Tunnel environment-oriented adaptive positioning correction method and mobile positioning equipment

Abstract

The invention discloses an adaptive positioning correction method for a tunnel environment, and the method comprises the steps: preliminarily restoring the position of a mobile device to a tunnel axis through inertial navigation, and employing inherent static features (such as a cable support, a fireproof door, and a duct piece joint) in a tunnel as a reference; and the automatic correction of the SLAM accumulative error is realized by measuring the distance from the feature object. And meanwhile, an intelligent filtering algorithm is adopted to effectively suppress dynamic object interference, so that the positioning precision and the system robustness are remarkably improved, and reliable technical support is provided for intelligent operation and maintenance of the cable tunnel.

Description

Technical Field

[0001] This application relates to the field of tunnel positioning technology, and in particular to an adaptive positioning correction method and mobile positioning device for tunnel environments. Background Technology

[0002] In enclosed or semi-enclosed environments such as cable tunnels, positioning technologies typically face challenges such as signal blockage, multipath effects, and dynamic interference. Common positioning methods include inertial navigation systems (INS), simultaneous localization and mapping (SLAM) technology, wireless signal positioning (such as UWB, Wi-Fi, and LiFi), and visual / laser scanning modeling methods. INS calculates motion trajectories using sensors such as accelerometers and gyroscopes, but it suffers from accumulated errors requiring external correction. SLAM technology builds environmental maps and estimates its own position in real time using sensor data, but it is susceptible to error accumulation in long tunnels or environments with sparse features. Wireless positioning relies on base station or beacon deployments, and its effectiveness is limited in areas without signal coverage. Visual and laser methods rely on environmental feature matching, but they are sensitive to changes in lighting, dynamic objects, and repetitive structures. Each of these technologies has its advantages and disadvantages, and in practical applications, multi-source fusion is often necessary to improve robustness and accuracy.

[0003] Chinese patent application CN201610980005 discloses a cable tunnel positioning device and method based on LiFi technology. This technical solution involves arranging multiple optical signal transmitters with unique address codes within the cable tunnel. The location code information is transmitted using a modulation method based on the brightness variations of visible light. A mobile optical signal receiver then analyzes the address codes contained in the optical signals in real time, achieving precise positioning. The greatest advantage of this method is its full utilization of existing lighting infrastructure within the tunnel, achieving an integrated design of communication and positioning. It features high positioning accuracy, low transmission delay, and strong resistance to electromagnetic interference. However, the limitations of this method are also significant: First, the system must rely on pre-deployed dedicated optical beacons, increasing equipment investment and maintenance costs; second, visible light communication has high requirements for the direct path of light, and the signal is easily interrupted in environments with smoke, dust, or equipment obstruction, leading to decreased reliability; furthermore, this method cannot adapt to structural changes in the tunnel environment, lacks self-adaptive capabilities, and cannot achieve truly autonomous positioning without the need for markers.

[0004] Chinese patent application CN202210534669 discloses a cable tunnel positioning method based on image edge contours and hash coding. This method acquires tunnel environment images using a camera, extracts edge contour features from the environment using image processing algorithms, and then hash-codes the edge contour images to generate feature sequences. These sequences are then matched against a pre-established image location coding library to achieve location identification. The advantage of this technical solution is that it does not require additional hardware deployment, utilizing only visual information for positioning, thus reducing system complexity. However, its disadvantages are also significant: this method is highly dependent on image quality; in situations with insufficient tunnel lighting, repetitive texture features, or a large number of dynamic objects (such as moving personnel or equipment), edge feature extraction is easily interfered with, leading to matching failures; hash coding is sensitive to changes in viewing angle, and error accumulation is prone to occur during long-term operation; furthermore, this method requires the pre-establishment of a complete image location coding library, resulting in high costs for database creation and maintenance.

