ​​constructed based on this. .

[0024] As a further aspect of the present invention: step S4 specifically includes:

[0025] S41. The three-dimensional point cloud collected by the lidar in the scanning frame is mapped point by point to the two-dimensional distance image space through spherical projection transformation to obtain the pixel coordinates (u, v) of each laser point in the distance image;

[0026] S42. Construct a five-channel input feature tensor I from the laser point attributes projected to each valid pixel coordinate. k The five-dimensional feature vector at pixel (u, v) is:

[0027] ;

[0028] Where, x i、 y i z i Indicates the three-dimensional spatial coordinates of the i-th laser point in its lidar coordinate system; intensity i r represents the reflection intensity of the i-th laser point; i This represents the distance from the i-th laser point to the origin of the lidar; r max and I max These are the nominal values ​​of the maximum detection range and maximum reflection intensity of the lidar, respectively.

[0029] S43. The five-channel input feature tensor is fed into a lightweight semantic segmentation encoder with the inverse residual bottleneck module as the basic unit for multi-scale hierarchical feature extraction; a multi-branch heterogeneous context aggregation module is embedded at the end of the encoder to generate a multi-scale context-enhanced semantic feature map.

[0030] S44. The semantic feature map is fed into a lightweight decoder, and after step-by-step upsampling and concatenation of the shallow feature map with the dimension of the feature map, channel compression and fusion are performed to output a pixel-by-pixel semantic feature map restored to the original spatial resolution.

[0031] S45. The pixel-by-pixel semantic feature map is mapped to three semantic categories using a classification head. After normalization, the posterior probability estimate of each pixel belonging to each category is obtained, and the three-level semantic segmentation mask is generated. In this context, H and W represent the height and width of the mask, respectively; label 1 represents the static background class, label 2 represents the potential dynamic class, and label 3 represents the uncertain class.

[0032] As a further aspect of the present invention: the two-stage filtering mechanism specifically includes:

[0033] S51. Based on the spherical projection index mapping relationship, the three-level semantic segmentation mask is traced back to the three-dimensional point cloud space, and each laser point in the original point cloud dataset is assigned a corresponding category label; the static background point set P with label 1 is retained. static The set of potential dynamic points labeled 2 and the set of uncertain points labeled 3 are combined to form the set of points to be verified, P. verify ;

[0034] S52, For each point p to be verified jThe predicted pose of the error-state Kalman filter is transformed to the global world coordinate system, and the nearest neighbor set Q is obtained in the local sub-map. near ;

[0035] When the number of nearest neighbors meets a preset threshold, the local geometric structure features of the nearest neighbor set are extracted, and eigenvalue decomposition is performed on the covariance matrix to obtain descending eigenvalues. ,by The corresponding eigenvectors are used as local surface normal vector estimates;

[0036] Define the local flatness response value:

[0037] ;

[0038] when When the value is less than a preset threshold, it is determined that no effective plane has been formed in the neighborhood; when When the distance to the point to be verified is greater than or equal to a preset threshold, the neighborhood is determined to form a valid plane, and the distance from the point to be verified to this local fitting plane is used as the geometric consistency residual. ;

[0039] Based on the geometric residual distribution of the static background point set, the noise scale σ is estimated using the robust statistic of median absolute deviation. mad Construct an adaptive dynamic discrimination threshold , l is the sensitivity coefficient. j The semantic label value of the j-th point; the point to be verified, p j Perform geometric consistency determination and label correction: ;

[0040] S53. Extract all laser points that are finally marked as static labels after two-stage filtration, and output the purified static point cloud set P. clean .

[0041] As a further aspect of the present invention, step S6 specifically includes:

[0042] S61. Construct a 30-dimensional extended error state vector. The independent accelerometer bias of each array inertial measurement unit and the three-dimensional angular acceleration estimation error calculated recursively from the lever arm constraint residual are incorporated into the state space:

[0043] ;

[0044] Where δp is the position error, δv is the velocity error, δθ is the attitude angle error, and δb is the position error. a0 to δb a3 These represent the accelerometer zero bias errors of the four inertial measurement units, δb. gFor gyroscope zero bias error, This represents the angular acceleration estimation error. This represents the error in the gravity vector.

[0045] S62, The twelve-dimensional process noise adaptive parameter vector A 30-dimensional process noise covariance matrix is ​​injected using a block-based diagonal structure mapping. In the process, the position, velocity, attitude, and noise sub-blocks corresponding to the zero bias of each accelerometer are dynamically updated;

[0046] S63, using the purified static point cloud set P clean Observation update for error state Kalman filtering, observation noise covariance R k The noise scale σ obtained based on the robust estimation of the median absolute deviation mad The online configuration, together with the adaptive injection of process noise covariance in step S62, constitutes a two-way collaborative adaptive mechanism for process noise and observation noise.

[0047] As a further aspect of the present invention, step S7 specifically includes:

[0048] S71. The nominal state propagated by inertial measurement is corrected using the error state posterior update amount, and the six-degree-of-freedom pose estimate of the carrier at the current moment is output.

[0049] S72. Using the corrected six-degree-of-freedom pose estimation, transform each laser point in the purified static point cloud set to the global world coordinate system and insert it into the global map; at the same time, associate and store the semantic segmentation category label corresponding to the scan frame with the global map point.

[0050] As a further aspect of the present invention: the point cloud registration in step S6 and the geometric verification in step S5 share a local geometric structure: the effective planar neighborhood and local surface normal vector determined in step S52 are reused as geometric primitives for point cloud registration; the insertion frequency of the global map in step S7, the point cloud registration update frequency in step S6, and the encoder network inference frequency in step S3 are kept synchronized.

