Unmanned aerial vehicle positioning method and system for urban rail transit vehicle inspection

Abstract

The invention discloses an unmanned aerial vehicle positioning method and system for urban rail transit vehicle inspection, and relates to the technical field of rail transit vehicle inspection. The method comprises the following steps: step 1, arranging bar code labels with unique coding characteristics at key parts of a bogie, a vehicle door and a pantograph of an urban rail transit vehicle; step 2, the unmanned aerial vehicle carries out self-inspection and preset inspection path parameter loading, and internal and external parameters of the depth camera and the scanning angle and resolution of the laser radar are calibrated; and step 3, based on initial pose information provided by the inertial measurement unit and the visual odometer, flying to a target correction point along a preset path. The position deviation is calculated in real time through a kernel correlation filtering algorithm, and accurate adjustment of the attitude of the unmanned aerial vehicle is realized by using multi-channel PID servo control, so that the accumulative error of the SLAM technology is effectively inhibited, and the centimeter-level positioning accuracy in long-distance inspection is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of rail transit vehicle inspection technology, and in particular to a drone positioning method and system for urban rail transit vehicle inspection. Background Technology

[0002] With the rapid expansion of urban rail transit networks, routine vehicle inspections have become a core component in ensuring operational safety and efficiency. Traditional inspections rely heavily on manual labor during nighttime inspection windows, facing challenges such as high workload, tight deadlines, and a high risk of missed inspections, making it difficult to meet the high-density, high-frequency maintenance demands. In recent years, ground inspection robots have been introduced for inspecting undercarriage components, alleviating some of the manpower burden. However, their mobility is limited by tracks or planar structures, making it impossible to effectively cover critical areas such as the side walls, passenger compartments, and roof, resulting in an overall insufficient level of automation.

[0003] Among them, drones, with their advantages of strong three-dimensional spatial mobility, flexible vertical take-off and landing, and hovering observation capabilities, are considered an ideal vehicle for breaking through existing blind spots in inspections. However, rail transit depots are mostly enclosed indoor environments where GPS signals are severely attenuated or even completely absent, and there are numerous sources of electromagnetic interference in the surrounding area, causing conventional satellite navigation to fail. Although Simultaneous Localization and Mapping (SLAM) technology can provide pose estimation in the absence of GPS, in the depot environment with high structural repetition and sparse texture features, it is susceptible to cumulative errors and interference from dynamic obstacles, resulting in significant positioning drift and difficulty in maintaining centimeter-level path accuracy, which in turn affects the accurate focusing and image acquisition of key components such as bogies and pantographs.

[0004] In existing technologies, positioning schemes relying solely on inertial measurement units (IMUs) or visual odometry suffer from insufficient stability, while map matching methods based on high-precision lidar face problems such as high cost and computational burden. Furthermore, the lack of an active correction mechanism integrated with the vehicle's structural features makes it difficult for UAVs to continuously calibrate their position during long-distance inspections. Therefore, there is an urgent need for a highly robust positioning method that integrates environmental perception, target recognition, and closed-loop feedback control to achieve stable, accurate, and repeatable automatic inspection flights in GPS-free indoor rail transit scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a UAV positioning method and system for urban rail transit vehicle inspection. In a first aspect, the present invention provides a drone positioning method for urban rail transit vehicle inspection, comprising: Step 1: Install barcode tags with unique coding features on key parts of the urban rail transit vehicle, such as the bogie, doors, and pantograph, to form a network of correction points covering the entire outer surface of the vehicle. Step 2: Load the UAV self-test and preset inspection path parameters, and calibrate the internal and external parameters of the depth camera and the scanning angle and resolution of the LiDAR; Step 3: The UAV receives the stepping flight command from the ground control station and flies along the preset path to the target correction point based on the initial pose information provided by the inertial measurement unit and visual odometry. Step 4: The target tracking algorithm based on the kernel correlation filter principle constructs a cyclic matrix in the frequency domain, solves for the center position of the barcode label, and compares it with the preset theoretical position to generate a position deviation vector; Step 5: Input the position deviation vector into the proportional-integral-derivative servo control algorithm to adjust the pitch angle, roll angle, yaw angle and vertical altitude of the UAV; Step 6: After completing the multimodal data acquisition of the current correction point, automatically fly to the next inspection point and repeat steps 3 to 5 until the inspection of key parts of the entire train is completed. Step 7: During the inspection, the collected multimodal data is transmitted to the ground control station. After the inspection is completed, the UAV autonomously returns to the charging station and performs data integrity verification.