[0005] Chinese patent application number CN202210246542 discloses a cable tunnel positioning system based on comprehensive analysis of positioning data. The system consists of a positioning base station master station, positioning base station terminals, a data analysis server, and a handheld positioning device. It achieves signal coverage through wireless networking and uses a multi-base station ranging and data fusion algorithm to achieve precise positioning. The system's advantages include a large coverage area, high positioning reliability, and low power consumption. However, this technical solution also has significant drawbacks: it requires the deployment of a large number of base station devices, leading to high system construction and maintenance costs; in complex tunnel environments, wireless signals are easily affected by multipath effects and electromagnetic interference, causing a decrease in positioning accuracy; furthermore, the system struggles to effectively filter interference signals from moving objects, resulting in insufficient positioning stability in environments with frequent personnel and equipment movement.

[0006] Chinese patent application CN202311281940 discloses a cable tunnel positioning method based on 3D modeling and image recognition technology. This method first uses a laser scanner to perform a detailed scan of the tunnel, establishing a high-precision 3D tunnel model on a computer terminal. By selecting and locating typical reference objects within the tunnel, a reference coordinate system is established. Then, a 3D map marked with positioning information is generated based on the 3D model. Finally, by capturing real-time photos of the tunnel, a 3D point cloud is reconstructed using deep learning technology and matched with the pre-established model to achieve location identification. The advantage of this method is that it can provide accurate spatial coordinate information and rich environmental data. However, its limitations include: reliance on high-precision laser scanning equipment, resulting in high hardware costs; high computational complexity in 3D modeling and point cloud matching, placing strict demands on processor performance; difficulty in adapting to environmental changes such as lighting variations and structural corrosion within the tunnel; and, in particular, a lack of effective filtering mechanisms for interference from dynamic objects, which can easily lead to mismatches in practical applications.

[0007] Chinese patent application CN202410936792 discloses a tunnel positioning method based on the fusion of MR fingerprints and inertial navigation. This method acquires multiple network parameters (MR measurement reports) reported by the user terminal, combines them with predicted location data output by the inertial navigation system, matches the best fingerprint data from an MR fingerprint database, and achieves positioning correction through similarity analysis and data fusion. The advantage of this technical solution is that it eliminates the need for separate positioning equipment, achieving good economy and flexibility by utilizing existing network infrastructure. However, this method also has significant drawbacks: it is highly dependent on the quality of wireless signal coverage, and its positioning performance significantly degrades in areas with weak or abrupt signal changes; the inherent cumulative error of the inertial navigation system is only corrected by MR data, with limited effectiveness, and significant positioning drift may still occur in long-distance tunnels; furthermore, the MR fingerprint database requires regular updates and maintenance, increasing operating costs.

[0008] The aforementioned existing technologies all suffer from problems to varying degrees, such as reliance on external markers, the need for dedicated equipment deployment, poor environmental adaptability, and the inability to effectively overcome error accumulation and dynamic interference.

[0009] High-precision positioning of cable tunnels is crucial for intelligent inspection and safe operation and maintenance. With the continuous expansion of urban underground pipe networks, traditional positioning methods relying on manual markers or dedicated beacons are no longer sufficient to meet the application needs of long-distance and complex environments. In recent years, advanced technologies such as Simultaneous Localization and Mapping (SLAM), inertial navigation, and multi-sensor fusion have been gradually applied to tunnel environments, driving the development of positioning technology towards automation and intelligence. However, in practical applications, due to the characteristics of cable tunnels, such as repetitive structural features, significant changes in lighting conditions, and numerous dynamic interference factors, existing technologies still face significant challenges: SLAM systems accumulate errors during long-distance operation, leading to a decrease in positioning accuracy; reliance on pre-deployed hardware markers increases system maintenance costs and engineering complexity; and dynamic objects such as personnel and equipment within the tunnel can easily cause visual feature mismatches, affecting the accuracy and reliability of the positioning system. Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this application provides an adaptive positioning correction method and mobile positioning device for tunnel environments. Its advantages are that it innovatively integrates inertial navigation and SLAM technology, utilizes the inherent structural characteristics of tunnels for error correction and dynamic interference filtering, and achieves high-precision adaptive positioning that does not rely on markers.