[0051] This invention also proposes a Transformer-based laser inertial SLAM system, comprising:

[0052] The data acquisition module consists of an array of inertial measurement units arranged in a spatial pyramid and a lidar, and is used to perform data acquisition as described in step S1 of claim 1.

[0053] The angular acceleration calculation and adaptive parameter regression module is used to perform the lever arm residual recursive calculation of three-dimensional angular acceleration in step S2 as described in claim 1, and the channel structure perception attention encoder network in step S3 as described in claim 1 outputs a twelve-dimensional process noise adaptive parameter vector.

[0054] The semantic dynamic filtering module is used to perform the spherical projection and lightweight semantic segmentation network to generate a three-level semantic segmentation mask in step S4 as described in claim 1, and to perform the local flatness response value and adaptive dynamic discrimination threshold filtering in step S5 as described in claim 1, and output a purified static point cloud set.

[0055] The error state Kalman filter tightly coupled mapping module uses a 30-dimensional extended error state vector as the filtering space. It maps the 12-dimensional process noise adaptive parameter vector into the process noise covariance matrix in a block diagonal structure, merges the purified static point cloud set to perform recursive updates, and outputs high-frequency poses and incrementally constructs a global map using a dual-end collaborative adaptive mechanism.

[0056] As a further aspect of the present invention: the array inertial measurement unit includes one main inertial measurement unit and three auxiliary inertial measurement units; the main inertial measurement unit is installed directly above the lidar, and the three auxiliary inertial measurement units are respectively installed on the front sides and rear end of the fixed device, and the four inertial measurement units together form a pyramid-shaped array in three-dimensional space; the lidar is installed above the center point of the fixed device and the center of the carrier; the lidar and the four inertial measurement units are all rigidly connected to the fixed device, and the data output ports are all connected to the communication interface of the industrial control computer.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] This invention establishes physically observable conditions for rigid body angular acceleration through a spatial pyramid layout of arrayed inertial measurement units (IMUs). The lever arm constraints between four non-collinear IMUs constitute a redundant set of measurement equations, which mathematically possess a unique solution, thus overcoming the inherent limitation that a single IMU cannot directly sense angular acceleration. The calculated three-dimensional angular acceleration vector is incorporated into the filtered state space in real time as an independent physical observation, enabling the attitude prediction model to explicitly respond to the rate of change of the carrier's angular motion, fundamentally suppressing the tendency of inertial recursion divergence under high-maneuvering conditions.

[0059] Based on the physical redundancy provided by the array IMUs, this invention constructs a 27-dimensional joint feature vector to uniformly encode the raw measurements and calculated angular accelerations from the four IMUs. The channel structure-aware attention network divides the bias matrix into corresponding sub-blocks according to the IMU spatial topology. The initial values ​​of each sub-block are determined by the true Euclidean distance between the IMUs through radial basis function decay, thus constraining and guiding the learning process of the attention weights by the physical configuration. After performing long-range dependency modeling on the temporal feature sequence within the sliding window, the network outputs in real-time 12-dimensional parameters to dynamically adjust the diagonal elements of the noise covariance matrix of the filtering process. This allows the filter's confidence assessment of the predicted noise to adaptively update with changes in the vibration intensity and motion state of the carrier, overcoming the defect of state estimation distortion under high vibration conditions with fixed noise parameters in wells.

[0060] In the point cloud processing branch, this invention generates three categories: static background, potential dynamic, and uncertain. This three-level classification mechanism explicitly separates regions with insufficient semantic confidence, avoiding misclassifications introduced by forcibly classifying low-confidence pixels into static or dynamic categories in binary classification mechanisms. For points that cannot be reliably determined by the semantic branch, this invention further utilizes the local point cloud geometry for verification. It calculates the local flatness response value through eigenvalue decomposition and constructs an adaptive discrimination threshold by combining it with the robust statistics of median absolute deviation, enabling points that do not belong to dynamic objects to be corrected and recovered.

[0061] The processing results of the above branches are tightly coupled and fused within the error state Kalman filter framework: angular acceleration physical observations ensure the structural integrity of the prediction model, adaptive noise covariance provides dynamic adjustment capability in the prediction stage, and pure static point clouds verified by both semantics and geometry provide reliable external measurements for observation updates. The synergistic effect of these three elements enables the system to maintain both long-term stability of pose estimation and geometric purity of map construction in long-distance, high-vibration, and high-dynamic environments downhole. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0063] Figure 2 This is a schematic diagram of the module structure of the system of the present invention;

[0064] Figure 3 This is a schematic diagram showing the layout and installation of the sensor of the present invention on a mobile device carrier;

[0065] Figure 4 This is a two-dimensional trajectory comparison diagram of the method of the present invention and various comparison algorithms;

[0066] Figure 5 The graph shows the cumulative distribution function (CDF) curves of the absolute trajectory error between the method of this invention and various comparison algorithms.

[0067] Figure 6 The graph shows the changes in the errors of each axis of the method of this invention and the comparison algorithms over time.

[0068] In the diagram: 1. LiDAR; 2. Main IMU0; 3. IMU1; 4. IMU2; 5. IMU3; 6. Mobile device carrier; 7. Wheels. Detailed Implementation

[0069] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0070] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0071] I. System Overall Architecture and Module Division

[0072] The Transformer-based laser inertial SLAM system proposed in this invention, such as... Figure 2 As shown, it consists of the following four functional modules:

[0073] The data acquisition module, consisting of an array of inertial measurement units arranged in a spatial pyramid and a lidar, is used to perform the data acquisition function in step S1. This module is responsible for synchronously acquiring angular velocity and linear acceleration data from the four IMUs, as well as 3D point cloud data from the lidar, and transmitting the acquired multi-source sensor data to the industrial control computer for processing.