[0006] Preferably, step 1 includes: The barcode labels are made of highly reflective materials to ensure high readability even in low light or partially obscured conditions. The barcode labels are deployed according to the grid principle, with a correction point set every 3m to 8m along the longitudinal direction of the car body. A total of no less than 20 key points are deployed throughout the train, covering the bottom of the bogie, the door rails, the pantograph base, and the vulnerable or critical functional components of the air conditioning unit casing. Each barcode label is assigned a globally unique ID, which is bound to the vehicle model, group number, component type and theoretical three-dimensional coordinates, and stored in the inspection database of the ground control station; The barcode label is coated with an anti-ultraviolet and anti-fouling coating.

[0007] Preferably, step 2 includes: The UAV self-test process includes power supply voltage detection, motor health status diagnosis, communication link connectivity test, and sensor data synchronization clock verification. The UAV is equipped with a lidar, a depth camera, motors, a stabilization module, and an onboard computer; the onboard computer supports multi-threaded parallel processing of sensor data streams; the depth camera simultaneously outputs color images and depth maps; the lidar has a horizontal field of view of 360 degrees, a vertical field of view of 30 degrees, and an effective detection range of up to 20 meters. Sensor calibration was completed in a dedicated calibration field. The depth camera obtained the intrinsic parameter matrix and distortion coefficients using the Zhang Zhengyou checkerboard method, and the extrinsic parameters were aligned with the UAV's body coordinate system using a rigid body transformation matrix. The lidar used plane fitting and ICP registration algorithms to unify the UAV's coordinate system with the depth camera's coordinate system, forming a fused perception coordinate system.

[0008] Preferably, step 3 includes: The stepping flight command includes the next target point ID, theoretical coordinates, desired hovering altitude, and flight speed; the UAV inertial measurement unit outputs triaxial acceleration and angular velocity, which are then filtered by complementary filtering and zero-velocity correction to generate short-term stable pose increments; the visual odometry processes continuous frame depth images, extracts ORB feature points and performs optical flow tracking, solves relative motion using the PnP (Perspective-n-Point, PnP) algorithm to compensate for the drift of the inertial measurement unit, and generates initial pose information based on the pose increments; The UAV calculates navigation vectors based on initial pose information and target correction point coordinates, generates quaternion attitude commands, and adjusts the rotational speeds of each rotor to achieve smooth trajectory tracking. The UAV's flight speed between points is set between 0.5 m / s and 2 m / s, dynamically adjusted according to environmental complexity: high-speed mode is used in open areas, and the speed is reduced to below 0.8 m / s in narrow passages or near obstacles. Simultaneously, the UAV continuously runs a real-time obstacle avoidance algorithm based on LiDAR, automatically generating a detour path upon detecting an obstacle. This real-time obstacle avoidance algorithm uses a data structure combining a grid map and an octree, updating obstacle occupancy probabilities in real time, and employs an improved A / B algorithm for path planning. The algorithm minimizes the detour distance while ensuring safety; when the drone enters the target correction point within 3m, it triggers the barcode label recognition preparation state, the depth camera switches to high gain mode, and the lidar focuses on the cone-shaped area 2m in front.