[0011] The technical solution of this application is as follows:

[0012] On the one hand, this application provides an adaptive positioning correction method for tunnel environments, including the following steps:

[0013] S1: The mobile device is equipped with an inertial measurement unit, which collects acceleration and angular velocity data and uses a dead reckoning algorithm to make a preliminary estimate of the device's position and attitude in the tunnel;

[0014] S2: Project the initially estimated equipment movement trajectory onto the tunnel axis direction to form a preliminary positioning result under axis constraints;

[0015] S3: Use visual sensors and lidar to collect images and reflection duration data, identify inherent static features through deep learning models and point cloud processing algorithms, and measure the relative distance between the device and the features;

[0016] S4: Match and compare the observed distance of the feature with the theoretical position of the same feature in the pre-constructed 3D tunnel model. For each successfully matched feature, calculate the difference between its observed distance and theoretical distance to form an observation error vector. Use the observation error vector data to correct the SLAM cumulative error in real time.

[0017] Furthermore, in step S4, the observation error vector is obtained in the following way:

[0018] Let the theoretical location of a certain feature in the tunnel model be... The device pose estimated by the current SLAM Transform it to the current sensor coordinate system to obtain the predicted observation position. The actual observation location is The observation error is defined as follows: , This represents the transformation matrix from the world coordinate system to the camera coordinate system.

[0019] Furthermore, in step S4, the extended Kalman filter algorithm is used to incorporate the error vector into the state update process to correct the device pose; in the prediction step, the current state is predicted based on the data from the inertial measurement unit using the system state equation. and its covariance The above observation errors As measurement information, the Kalman gain is calculated. Correct the predicted state:

[0020] = +K·e

[0021]

[0022] in, This represents the posterior state estimate, which is the state estimate obtained after incorporating the observations at time k. It is the posterior state covariance. It is the identity matrix. The Jacobian matrix of the observation model with respect to the state vector represents the sensitivity of the observations to state changes.

[0023] Furthermore, it also includes step S5:

[0024] Maintain a continuously updated local feature map, which contains all feature points identified in the past few frames and their 3D positions; when a new frame of data arrives, perform feature matching to associate the features observed in the current frame with the features in the local map;

[0025] For each matched feature point, its motion consistency is checked; in consecutive frames, features whose reprojection error value is within a set threshold range are judged as static features, otherwise they are judged as non-static features; for non-static features, they are removed from the feature matching library in the current frame and several subsequent frames. At the same time, the non-static feature information is fed back to the SLAM front end, and the non-static features are deleted from the local map.

[0026] Furthermore, it also includes step S6:

[0027] By integrating historical and current corrected pose data, a globally consistent high-precision map is constructed, and the real-time position of the device is optimized.

[0028] A pose graph model is constructed using keyframe poses as nodes and inter-frame relative transformations as edges. Whenever a new corrected keyframe is added, it is added to the graph, and a constraint relationship with existing nodes is established.

[0029] By solving a nonlinear optimization problem with the objective of minimizing the weighted sum of squared errors of all edge constraints, all historical poses are jointly optimized to obtain a globally consistent optimal trajectory and map; the optimization result is output as the final positioning information of the device.

[0030] Furthermore, the keyframes of the SLAM system are determined using the following method:

[0031] The current frame is marked as a keyframe when the device's movement distance relative to the previous keyframe exceeds a set threshold; the current frame is marked as a keyframe when the device's rotation angle relative to the previous keyframe exceeds a set threshold; if the number of new features observed in the current frame is significantly greater than that of existing keyframes, or the feature overlap with the previous keyframe is less than a set threshold, then it is set as a keyframe; to ensure system real-time performance, if several consecutive frames do not meet the above conditions, the current frame is forcibly set as a keyframe to avoid error accumulation due to prolonged periods without keyframes. The current frame is marked as a keyframe when any of the conditions are met.

[0032] Furthermore, step S7 is included: outputting the final corrected positioning information and displaying the device position in real time in the three-dimensional tunnel model.