[0074] The angular acceleration calculation and adaptive parameter regression module is used to perform the recursive calculation of the three-dimensional angular acceleration from the lever arm residual in step S2, and outputs a twelve-dimensional process noise adaptive parameter vector from the channel structure-aware attention encoder network in step S3. This module is based on the spatial configuration constraints of the array IMU and the kinematic principles of the lever arm. It utilizes the measurement redundancy among the four IMUs to construct acceleration consistency constraint residuals and calculates the rigid body's three-dimensional angular acceleration in real time. Simultaneously, it performs temporal joint encoding of the raw IMU measurements and the calculated angular acceleration, and uses a channel structure-aware multi-head self-attention network to perform temporal correlation modeling of the joint feature sequence within the sliding window, outputting adaptive adjustment parameters for the process noise covariance matrix.

[0075] The semantic dynamic filtering module performs step S4, which involves generating a three-level semantic segmentation mask using spherical projection and a lightweight semantic segmentation network, and step S5, which involves filtering based on local flatness response values ​​and an adaptive dynamic discrimination threshold, outputting a cleaned static point cloud set. This module first maps the 3D point cloud acquired by the LiDAR to a 2D distance image space through spherical projection transformation. It then uses a lightweight semantic segmentation network for pixel-by-pixel classification, generating a three-level semantic segmentation mask containing static background, potential dynamic, and uncertain classes. Subsequently, it constructs a two-level filtering mechanism guided by semantic priors and constrained by local geometric consistency. The first level of gating filtering is applied to the semantic segmentation results, and the second level of geometric verification, based on local flatness response values ​​and an adaptive dynamic discrimination threshold, is applied to the points to be verified after the first level of filtering. Finally, it outputs a cleaned static point cloud set free of dynamic interference.

[0076] The error state Kalman filter tightly coupled mapping module uses a 30-dimensional extended error state vector as the filtering space. It maps the 12-dimensional process noise adaptive parameter vector into the process noise covariance matrix in a block-diagonal structure, and performs recursive updates by fusing the cleaned static point cloud set. This outputs high-frequency pose and incrementally constructs a global map. The module incorporates the independent accelerometer biases of each array IMU and the 3D angular acceleration estimation errors recursively calculated from the lever arm constraint residuals into the 30-dimensional state space. It dynamically updates the process noise covariance matrix using adaptive adjustment parameters and performs prior prediction recursion. Simultaneously, it uses the cleaned static point cloud to perform point cloud registration and observation updates. The observation noise covariance is set online based on the noise scale obtained from robust estimation of the median absolute deviation. Together with the process noise adaptive injection, this constitutes a dual-end collaborative adaptive mechanism, ultimately outputting the carrier's six-degree-of-freedom high-frequency pose and constructing a global point cloud map after removing dynamic features.

[0077] II. System Hardware Configuration

[0078] like Figure 3 As shown in the figure, the hardware configuration of the laser inertial SLAM system based on Transformer provided in this embodiment of the invention is as follows:

[0079] In the structural carrier and computing unit section, the chassis of the mobile device carrier 6 is equipped with wheels 7, an industrial control computer is fixedly installed inside the body of the mobile device carrier 6, and a fixing device is provided on the top of the mobile device carrier 6.

[0080] In the sensor installation layout, the lidar 1 is installed at the center point of the fixed device and above the center of the mobile device carrier 6, and the main IMU0 (2) is installed directly above the lidar 1; auxiliary IMU1 (3) and auxiliary IMU2 (4) are distributed and installed at the front end of the fixed device facing the direction of movement of the mobile device carrier 6, and auxiliary IMU3 (5) is installed at the rear end of the fixed device. The main IMU0 (2) and the three auxiliary IMUs together form a pyramid-shaped array in three-dimensional space, and the spatial position vectors of the four IMUs satisfy the non-collinearity condition to ensure that the lever arm constraint equation has a unique solution.

[0081] In the system equipment connection section, the lidar 1 and the four IMUs are all rigidly connected to the fixed device, and the data output ports are all connected to the communication interface of the industrial control computer to enable the industrial control computer to acquire array IMU data and 3D point cloud data in real time.

[0082] III. Method Implementation Steps:

[0083] like Figure 1 As shown, the method flow of the present invention is as follows:

[0084] Step S1: Data Acquisition

[0085] Angular velocity and linear acceleration measurement data were acquired from an array of IMUs arranged in a spatial pyramid layout, along with 3D point cloud data acquired by a lidar. In this embodiment, the IMU output frequency is 200Hz, and the lidar output frequency is 10Hz.

[0086] Step S2: Angular acceleration calculation

[0087] Based on the spatial configuration constraints of the array IMU and the kinematic principle of the lever arm, the acceleration consistency constraint residual is constructed by utilizing the measurement redundancy between the main inertial measurement unit and the auxiliary inertial measurement unit. The three-dimensional angular acceleration vector of the rigid body is calculated in real time by recursively estimating the optimal value of the residual.

[0088] The accelerations of the auxiliary inertial measurement unit and the main inertial measurement unit satisfy the following lever arm relationship:

[0089] ;

[0090] in, Let be the acceleration of the i-th auxiliary IMU (i=1,2,3). Main IMU accelerometer Main IMU angular velocity, r i For spatial location, This represents the angular acceleration obtained by constructing an acceleration constraint residual function and recursively estimating the optimal value.