[0009] Preferably, step 4 includes: After the barcode label is acquired, it first undergoes adaptive histogram equalization and Gaussian filtering to suppress uneven lighting and noise. Next, a sliding window and Haar-like feature classifier are used for coarse localization of the barcode label. Morphological closing operations are then used to connect broken edges caused by lighting, shadows, or partial occlusion in the coarsely localized area, accurately outlining the barcode label boundary. Then, a target tracking algorithm based on kernel correlation filtering is used to construct a cyclic matrix in the frequency domain. Discrete Fourier transform is used to convert the convolution operation into point-by-point multiplication to obtain the center position of the barcode label. Finally, the decoding module binarizes, denoises, and measures the bar width of the center position of the barcode label to parse out a unique ID code, which is then compared and verified with a pre-stored database. After successful verification, the three-dimensional coordinates of the barcode center point in the camera coordinate system are calculated using a pinhole imaging model back projection and compared with a preset theoretical position to generate a position deviation vector.

[0010] Preferably, step 5 includes: The position deviation vector is decomposed into four control channels: roll angle control, pitch angle control, vertical thrust control, and yaw angle control; each channel independently runs a proportional-integral-derivative servo control algorithm, the mathematical expression of which is: ; In the formula, For the first Output of each control cycle For the current deviation, The sampling period This is the proportionality coefficient. For the integral coefficient, These are the differential coefficients, and the control period is 50ms. This is a proportional term used to correct for deviations; This is the integral term, used to eliminate steady-state error; This is the differential term, used to suppress overshoot and oscillation; The system continuously monitors the normalized position and scale factor of the barcode label in the image. When the normalized position deviation is less than 0.05 and the scale factor fluctuation is less than 5%, it determines that the precise positioning is completed and triggers the high-definition image acquisition command. At this time, the depth camera switches to the highest resolution mode, and the LiDAR performs local high-density scanning to acquire multimodal data of key components, ensuring that the shooting angle and focal length meet the geometric requirements of defect detection.

[0011] Preferably, step 6 includes: The ground control station receives the actual positioning coordinates of each point in real time and compares them globally with the theoretical path. If a systematic offset is found, path offset compensation parameters can be dynamically issued to update the theoretical coordinates of subsequent points. The system provides a visual display of the drone's real-time location, flight path, sensor data, and barcode recognition status, and supports remote updates of inspection path parameters and emergency intervention control.

[0012] Preferably, step 7 includes: Multimodal data is compressed using H.265 encoding, supplemented by forward error correction codes, to ensure reliable transmission in complex electromagnetic environments; After the inspection is completed, the drone receives the return command and, based on the last known inspection location and the local map constructed by LiDAR SLAM, plans the shortest safe path back to the charging station. The top of the charger nest is equipped with high-contrast visual markings for final docking of drones; the charger nest has an automatic docking function, the charging interface is magnetically designed, a single full charge takes 40 minutes, and it integrates environmental monitoring sensors.

[0013] Preferably, it also includes an analysis module based on historical inspection data; the analysis module uses machine learning algorithms to predict defects in the multimodal data of collected key components and generates inspection reports and maintenance suggestions; the analysis module is deployed at the ground control station, uses a convolutional neural network model, takes multimodal data of key components as input, and outputs the probability distribution of cracks, corrosion, and loosening defects to assist maintenance personnel in decision-making.