[0033] This invention provides an automatic positioning and correction method for cable tunnels that does not rely on markers. The method uses inertial navigation to initially restore the position of the mobile device to the tunnel axis, and utilizes inherent static features within the tunnel (such as cable supports, fire doors, and segment joints) as references. Automatic correction of SLAM cumulative errors is achieved by measuring the distance to these features. Simultaneously, an intelligent filtering algorithm effectively suppresses interference from dynamic objects, significantly improving positioning accuracy and system robustness, providing reliable technical support for the intelligent operation and maintenance of cable tunnels.

[0034] On the other hand, this application provides a mobile positioning device, which is equipped with an inertial measurement unit, a visual sensor, a lidar and a core processing unit, wherein the core processing unit includes a central processing unit and an extended Kalman filter algorithm module;

[0035] The process involves several steps: First, the central processing unit (CPU) collects acceleration and angular velocity data from the mobile device. Based on dead reckoning algorithms, it initially estimates the device's position and attitude within the tunnel. The estimated trajectory is then projected onto the tunnel's axis, forming a preliminary positioning result under axial constraints. Next, visual sensors and lidar are used to acquire image and reflection duration data. Deep learning models and point cloud processing algorithms are employed to identify inherent static features and measure the relative distance between the device and these features. Finally, the observed distance of each feature is matched and compared with its theoretical position in a pre-constructed 3D tunnel model. For each successfully matched feature, the difference between its observed and theoretical distances is calculated, forming an observation error vector. Finally, an extended Kalman filter module uses this observation error vector data to correct the accumulated SLAM error in real time.

[0036] In summary, the beneficial effects of this application are as follows:

[0037] 1. Effectively suppresses SLAM accumulated errors and improves long-distance positioning accuracy;

[0038] Existing SLAM schemes are prone to accumulating errors in long-distance tunnels with repetitive structures and lack effective online correction mechanisms. This invention achieves preliminary axis-constrained positioning through an inertial navigation system, and combines lidar and visual sensors to accurately measure and match stable features such as cable supports and tunnel segment joints. The observed distances of these features are compared with the theoretical distances in the tunnel model, and SLAM pose errors are calculated and corrected in real time, thus significantly suppressing drift and achieving stable, high-precision positioning in tunnels hundreds of meters long.

[0039] 2. Eliminate dependence on external markers, enhancing applicability and economy;

[0040] Unlike methods that require pre-deployment of QR codes, UWB base stations, or LiFi beacons, this invention directly utilizes existing structural features in the tunnel as natural references, eliminating the need for additional deployment and maintenance of marking equipment. This not only reduces implementation costs but also avoids positioning failures caused by damaged, contaminated, or obstructed markers, significantly improving applicability and reliability in real-world operation and maintenance environments.

[0041] 3. Possesses dynamic interference recognition capability, improving robustness in complex environments.

[0042] Traditional visual SLAM solutions are prone to feature mismatches when disturbed by dynamic objects (such as people and equipment). This invention, through consistency checks and motion attribute analysis within a multi-sensor fusion framework, can effectively distinguish between dynamic objects and static features, and automatically filter out interference information during pose estimation, thus maintaining positioning stability and continuity even in scenarios with frequent personnel and equipment activity, such as inspections and construction. Attached Figure Description

[0043] Figure 1 This is a schematic diagram illustrating the steps of the adaptive positioning correction method for tunnel environments proposed in this application;

[0044] Figure 2 This is a schematic diagram of static feature recognition and ranging in this application;

[0045] Figure 3 This is a schematic diagram of the mobile positioning device of this application. Detailed Implementation

[0046] The specific embodiments of this application are described in detail below with reference to the accompanying drawings.