[0091] Step S3: Adaptive Parametric Regression

[0092] A joint representation of multi-source inertial information is constructed. The angular velocity and linear acceleration data of the array IMU are coupled with the three-dimensional angular acceleration vector for temporal joint encoding. A time-dependent attention encoding mechanism is used to perform long-term time-series correlation modeling on the joint representation sequence within a preset sliding window. The adaptive adjustment parameter set of the process noise covariance matrix is ​​transmitted back in real time.

[0093] Specifically:

[0094] Step S31: Define the six-dimensional measurement vector of the i-th inertial measurement unit at the k-th sampling time as:

[0095] ;

[0096] Wherein, the superscripts x, y, and z represent the projection components on the corresponding axes of the sensor body coordinate system, i=0 corresponds to the main IMU in step S2, and i=1,2,3 correspond to the three auxiliary IMUs in step S2.

[0097] The measurement vectors of the four inertial measurement units are stacked vertically in a fixed spatial topology order from the primary IMU to the auxiliary IMU to construct a 24-dimensional array IMU joint measurement vector. :

[0098] ;

[0099] The three-dimensional angular acceleration vector at time k obtained by recursive optimal estimation of the acceleration consistency constraint residual in step S2 is denoted as... The joint measurement vector of the 24-dimensional array IMU With the three-dimensional angular acceleration vector By concatenating along the vertical axis, a 27-dimensional joint feature vector is constructed at time k:

[0100] .

[0101] Step S32: Construct a time-dependent sensing sliding window input matrix. With a preset sliding window length as a constraint, perform a time-series expansion on the 27-dimensional joint feature vector, and trace back from the current time to the previous time to construct an N×27-dimensional time-series input matrix.

[0102] Step S33: Apply temporal position encoding to generate a sine-cosine alternating position encoding matrix with the same dimension as the temporal input matrix, and add it element by element to the temporal input matrix to obtain an encoded input sequence matrix carrying temporal position information.

[0103] Step S34: Perform channel structure-aware attention encoding and parameter regression, and send the encoded input sequence matrix into a multi-layer channel structure-aware multi-head self-attention encoder network; for the self-attention calculation process in the network, construct a channel structure-aware bias matrix according to the block composition rules of the 27-dimensional joint feature vector, and divide it into multiple structural sub-blocks, which are used to encode the self-interaction bias within each inertial measurement unit, the cross-attention bias between different inertial measurement units, and the cross-modal interaction bias between the inertial measurement unit and the angular acceleration channel;

[0104] Each sub-block of the channel structure sensing bias matrix The initial value is based on the spatial non-collinear position vectors of the four inertial measurement units in step S2. and Euclidean distance between Physically guided initialization is performed using radial basis function decay mapping.

[0105] ;

[0106] in, For the spatial distance attenuation scale parameter, 1 {6×6} It is a matrix of all ones; each sub-block can be learned and updated end-to-end through gradient backpropagation during the training phase;

[0107] After layer-by-layer feature encoding by the network, global average pooling is performed on the final layer output, and dimension mapping is performed through a fully connected regression network to output a twelve-dimensional adaptive process noise parameter vector. This vector is then used to construct the process noise covariance diagonal matrix required for the error state Kalman filter framework. .

[0108] Step S4: Semantic segmentation mask generation

[0109] Based on spherical projection geometric transformation, 3D point clouds are mapped to 2D distance image space. A multi-scale hole context aggregation mechanism is used to perform static reliability assessment on the distance image, generating a three-level scene semantic decoupling probability map containing static high-confidence regions, dynamic suspicious regions, and uncertain regions.

[0110] Specifically:

[0111] Step S41: Perform spherical projection transformation of the three-dimensional point cloud. The three-dimensional point cloud dataset collected by the lidar in the scanning frame is first transformed from the Cartesian coordinate system to the spherical coordinate system to obtain the radial distance, horizontal azimuth angle and vertical pitch angle. Then, the corresponding two-dimensional pixel coordinates (u, v) are obtained through the angle-to-pixel mapping transformation.

[0112] Step S42: Construct a five-channel input feature tensor I from the laser point attributes projected to each effective pixel coordinate. k The five-dimensional feature vector at pixel (u, v) is:

[0113] ;

[0114] Where, x i、 y i z i Indicates the three-dimensional spatial coordinates of the i-th laser point in its lidar coordinate system; intensity i r represents the reflection intensity of the i-th laser point; i This represents the distance from the i-th laser point to the origin of the lidar; r max and I max These are the maximum detection range and the nominal value of the maximum reflection intensity of the lidar, respectively. Each component corresponds to the normalized Cartesian coordinate three components, the normalized reflection intensity, and the normalized radial distance, respectively.

[0115] Step S43: Multi-scale feature extraction is performed using a lightweight semantic segmentation encoder; the five-channel input feature tensor is fed into a lightweight semantic segmentation encoder with an inverse residual bottleneck module as the basic unit for multi-scale hierarchical feature extraction; after performing depthwise separable convolution, the encoder uses a compressed excitation channel attention mechanism to adaptively recalibrate each feature channel.

[0116] Furthermore, a dilated spatial pyramid pooling module is embedded at the end of the encoder to perform heterogeneous context capture in parallel with five branches. Four of these branches employ 3×3 dilated convolutions with dilation rates d1=1, d2=6, d3=12, and d4=18, respectively, while the fifth branch performs global average pooling. The outputs of the five branches are concatenated along the channels and then fused through pointwise convolutions to generate a multi-scale context-enhanced semantic feature map F. aspp .