[0014] Secondly, the present invention also provides a UAV positioning system for urban rail transit vehicle inspection, which applies a UAV positioning method for urban rail transit vehicle inspection as described above. The system includes an identification deployment module, a flight platform module, a navigation control module, a visual recognition module, a servo positioning module, and a ground control module. The identification deployment module is used to deploy barcode labels with unique codes on the surfaces of key components such as the vehicle bogie, doors, and pantograph, forming a network of correction points covering the entire outer surface of the vehicle. The flight platform module is used to carry sensor equipment and perform autonomous flight missions, including performing system self-checks, loading inspection paths, and completing multi-sensor calibration. The navigation control module is used to generate flight commands based on preset paths and sensor fusion data to achieve point-to-point autonomous navigation and real-time obstacle avoidance. The visual recognition module is used to collect barcode labels, decode and locate them, and calculate the position deviation vector between the actual position and the theoretical coordinates. The servo positioning module is used to receive the position deviation vector and adjust the attitude of the UAV through a multi-channel PID control algorithm to achieve centimeter-level precise hovering at the inspection point. The ground control module includes an analysis module for mission planning, data reception and storage, path correction decisions, and controlling the UAV's return and data verification.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention constructs a reliable visual positioning reference network by deploying uniquely coded barcode tags on the surface of key components of the inspected vehicle, providing the UAV with an absolute position reference independent of external satellite signals. During inspection, the UAV deeply integrates data from multiple sensors, including inertial measurement units, visual odometry, and lidar, achieving robust initial positioning in environments with high structural repeatability and sparse texture features. It also employs a closed-loop correction mechanism based on barcode recognition, using a kernel correlation filtering algorithm to calculate position deviation in real time and utilizing multi-channel PID servo control to precisely adjust the UAV's attitude, effectively suppressing the accumulated error of SLAM technology and ensuring centimeter-level positioning accuracy in long-distance inspections. Furthermore, this invention features global path dynamic correction and intelligent obstacle avoidance capabilities, significantly improving operational safety and system robustness in complex environments.

[0016] In summary, this invention not only reduces reliance on high-cost navigation equipment, but also achieves full automation and repeatability of the inspection process, providing an efficient and reliable technical means for the operation and maintenance of urban rail transit vehicles, and has significant engineering application value. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a drone positioning method for urban rail transit vehicle inspection.

[0018] Figure 2 This is a schematic diagram of a drone positioning system used for the inspection of urban rail transit vehicles. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0020] Example 1 Please see Figure 1This embodiment provides a UAV positioning method for urban rail transit vehicle inspection, including: Step 1: Pre-set inspection routes and signage layout; deploy barcode labels with unique coding characteristics on key parts of the urban rail transit vehicle such as bogies, doors, and pantographs to form a network of correction points covering the entire outer surface of the vehicle. Specifically, the barcode labels are made of high-reflectivity material, measuring 10cm × 5cm, with a coding capacity of 10,000 unique identifiers, ensuring high readability even under low light or partial obstruction conditions. The affixing position is precisely marked according to the vehicle design drawings, with the lateral offset from the track centerline controlled within ±2cm to match the lateral tolerance zone of the UAV's preset flight trajectory. The contrast between the barcode label and the background is greater than 80%, enhancing edge sharpness through black-and-white contrast and improving image segmentation and decoding success rates. The barcode labels are deployed according to a grid principle, with a correction point set every 3m to 8m along the longitudinal direction of the vehicle body, totaling no fewer than 20 key points throughout the train, covering vulnerable or critical functional component areas such as the bottom of the bogie, door rails, pantograph base, and air conditioning unit casing. Each barcode label is assigned a globally unique ID, which is bound to the vehicle model, train formation number, component type, and theoretical three-dimensional coordinates, and stored in the inspection database of the ground control center as a benchmark reference for subsequent position error calculations. The barcode labels are installed using industrial-grade adhesive backing or magnetic bases, ensuring long-term stability under vibration, temperature and humidity changes. The surface is coated with an anti-UV and anti-fouling coating to extend service life.

[0021] Step 2: UAV initialization and sensor calibration; The UAV first performs a self-test, then loads the preset inspection path parameters, and calibrates the internal and external parameters of the depth camera and the scanning angle and resolution of the LiDAR; Specifically, the drone adopts a quadcopter configuration, providing redundant power output and attitude stability. It has a maximum payload capacity of 2kg and a flight time of 30 minutes, meeting the energy requirements for a single complete train inspection. The self-test process includes: power supply voltage detection, motor health status diagnosis, communication link connectivity testing, and sensor data synchronization clock verification.