[0047] Example 1: An adaptive positioning correction method for tunnel environments. This method sequentially includes seven processes: initial inertial navigation positioning, tunnel axis projection, static feature recognition and ranging, SLAM error calculation and correction, dynamic interference filtering, multi-frame data fusion, and positioning result output. A mobile device equipped with a multi-sensor module estimates its pose initially through inertial navigation while traveling through the tunnel, and uses inherent static features of the tunnel for distance measurement, providing data support for error correction. By calculating the deviation between the observed distance and the theoretical distance of the features, the accumulated SLAM error is corrected in real time, and a filtering algorithm is used to suppress interference from dynamic objects, ultimately outputting high-precision positioning information. Each process is described in detail below:

[0048] S1 Inertial Navigation Preliminary Positioning: The mobile device is equipped with an inertial measurement unit, which collects acceleration and angular velocity data and estimates the device's position and attitude in the tunnel based on dead reckoning algorithms; this process provides initial attitude estimation for subsequent feature matching and error correction.

[0049] S2 Tunnel Axis Projection: To suppress inertial navigation drift in the vertical direction, the initially estimated equipment motion trajectory is projected onto the tunnel axis direction to form a preliminary positioning result under axis constraints;

[0050] S3 Static Feature Recognition and Ranging: For example Figure 2 As shown, visual sensors (cameras) and lidar are used to collect images and reflection duration data. Deep learning models and point cloud processing algorithms are used to identify inherent static features, including cable brackets, fire doors, and pipe segment joints, and the relative distance between the equipment and the features is measured.

[0051] S4 SLAM Error Calculation and Correction: The observed distance of a feature is matched and compared with the theoretical position of the same feature in the pre-constructed 3D tunnel model. For each successfully matched feature, the difference between its observed distance and theoretical distance is calculated to form an observation error vector. The observation error vector data is used to correct the SLAM cumulative error in real time.

[0052] The observation error vector is obtained as follows: Let the theoretical position of a certain feature in the tunnel model be... The device pose estimated by the current SLAM Transform it to the current sensor coordinate system to obtain the predicted observation position. The actual observation location is The observation error is defined as follows: , This represents the transformation matrix from the world coordinate system to the camera coordinate system.

[0053] Subsequently, the extended Kalman filter algorithm is used to incorporate this error vector into the state update process to correct the device pose. The state vector of the extended Kalman filter typically includes the device's position, attitude, velocity, and sensor bias, denoted as […]. In the prediction step, the current state is predicted based on data from the inertial measurement unit using the system state equation. and its covariance The above observation errors As measurement information, the Kalman gain is calculated. Correct the predicted state:

[0054] = +K·e

[0055]

[0056] in, This represents the posterior state estimate, which is the state estimate obtained after incorporating the observations at time k. It is the posterior state covariance. It is the identity matrix. The Jacobian matrix of the observation model with respect to the state vector represents the sensitivity of the observations to state changes.

[0057] S5 Dynamic Interference Filtering:

[0058] Dynamic interference filtering is mainly achieved through multi-frame data association and motion consistency verification. First, a continuously updated local feature map is maintained, which contains all feature points identified in the past few frames and their 3D positions. When a new frame of data arrives, feature matching is performed, associating the features observed in the current frame with the features in the local map.

[0059] For each matched feature point, check its motion consistency; the core discrimination method is: the motion trajectory of static feature points in multiple frames should be completely explained by the camera's own motion; while the motion trajectory of dynamic feature points will deviate from the trajectory induced by the camera motion.

[0060] Specifically, the reprojection error can be determined by calculating its temporal stability: in consecutive frames, the reprojection error of static features (the difference between the observed position and the position projected according to the camera pose) should remain small and stable; while the reprojection error of dynamic features will exhibit abnormal or abrupt patterns. In this embodiment, in consecutive frames, features whose reprojection error values ​​are within a set threshold range are determined to be static features, otherwise they are determined to be non-static features. For non-static features, they are removed from the feature matching library in the current frame and several subsequent frames. At the same time, the non-static feature information is fed back to the SLAM front end, and the non-static features are deleted from the local map, thereby ensuring from the source that the features used for mapping and localization are all static and reliable features, greatly enhancing the robustness of the system in complex dynamic environments.