[0117] Step S44: Spatial resolution is restored using a lightweight decoder, and the semantic feature map F is... aspp The data is fed into a lightweight decoder and undergoes bilinear interpolation upsampling at multiple decoding stages. After upsampling, the shallow feature maps of the corresponding layers in the encoder are concatenated with the upsampled feature maps along the channel dimension via skip connections. Pointwise convolution is then used to compress and fuse the concatenated features. The output is a pixel-wise semantic feature map F restored to its original spatial resolution. dec .

[0118] Step S45: Generate a three-level scene semantic decoupling probability map, and process the pixel-by-pixel semantic feature map F. decBy mapping the convolution class header to three preset semantic categories, and after normalization, the posterior probability estimates p1, p2, and p3 of each pixel belonging to the three categories are obtained. The sum of the three values ​​is 1.

[0119] A semantic segmentation mask is generated by taking the class index with the highest probability from the posterior probability vector of each pixel location:

[0120] ;

[0121] in, In the diagram, H and W represent the height and width of the mask, respectively; label 1 is the static background class, label 2 is the potential dynamic class, and label 3 is the uncertain class.

[0122] Step S5: Two-stage filtering mechanism

[0123] A two-level filtering mechanism is constructed, which combines semantic prior guidance and motion geometry consistency constraints. The first-level semantic gating filtering is performed using a three-level scene semantic decoupling probability graph, and the second-level geometric verification based on rigid body motion consistency metric is performed on suspicious regions, so as to achieve reliable removal of dynamic features and high-fidelity preservation of static background point clouds.

[0124] Specifically:

[0125] Step S51: Perform the first-level coarse-grained dynamic filtering guided by semantic priors. Based on the spherical projection index mapping relationship, backtrack the semantic segmentation mask generated in step S45 to the 3D point cloud space, and assign corresponding category labels to each laser point in the original point cloud dataset; retain the static background point set P with label 1. static The set of potential dynamic points labeled 2 and the set of uncertain points labeled 3 are combined to form the set of points to be verified, P. verify It is then sent to the second level of fine-grained verification.

[0126] Step S52: Perform a second-level refined verification based on local geometric structure constraints for each point p to be verified. j The predicted pose of the error-state Kalman filter is transformed to the global world coordinate system, and the nearest neighbor set Q is obtained in the local sub-map. near ;

[0127] When the number of nearest neighbors meets a preset threshold, the local geometric structure features of the nearest neighbor set are extracted, and eigenvalue decomposition is performed on the covariance matrix to obtain descending eigenvalues. ,by The corresponding eigenvectors are used as local surface normal vector estimates;

[0128] Define the local flatness response value:

[0129] ;

[0130] when When the value is less than a preset threshold, it is determined that no effective plane has been formed in the neighborhood; when When the distance to the point to be verified is greater than or equal to a preset threshold, the neighborhood is determined to form a valid plane, and the distance from the point to be verified to this local fitting plane is used as the geometric consistency residual. ;

[0131] Based on the geometric residual distribution of the static background point set, the noise scale σ is estimated using the robust statistic of median absolute deviation. mad :

[0132] ;

[0133] Where 1.4826 is the asymptotic consistency coefficient from MAD to Gaussian standard deviation, from which an adaptive dynamic discrimination threshold is constructed. , l is the sensitivity coefficient. j The semantic label value of the j-th point; the point to be verified, p j Perform geometric consistency determination and label correction: .

[0134] Step S53: Output the purified static point cloud set. Extract all laser points that are finally marked as static tags after two-stage cascaded filtering, and output the purified static point cloud set P. clean .

[0135] Step S6: ESKF tightly coupled fusion

[0136] The adaptive adjustment parameter set is constructed as a process noise covariance matrix and injected into the prior prediction stage of the error state Kalman filter. The static background point cloud filtered in step S5 is registered with the current frame point cloud to match the residual-driven posterior measurement update.

[0137] Specifically:

[0138] Step S61: Construct a 30-dimensional extended error state vector The independent accelerometer bias of each array IMU in step S2, and the three-dimensional angular acceleration estimation error calculated recursively from the lever arm constraint residual, are incorporated into the state space:

[0139] ;

[0140] Where δp is the position error, δv is the velocity error, δθ is the attitude angle error, and δb is the position error. a0 to δb a3 These represent the accelerometer zero bias errors of the four inertial measurement units, δb. g For gyroscope zero bias error, This represents the angular acceleration estimation error. This represents the error in the gravity vector.

[0141] Step S62: Perform adaptive process noise injection and prior prediction, and input the twelve-dimensional process noise parameter vector output from the network terminal in step S34. The process noise covariance matrix is ​​injected into the 30-dimensional process noise covariance matrix according to the block diagonal structure to dynamically update the noise sub-blocks corresponding to the position, velocity, attitude and the zero bias of the main accelerometer; using the injected and updated process noise covariance matrix and the state transition matrix constructed based on the solved angular acceleration, the prior prediction recursion of the extended error state Kalman filter is performed.

[0142] Step S63: Utilize the purified static point cloud set P output in step S53. clean Perform filtered observation updates, with the observation noise covariance based on the noise scale obtained in step S52 using robust median absolute deviation estimation. The online settings, together with the adaptive injection of process noise in step S62, constitute a two-way collaborative adaptive mechanism for process noise and observation noise.

[0143] Step S7: Pose Output and Mapping

[0144] Based on the fusion estimation results of the error state Kalman filter, the six-degree-of-freedom high-frequency pose of the carrier is output, and a global clean point cloud map after removing dynamic features is constructed incrementally at the same time.