[0022] The UAV is equipped with a LiDAR, depth camera, motors, a stabilization module, and an onboard computer. The onboard computer has 8GB of RAM and 256GB of storage, runs a real-time operating system, and supports multi-threaded parallel processing of sensor data streams. The depth camera has a resolution of 1920×1080 pixels, a frame rate of 30 frames per second, and simultaneously outputs color images and depth maps, with a depth measurement range of 0.5m to 5m. The LiDAR has a scanning frequency of 10Hz, a horizontal field of view of 360 degrees, a vertical field of view of 30 degrees, a ranging accuracy of ±3cm, and an effective detection range of 20m. Sensor calibration is performed in a dedicated calibration field. The depth camera obtains its intrinsic parameter matrix and distortion coefficients using the Zhang Zhengyou checkerboard method, and its extrinsic parameters are aligned with the UAV's coordinate system using a rigid body transformation matrix. The LiDAR uses plane fitting and ICP registration algorithms to unify the UAV's coordinates with the depth camera's coordinate system, forming a fused perception coordinate system. Calibration parameters are stored in the onboard computer's configuration file and are automatically loaded upon each startup, ensuring spatial consistency of multi-source data.

[0023] Step 3: Autonomous flight and point approach; The UAV receives the step flight command from the ground control station and flies along the preset path to the recognition range of the target correction point based on the initial pose information provided by the inertial measurement unit and visual odometry. Specifically, the stepping flight command includes the next target point ID, theoretical coordinates, desired hovering altitude, and flight speed; the UAV inertial measurement unit outputs triaxial acceleration and angular velocity at a frequency of 200Hz, which are then filtered by complementary filtering and zero-velocity correction to generate short-term stable pose increments; the visual odometry processes continuous frame depth images at a frequency of 10Hz, extracts ORB feature points and performs optical flow tracking, solves relative motion using the PnP (Perspective-n-Point, PnP) algorithm to compensate for the drift of the inertial measurement unit, and generates initial pose information based on the pose increments.

[0024] The UAV calculates navigation vectors based on initial pose information and target correction point coordinates, generates quaternion attitude commands, and adjusts the rotational speeds of each rotor to achieve smooth trajectory tracking. The UAV's flight speed between points is set between 0.5 m / s and 2 m / s, dynamically adjusted according to environmental complexity: high-speed mode is used in open areas, and the speed is reduced to below 0.8 m / s in narrow passages or near obstacles. Simultaneously, the UAV continuously runs a real-time obstacle avoidance algorithm based on LiDAR, with a scanning radius of 5m and a detection latency of less than 100ms for dynamic obstacles. It can automatically generate a detour path after detecting an obstacle, with a minimum safe distance set at 0.5m. The real-time obstacle avoidance algorithm uses a data structure combining a grid map and an octree, updating obstacle occupancy probabilities in real time. Path planning employs an improved A / B algorithm. The algorithm minimizes the detour distance while ensuring safety. When the drone enters within 3 meters of the target correction point, it triggers barcode recognition preparation, switches the depth camera to high-gain mode, and focuses the LiDAR on a 2-meter cone-shaped area in front, providing high-quality input data for subsequent operations.

[0025] Step 4: Barcode label recognition and position error calculation; Barcode labels are acquired in real time using a depth camera. The target tracking algorithm based on the kernel correlation filter principle constructs a cyclic matrix in the frequency domain and solves for the center position of the barcode label. The position is then compared with the preset theoretical position to generate a position deviation vector. Specifically, after the barcode label is acquired, it first undergoes adaptive histogram equalization and Gaussian filtering to suppress uneven illumination and noise. Next, a sliding window and Haar-like feature classifier are used to coarsely locate the barcode label, and morphological closing operations are applied to the coarsely located region to connect broken edges caused by illumination, shadows, or partial occlusion, accurately outlining the barcode label boundary. Then, a target tracking algorithm based on kernel correlation filtering is used to construct a cyclic matrix in the frequency domain, and the convolution operation is converted into point-by-point multiplication using discrete Fourier transform to obtain the center position of the barcode label. Finally, the decoding module binarizes, denoises, and measures the bar width of the center position of the barcode label to parse out a unique ID code, which is then compared and verified with a pre-stored database.