[0061] S6 multi-frame data fusion:

[0062] Multi-frame data fusion integrates historical and current corrected pose data to build a globally consistent high-precision map and optimize the device's real-time position.

[0063] First, this method determines keyframes for the SLAM system using the following strategies: When the device's movement distance relative to the previous keyframe exceeds a set threshold, the current frame is marked as a keyframe; when the device's rotation angle relative to the previous keyframe exceeds a set threshold, the current frame is marked as a keyframe; if the number of new features observed in the current frame is significantly greater than that of existing keyframes, it is determined by a pre-set difference threshold or ratio threshold, or if the feature overlap with the previous keyframe is less than a set threshold, then it is marked as a keyframe; to ensure system real-time performance, if several consecutive frames do not meet the above conditions, the current frame is forcibly marked as a keyframe to avoid error accumulation due to prolonged periods without keyframes. The current frame is marked as a keyframe when any of the above conditions are met.

[0064] A pose graph model is constructed using keyframe poses as nodes and inter-frame relative transformations as edges. Whenever a new corrected keyframe is added, it is added to the graph, and a constraint relationship with existing nodes is established.

[0065] By solving a nonlinear optimization problem with the objective of minimizing the weighted sum of squared errors of all edge constraints, all historical poses are jointly optimized to obtain a globally consistent optimal trajectory and map; the optimization result is output as the final positioning information of the device.

[0066] S7 Positioning Result Output: Outputs the final corrected positioning information and displays the device position in real time in the 3D tunnel model, providing accurate spatial reference for operation and maintenance.

[0067] Example 2: A mobile positioning device, referenced Figure 3 The mobile positioning device is equipped with an inertial measurement unit, a visual sensor, a lidar and a core processing unit. The core processing unit includes a central processing unit, an extended Kalman filter algorithm module and a dynamic interference filtering and data fusion module.

[0068] The process involves several steps: The central processing unit (CPU) collects acceleration and angular velocity data from the mobile device and uses a dead reckoning algorithm to initially estimate the device's position and attitude within the tunnel. The estimated trajectory is projected onto the tunnel's axis to form a preliminary positioning result under axial constraints. Visual sensors and lidar are used to acquire image and reflection duration data. Deep learning models and point cloud processing algorithms are employed to identify inherent static features and measure the relative distance between the device and these features. The observed distances of these features are matched and compared with their theoretical positions in a pre-constructed 3D tunnel model. For each successfully matched feature, the difference between its observed and theoretical distances is calculated to form an observation error vector. An extended Kalman filter module uses this observation error vector data to correct the accumulated SLAM error in real time. A dynamic interference filtering and data fusion module is used for dynamic interference and multi-frame data fusion. The specific steps are the same as in Example 1.

[0069] The mobile positioning device is also equipped with a power and management module and a data communication interface, which is used to output positioning results.

[0070] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the protection scope of this application.

Claims

1. An adaptive positioning correction method for tunnel environments, characterized in that, Includes the following steps: S1: The mobile device is equipped with an inertial measurement unit, which collects acceleration and angular velocity data and uses a dead reckoning algorithm to make a preliminary estimate of the device's position and attitude in the tunnel; S2: Project the initially estimated equipment movement trajectory onto the tunnel axis direction to form a preliminary positioning result under axis constraints; S3: Use visual sensors and lidar to collect images and reflection duration data, identify inherent static features through deep learning models and point cloud processing algorithms, and measure the relative distance between the device and the features; S4: Match and compare the observed distance of the feature with the theoretical position of the same feature in the pre-constructed 3D tunnel model. For each successfully matched feature, calculate the difference between its observed distance and theoretical distance to form an observation error vector. The cumulative error of SLAM is corrected in real time using observation error vector data.

2. The adaptive positioning correction method for tunnel environments according to claim 1, characterized in that, In step S4, the observation error vector is obtained in the following way: Let the theoretical location of a certain feature in the tunnel model be... The device pose estimated by the current SLAM Transform it to the current sensor coordinate system to obtain the predicted observation position. The actual observation location is The observation error is defined as follows: , This represents the transformation matrix from the world coordinate system to the camera coordinate system.