[0145] Specifically:

[0146] Step S71: Perform state correction and high-frequency pose output. Use the error state posterior update amount output in step S63 to correct the nominal state recursively derived from the inertial measurement and directly output the carrier six-degree-of-freedom pose estimate after tight coupling and fusion at the current moment.

[0147] Step S72: Construct a clean static global map. Using the corrected six-degree-of-freedom pose estimation, transform the clean static point cloud set output in step S53 to the global world coordinate system and insert it into the global map. At the same time, associate and store the semantic segmentation category label corresponding to the scan frame with the global map points.

[0148] IV. Preferred Implementation Method

[0149] In a preferred embodiment of the present invention, the point cloud registration in step S6 shares a local geometric structure with the geometric verification in step S5. Specifically, the effective planar neighborhood and the normal vector obtained by decomposing the local flatness response value ρ in step S52 are directly reused as the point-to-planar geometric primitives required for registration, avoiding repeated neighborhood geometric structure extraction on the cleaned point cloud. The insertion frequency of the global map in step S7, the point cloud registration update frequency in step S6, and the encoder network inference frequency in step S3 are kept synchronized.

[0150] V. Experimental Verification

[0151] To verify the effectiveness and robustness of the method proposed in this invention, five comparative algorithms were introduced for comprehensive evaluation:

[0152] FAST-LIO2: Baseline tightly coupled laser inertial odometry calculation method;

[0153] FAST-LIO-A: An ablation algorithm that uses an array IMU but with a fixed process noise covariance parameter;

[0154] LSTM: Replace the adaptive module of this invention with the ablation algorithm of a traditional recurrent neural network;

[0155] A-LOAM: Pure laser odometry;

[0156] LIO-SAM: A tightly coupled algorithm based on factor graph optimization.

[0157] The experiment constructed a dynamic environment simulating a coal mine roadway in the Gazebo simulation environment, including straight sections, curves, and intersections. Dynamic objects (mine cars and miners) traversed the scene at random speeds. Evaluation was performed using the EVO tool to calculate the root mean square error of the absolute trajectory error and the 95th percentile of the cumulative distribution function.

[0158] Experimental results are as follows Figures 4 to 6 As shown:

[0159] Trajectory Comparison ( Figure 4The trajectory obtained by the method of this invention closely matches the real trajectory, and it can maintain stable and accurate pose even in areas where the motion state changes, such as turning. This result effectively demonstrates the effectiveness of the dual-end collaborative adaptive mechanism introduced by the method at both the filtering prediction and observation ends. At the prediction end, the spatial configuration of the array IMU is used to calculate angular acceleration, and the process noise parameters are adaptively adjusted by the attention network, which enhances the system's adaptability to complex motion. At the observation end, dynamic point clouds are eliminated through semantic and geometric cascade filtering, providing reliable static geometric features for pose fusion. In contrast, the comparison algorithm relies more on fixed noise parameters and has a limited mechanism for eliminating dynamic environmental features. Therefore, under the same working conditions, different degrees of trajectory drift and error accumulation are more likely to occur.

[0160] CDF curve ( Figure 5 In a 95% probability error assessment, this invention controls the error to 0.076 m, compared to 2.051 m for the baseline algorithm FAST-LIO2 and 1.433 m for the array extension algorithm FAST-LIO2-A, and 1.711 m for the conventional pure laser algorithm A-LOAM, thus reducing the global positioning error. Meanwhile, the LSTM ablation algorithm has an error as high as 2.826 m, confirming the effectiveness of the dual-end collaborative adaptive mechanism of this method and demonstrating its ability to effectively suppress sudden pose shifts caused by complex dynamic environments. Furthermore, the LIO-SAM algorithm based on a factor graph architecture achieves an accuracy of 0.056 m through global joint optimization of historical multi-frame poses; while this method, as a filtering framework relying on single-step recursion, still achieves a positioning accuracy of 0.076 m without global optimization, verifying that the method of this invention improves the overall positioning accuracy under the current filtering architecture.

[0161] Error curves of each axis ( Figure 6The present invention maintains the absolute error at the lowest level in the X, Y, and Z directions, and remains close to 0m for a long period. In the X and Y axes, FAST-LIO2, FAST-LIO-A, and A-LOAM all exhibited varying degrees of fluctuation and error accumulation, while the error of the LSTM ablation algorithm showed a continuous and dramatic increase over time. Therefore, the proposed method, through a dual-end collaborative adaptive mechanism, effectively suppresses error drift over time. In the Z-axis direction, FAST-LIO2, FAST-LIO-A, A-LOAM, and LSTM all generated significant initial errors at the initial moment. Comparing the Z-axis error curves of LIO-SAM and the proposed method, the error is minimal from the initial moment, avoiding the drastic error fluctuations of the aforementioned algorithms during the initialization phase. Furthermore, throughout the entire subsequent running cycle, the Z-axis error of the proposed method remains minimal without any rebound. This reflects that the proposed method does not require a long initialization error elimination phase, verifying its strong ability to suppress accumulated errors in each axis throughout the entire motion process and its long-term stability in pose calculation.

[0162] The experimental data above fully demonstrate that the present invention achieves excellent positioning accuracy and mapping quality in dynamic degradation scenarios through a combination of techniques including array IMU angular acceleration sensing, Transformer adaptive process noise adjustment, three-level semantic segmentation, and two-level cascaded dynamic filtering.