[0026] Once the ID is successfully matched, the system calculates the three-dimensional coordinates of the barcode's center point in the camera coordinate system by back-projecting the pinhole imaging model. And compare the coordinates with the preset theoretical position. Compare and generate a positional deviation vector. ,in The horizontal deviation is within ±1cm, and the positioning accuracy is ±1cm. The horizontal and vertical deviations are accounted for, with a positioning accuracy of ±1cm. The positioning accuracy is ±2cm, with a vertical height deviation. The entire recognition and calculation process is executed in a pipeline manner on the onboard computer, with end-to-end latency controlled within 50ms to ensure the real-time performance of the control loop.

[0027] Step 5: Servo control and precise positioning; input the position deviation vector into the proportional-integral-derivative servo control algorithm to adjust the pitch angle, roll angle, yaw angle and vertical altitude of the UAV; Specifically, the position deviation vector is decomposed into four control channels: Mapped to roll angle control Mapped to pitch control Mapped to vertical thrust control, while the rotation angle deviation of the barcode label in the image is mapped to yaw angle control; each channel independently runs a proportional-integral-derivative servo control algorithm, the mathematical expression of which is: ； In the formula, For the first Output of each control cycle For the current deviation, The sampling period This is the proportionality coefficient. For the integral coefficient, These are the differential coefficients, and the control period is 50ms. This is a proportional term used to correct for deviations; This is the integral term, used to eliminate steady-state error; This is the differential term, used to suppress overshoot and oscillation. For example, when the drone's position deviates 0.1m to the right, The proportional term immediately generates a leftward control command with a force proportional to 0.1m, used to correct a 0.1m deviation to the right. The integral term checks and accumulates the previous deviations, while adding an additional leftward control force to ensure a complete return to the center point, rather than stopping very close to the right of the center point. The derivative term calculates the speed at which the drone moves to the left. If the speed is too fast, the derivative term generates a rightward "resistance" (equivalent to braking), allowing the drone to stop smoothly at the target point, thus enabling the drone to fly smoothly and accurately, eventually hovering on the theoretical flight path.

[0028] The system continuously monitors the normalized position (center pixel coordinates divided by image width and height) and scale factor (actual size to image size ratio) of the barcode label in the image. When the normalized position deviation is less than 0.05 and the scale factor fluctuation is less than 5%, it is determined that precise positioning is complete, triggering a high-definition image acquisition command. At this time, the depth camera switches to the highest resolution mode, and the LiDAR performs local high-density scanning to acquire multimodal data of key components, thereby ensuring that the shooting angle and focal length meet the geometric requirements for defect detection.

[0029] Step 6: Continuous inspection and path correction; After completing the multimodal data collection of the current correction point, the drone automatically flies to the next inspection point and repeats steps 3 to 5 until the inspection of all key parts of the train is completed. Specifically, the preset inspection path includes at least 20 key correction points, with an interval of 3m to 8m between adjacent points, dynamically adjusted according to the distribution density of vehicle components: denser points in the bogie area (3m interval), and sparser points in the central area of ​​the car body (8m interval). The drone's flight speed between points is 0.5m / s to 2m / s, and a single complete inspection takes no more than 25 minutes. The path correction mechanism is reflected on two levels: firstly, the precise positioning of each inspection point eliminates the accumulated error of the previous flight, achieving step-by-step calibration; secondly, the ground control station receives the actual positioning coordinates of each point in real time, performs a global comparison with the theoretical path, and if a systematic offset is found (such as an offset in the entire train's coordinate system), path offset compensation parameters can be dynamically issued to update the theoretical coordinates of subsequent points. The inspection task management software visually displays the drone's real-time position, flight trajectory, sensor data, and barcode recognition status, and supports remote updates of inspection path parameters and emergency intervention control. For example, if a barcode label at a certain location is damaged and cannot be identified, the drone can skip that inspection point or be instructed to perform a small-scale spiral search to attempt to re-capture it. All inspection events (such as successful positioning, obstacle avoidance triggering, and communication interruption) are recorded with timestamps and contextual data for post-event analysis and system optimization.