3. The adaptive positioning correction method for tunnel environments according to claim 1, characterized in that, In step S4, the extended Kalman filter algorithm is used to incorporate the error vector into the state update process to correct the device pose. In the prediction step, the current state is predicted based on the data from the inertial measurement unit using the system state equation. and its covariance The above observation errors As measurement information, the Kalman gain is calculated. Correct the predicted state: = +K·e ； in, This represents the posterior state estimate, which is the state estimate obtained after incorporating the observations at time k. It is the posterior state covariance. It is the identity matrix. This is the Jacobian matrix of the observation model for the state vector.

4. The adaptive positioning correction method for tunnel environments according to claim 1, characterized in that, It also includes step S5: Maintain a continuously updated local feature map, which contains all feature points identified in the past few frames and their 3D positions; when a new frame of data arrives, perform feature matching to associate the features observed in the current frame with the features in the local map; For each matched feature point, its motion consistency is checked; in consecutive frames, features whose reprojection error value is within a set threshold range are judged as static features, otherwise they are judged as non-static features; for non-static features, they are removed from the feature matching library in the current frame and several subsequent frames. At the same time, the non-static feature information is fed back to the SLAM front end, and the non-static features are deleted from the local map.

5. The adaptive positioning correction method for tunnel environments according to claim 1, characterized in that, It also includes step S6: By integrating historical and current corrected pose data, a globally consistent high-precision map is constructed, and the real-time position of the device is optimized. A pose graph model is constructed using keyframe poses as nodes and inter-frame relative transformations as edges. Whenever a new corrected keyframe is added, it is added to the graph, and constraints are established between it and existing nodes. By solving a nonlinear optimization problem with the objective of minimizing the weighted sum of squared errors of all edge constraints, all historical poses are jointly optimized to obtain a globally consistent optimal trajectory and map; the optimization result is output as the final positioning information of the device.

6. The adaptive positioning correction method for tunnel environments according to claim 5, characterized in that, Keyframes of the SLAM system are determined using the following method: The current frame is marked as a keyframe when the device's movement distance relative to the previous keyframe exceeds a set threshold; the current frame is marked as a keyframe when the device's rotation angle relative to the previous keyframe exceeds a set threshold; if the number of new features observed in the current frame is significantly greater than that of existing keyframes, or the feature overlap with the previous keyframe is less than a set threshold, then it is set as a keyframe; to ensure system real-time performance, if several consecutive frames do not meet the above conditions, the current frame is forcibly set as a keyframe to avoid error accumulation due to prolonged periods without keyframes. The current frame is marked as a keyframe when any of the conditions are met.

7. The adaptive positioning correction method for tunnel environments according to claim 1, characterized in that, It also includes step S7: outputting the final corrected positioning information and displaying the device position in real time in the three-dimensional tunnel model.

8. A mobile positioning device, characterized in that, The mobile positioning device is equipped with an inertial measurement unit, a vision sensor, a lidar and a core processing unit. The core processing unit includes a central processing unit and an extended Kalman filter algorithm module. The process involves several steps: First, the central processing unit (CPU) collects acceleration and angular velocity data from the mobile device. Based on dead reckoning algorithms, it initially estimates the device's position and attitude within the tunnel. The estimated trajectory is then projected onto the tunnel's axis, forming a preliminary positioning result under axial constraints. Next, visual sensors and lidar are used to acquire image and reflection duration data. Deep learning models and point cloud processing algorithms are employed to identify inherent static features and measure the relative distance between the device and these features. Finally, the observed distance of each feature is matched and compared with its theoretical position in a pre-constructed 3D tunnel model. For each successfully matched feature, the difference between its observed and theoretical distances is calculated, forming an observation error vector. Finally, an extended Kalman filter module uses this observation error vector data to correct the accumulated SLAM error in real time.

Images