[0163] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A laser inertial SLAM method based on Transformer, characterized in that, Includes the following steps: S1. Acquire angular velocity and linear acceleration data collected by an array of inertial measurement units arranged in a spatial pyramid, as well as three-dimensional point cloud data collected by lidar. S2. Based on the spatial configuration constraints of the array inertial measurement unit and the kinematic principle of the lever arm, the acceleration consistency constraint residual is constructed by utilizing the measurement redundancy between the main inertial measurement unit and the auxiliary inertial measurement unit. The three-dimensional angular acceleration vector of the rigid body of the carrier is calculated in real time by recursively estimating the optimal value of the residual. S3. The angular velocity and linear acceleration data of the array inertial measurement unit are coupled with the three-dimensional angular acceleration vector in a time sequence. The multi-head self-attention encoder network with channel structure awareness is used to perform time sequence correlation modeling on the joint feature sequence within the preset sliding window, and the adaptive adjustment parameter set of the process noise covariance matrix is ​​output. S4. Based on spherical projection geometric transformation, the three-dimensional point cloud is mapped to a two-dimensional distance image space, and a lightweight semantic segmentation network is used to perform pixel-by-pixel classification to generate a three-level semantic segmentation mask containing static background class, potential dynamic class and uncertain class. S5. Construct a two-level filtering mechanism with semantic prior guidance and local geometric consistency constraints. Use the three-level semantic segmentation mask to perform the first-level semantic gating filtering. Perform the second-level geometric verification on the points to be verified after the first-level filtering based on the local flatness response value and the adaptive dynamic discrimination threshold. Output a cleaned static point cloud set. S6. Map the adaptive adjustment parameter set to a process noise covariance matrix and inject it into the prior prediction stage of the error state Kalman filter, and use the purified static point cloud to perform point cloud registration and measurement update. S7. Based on the fusion estimation results of the error state Kalman filter, output the six-degree-of-freedom high-frequency pose of the carrier and construct a global point cloud map after removing dynamic features.

2. The laser inertial SLAM method based on Transformer according to claim 1, characterized in that, In step S2, the acceleration of the auxiliary inertial measurement unit and the acceleration of the main inertial measurement unit satisfy the following lever arm relationship: ； in, Let be the acceleration of the i-th auxiliary IMU (i=1,2,3). Main IMU accelerometer Main IMU angular velocity, r i For spatial location, This represents the angular acceleration obtained by constructing an acceleration constraint residual function and recursively estimating the optimal value.

3. The laser inertial SLAM method based on Transformer according to claim 1, characterized in that, Step S3 specifically includes: S31. Stack the six-dimensional measurement vectors of the four inertial measurement units at sampling time k vertically according to a fixed spatial topology from the main inertial measurement unit to the auxiliary inertial measurement unit to construct a joint measurement vector of the twenty-four-dimensional array inertial measurement units. The three-dimensional angular acceleration vector and the twenty-four-dimensional joint measurement vector are concatenated longitudinally to construct a twenty-seven-dimensional joint feature vector at time k. ; S32. Perform time-series expansion on the 27-dimensional joint feature vector with a preset sliding window length N to construct an N×27-dimensional time-series input matrix; S33. Apply timing position encoding to the timing input matrix to obtain an encoded input sequence matrix carrying timing position information; S34. The encoded input sequence matrix is ​​fed into a multi-layer channel structure-aware multi-head self-attention encoder network. The self-attention calculation in the encoder network introduces a channel structure-aware bias matrix. The bias matrix is ​​divided into several structural sub-blocks according to the block composition rules of the 27-dimensional joint feature vector. The initial value of each sub-block is physically guided and initialized based on the Euclidean distance between the spatial position vectors of the four inertial measurement units, and end-to-end learning and updating are performed during the training phase. After being encoded layer by layer by the encoder network, a 12-dimensional process noise adaptive parameter vector is output, and the process noise covariance diagonal matrix is ​​constructed based on this. .

4. The Transformer-based laser inertial SLAM method according to claim 3, characterized in that, Step S4 specifically includes: S41. The three-dimensional point cloud collected by the lidar in the scanning frame is mapped point by point to the two-dimensional distance image space through spherical projection transformation to obtain the pixel coordinates (u, v) of each laser point in the distance image; S42. Construct a five-channel input feature tensor I from the laser point attributes projected to each valid pixel coordinate. k The five-dimensional feature vector at pixel (u, v) is: ； Where, x i、 y i z i Indicates the three-dimensional spatial coordinates of the i-th laser point in its lidar coordinate system; intensity i r represents the reflection intensity of the i-th laser point; i This represents the distance from the i-th laser point to the origin of the lidar; r max and I max These are the nominal values ​​of the maximum detection range and maximum reflection intensity of the lidar, respectively. S43. The five-channel input feature tensor is fed into a lightweight semantic segmentation encoder with the inverse residual bottleneck module as the basic unit for multi-scale hierarchical feature extraction; a multi-branch heterogeneous context aggregation module is embedded at the end of the encoder to generate a multi-scale context-enhanced semantic feature map. S44. The semantic feature map is fed into a lightweight decoder, and after step-by-step upsampling and concatenation of the shallow feature map with the dimension of the feature map, channel compression and fusion are performed to output a pixel-by-pixel semantic feature map restored to the original spatial resolution. S45. The pixel-by-pixel semantic feature map is mapped to three semantic categories using a classification head. After normalization, the posterior probability estimate of each pixel belonging to each category is obtained, and the three-level semantic segmentation mask is generated. In this context, H and W represent the height and width of the mask, respectively; label 1 represents the static background class, label 2 represents the potential dynamic class, and label 3 represents the uncertain class.