[0030] Step 7: Data transmission and return to the charging station. During the inspection, the collected multimodal data is transmitted to the ground control station in real time. After the inspection is completed, the UAV autonomously returns to the charging station and performs data integrity verification.

[0031] Specifically, the communication module uses dual-band transmission of 2.4GHz and 5.8GHz, with a maximum transmission distance of 500m. It supports transmission control protocols and user datagram protocols, and the automatic reconnection time after communication interruption is less than 3 seconds. Multimodal data uses H.265 encoding compression, supplemented by forward error correction codes, to ensure transmission reliability in complex electromagnetic environments. After the inspection, the UAV receives a return command and, based on the last known precise positioning point and a local map constructed by LiDAR SLAM, plans the shortest safe path back to the charging station. The charging station is located at a fixed position in the vehicle depot, with high-contrast visual markers on top for the UAV to dock. The station has an automatic docking function, a magnetic charging interface, a single full charge time of 40 minutes, and integrates environmental monitoring sensors to detect the temperature and humidity inside the station. During the UAV's descent, it first uses LiDAR for coarse positioning to a position 1m above the station, then switches to a depth camera to identify visual markers and executes servo control similar to step 5 to achieve centimeter-level precise landing. After integration, the system automatically performs data integrity verification, comparing the hash values ​​of locally stored data with those of uploaded data. If any data is missing, it is retransmitted via a wired interface. Furthermore, it includes an analysis module based on historical inspection data. This module uses machine learning algorithms to predict defects in the multimodal data of collected key components and generates inspection reports and maintenance recommendations. Deployed at the ground control station, this analysis module employs a convolutional neural network model. The input is multimodal data of key components, and the output is the probability distribution of defects such as cracks, corrosion, and loosening, assisting maintenance personnel in decision-making.

[0032] Example 2 Please see Figure 2 This embodiment provides a UAV positioning system for urban rail transit vehicle inspection, which applies a UAV positioning method for urban rail transit vehicle inspection as described above. The system includes an identification deployment module, a flight platform module, a navigation control module, a visual recognition module, a servo positioning module, and a ground control module. The identification deployment module is used to deploy barcode labels with unique codes on the surfaces of key components such as the vehicle bogie, doors, and pantograph, forming a network of correction points covering the entire outer surface of the vehicle. The flight platform module is used to carry sensor equipment and perform autonomous flight missions, including performing system self-checks, loading inspection paths, and completing multi-sensor calibration. The navigation control module is used to generate flight commands based on preset paths and sensor fusion data to achieve point-to-point autonomous navigation and real-time obstacle avoidance. The visual recognition module is used to collect barcode labels, decode and locate them, and calculate the position deviation vector between the actual position and the theoretical coordinates. The servo positioning module is used to receive the position deviation vector and adjust the attitude of the UAV through a multi-channel PID control algorithm to achieve centimeter-level precise hovering at the inspection point. The ground control module includes an analysis module for mission planning, data reception and storage, path correction decisions, and controlling the UAV's return and data verification.

[0033] All content not described in detail in this specification is prior art known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited; conventional equipment can be used. Electrical control components not mentioned in this technical solution are not shown in the figures because they are prior art, and will not be described further here.