5. The laser inertial SLAM method based on Transformer according to claim 1, characterized in that, The two-level filtering mechanism specifically includes: S51. Based on the spherical projection index mapping relationship, the three-level semantic segmentation mask is traced back to the three-dimensional point cloud space, and each laser point in the original point cloud dataset is assigned a corresponding category label; the static background point set P with label 1 is retained. static The set of potential dynamic points labeled 2 and the set of uncertain points labeled 3 are combined to form the set of points to be verified, P. verify ; S52, For each point p to be verified j The predicted pose of the error-state Kalman filter is transformed to the global world coordinate system, and the nearest neighbor set Q is obtained in the local sub-map. near ; When the number of nearest neighbors meets a preset threshold, the local geometric structure features of the nearest neighbor set are extracted, and eigenvalue decomposition is performed on the covariance matrix to obtain descending eigenvalues. ,by The corresponding eigenvectors are used as local surface normal vector estimates; Define the local flatness response value: ； when When the value is less than a preset threshold, it is determined that no effective plane has been formed in the neighborhood; when When the distance to the point to be verified is greater than or equal to a preset threshold, the neighborhood is determined to form a valid plane, and the distance from the point to be verified to this local fitting plane is used as the geometric consistency residual. ; Based on the geometric residual distribution of the static background point set, the noise scale σ is estimated using the robust statistic of median absolute deviation. mad Construct an adaptive dynamic discrimination threshold , l is the sensitivity coefficient. j The semantic label value of the j-th point; the point to be verified, p j Perform geometric consistency determination and label correction: ; S53. Extract all laser points that are finally marked as static labels after two-stage filtration, and output the purified static point cloud P. clean .

6. The laser inertial SLAM method based on Transformer according to claim 1, characterized in that, Step S6 specifically includes: S61. Construct a 30-dimensional extended error state vector. The independent accelerometer bias of each array inertial measurement unit and the three-dimensional angular acceleration estimation error calculated recursively from the lever arm constraint residual are incorporated into the state space: ； Where δp is the position error, δv is the velocity error, δθ is the attitude angle error, and δb is the position error. a0 to δb a3 These represent the accelerometer zero bias errors of the four inertial measurement units, δb. g For gyroscope zero bias error, This represents the angular acceleration estimation error. This represents the error in the gravity vector. S62, The twelve-dimensional process noise adaptive parameter vector A 30-dimensional process noise covariance matrix is ​​injected using a block-based diagonal structure mapping. In the process, the position, velocity, attitude, and noise sub-blocks corresponding to the zero bias of each accelerometer are dynamically updated; S63, using the purified static point cloud set P clean Observation update for error state Kalman filtering, observation noise covariance R k The noise scale σ obtained based on the robust estimation of the median absolute deviation mad The online configuration, together with the adaptive injection of process noise covariance in step S62, constitutes a two-way collaborative adaptive mechanism for process noise and observation noise.

7. The laser inertial SLAM method based on Transformer according to claim 1, characterized in that, Step S7 specifically includes: S71. The nominal state propagated by inertial measurement is corrected using the error state posterior update amount, and the six-degree-of-freedom pose estimate of the carrier at the current moment is output. S72. Using the corrected six-degree-of-freedom pose estimation, transform each laser point in the purified static point cloud set to the global world coordinate system and insert it into the global map; at the same time, associate and store the semantic segmentation category label corresponding to the scan frame with the global map point.

8. The laser inertial SLAM method based on Transformer according to claim 1, characterized in that, In step S6, the point cloud registration and the geometric verification in step S5 share a local geometric structure: the effective planar neighborhood and local surface normal vectors determined in step S52 are reused as geometric primitives for point cloud registration; the insertion frequency of the global map in step S7, the update frequency of point cloud registration in step S6, and the inference frequency of the encoder network in step S3 are kept synchronized.

9. A Transformer-based laser inertial SLAM system, characterized in that, include: The data acquisition module consists of an array of inertial measurement units arranged in a spatial pyramid and a lidar, and is used to perform data acquisition as described in step S1 of claim 1. The angular acceleration calculation and adaptive parameter regression module is used to perform the lever arm residual recursive calculation of three-dimensional angular acceleration in step S2 as described in claim 1, and the channel structure perception attention encoder network in step S3 as described in claim 1 outputs a twelve-dimensional process noise adaptive parameter vector. The semantic dynamic filtering module is used to perform the spherical projection and lightweight semantic segmentation network to generate a three-level semantic segmentation mask in step S4 as described in claim 1, and to perform the local flatness response value and adaptive dynamic discrimination threshold filtering in step S5 as described in claim 1, and output a purified static point cloud set. The error state Kalman filter tightly coupled mapping module uses a 30-dimensional extended error state vector as the filtering space. It maps the 12-dimensional process noise adaptive parameter vector into the process noise covariance matrix in a block diagonal structure, merges the purified static point cloud set to perform recursive updates, and outputs high-frequency poses and incrementally constructs a global map using a dual-end collaborative adaptive mechanism.

10. The Transformer-based laser inertial SLAM system according to claim 9, characterized in that, The array inertial measurement unit includes one main inertial measurement unit and three auxiliary inertial measurement units. The main inertial measurement unit is installed directly above the lidar, and the three auxiliary inertial measurement units are respectively installed on the front sides and rear end of the fixed device. The four inertial measurement units together form a pyramid-shaped array in three-dimensional space. The lidar is installed above the center point of the fixed device and the center of the carrier. The lidar and the four inertial measurement units are all rigidly connected to the fixed device, and their data output ports are all connected to the communication interface of the industrial control computer.

Images