[0034] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A positioning method for unmanned aerial vehicle (UAV) inspection of urban rail transit vehicles, characterized in that: The application relates to a multi-modal data acquisition and inspection method for urban rail transit vehicles, and belongs to the technical field of urban rail transit inspection. Step 1: Bar code labels with unique coding features are arranged at key positions of bogies, doors and pantographs of urban rail transit vehicles to form a network of correction points covering the outer surface of the whole vehicle; Step 2: The unmanned aerial vehicle (UAV) is self-checked, preset path parameters are loaded, and the internal and external parameters of a depth camera and the scanning angle and resolution of a laser radar are calibrated; Step 3: The UAV receives step-by-step flight instructions from a ground control station, flies to a target correction point along a preset path based on initial pose information provided by an inertial measurement unit and a visual odometry, and compares the initial pose information with the target correction point to generate a position deviation vector; Step 4: A target tracking algorithm based on a kernel correlation filtering principle constructs a cyclic matrix in a frequency domain, solves the center position of the bar code label, and compares the center position with a preset theoretical position to generate the position deviation vector; Step 5: The position deviation vector is input into a proportional-integral-derivative servo control algorithm to adjust the pitch angle, roll angle, yaw angle and vertical height of the UAV; Step 6: After multi-modal data of a current correction point are acquired, the UAV automatically flies to a next inspection point, and steps 3-5 are repeatedly executed until the inspection of all key positions of the whole vehicle is completed; Step 7: The acquired multi-modal data are transmitted to the ground control station during the inspection, and the UAV returns to a charging nest and performs data integrity checking after the inspection is completed.

2. The unmanned aerial vehicle positioning method for urban rail transit vehicle inspection according to claim 1, characterized in that: The step 1 comprises the following steps: The bar code labels are made of high-reflectivity materials to ensure high readability under low illumination or partial shielding conditions; The bar code labels are arranged according to a gridding principle, one correction point is arranged every 3-8 m along the longitudinal direction of the vehicle body, and no less than 20 key points are arranged on the whole vehicle to cover the bottom of a bogie, a door slide rail, a pantograph base, an air conditioner housing and other areas of vulnerable or key functional components; Each bar code label is assigned a globally unique ID, the ID is bound with a vehicle model, a marshalling sequence number, a component type and a theoretical three-dimensional coordinate, and is stored in an inspection database of the ground control station; The bar code labels are coated with anti-ultraviolet and anti-fouling coatings.

3. The unmanned aerial vehicle positioning method for urban rail transit vehicle inspection according to claim 2, characterized in that: The step 2 comprises the following steps: The self-checking of the UAV comprises power voltage detection, motor health state diagnosis, communication link connectivity test and sensor data synchronous clock verification; The UAV is provided with a laser radar, a depth camera, a motor, a stabilization module and an onboard computer; the onboard computer supports multi-thread parallel processing of sensor data streams; the depth camera synchronously outputs color images and depth maps; the laser radar has a horizontal field of view of 360 degrees, a vertical field of view of 30 degrees and an effective detection distance of 20 m; Sensor calibration is completed in a special calibration field; the depth camera obtains an internal parameter matrix and distortion coefficients through a Zhang Zhengyou chessboard method, and an external parameter is aligned with a UAV body coordinate system through a rigid body transformation matrix; the laser radar aligns the coordinates of the UAV with the coordinates of the depth camera through a plane fitting and an ICP registration algorithm to form a fusion perception coordinate system.

4. The unmanned aerial vehicle positioning method for inspection of urban rail transit vehicles according to claim 3, characterized in that: The step 3 comprises the following steps: The step-by-step flight instruction comprises a next target point ID, a theoretical coordinate, an expected hovering height and a flight speed; a three-axis acceleration and angular velocity of an inertial measurement unit of the unmanned aerial vehicle are output, a short-term stable pose increment is generated after complementary filtering and zero speed correction; a visual odometer processes continuous frame depth images, extracts ORB feature points and performs optical flow tracking, a relative motion is solved through a PnP (Perspective-n-Point, PnP for short) algorithm, drift of the inertial measurement unit is compensated, and initial pose information is generated based on the pose increment; The unmanned aerial vehicle calculates a navigation vector according to initial pose information and target correction point coordinates, generates a four-element attitude instruction, adjusts the rotation speed of each rotor, and realizes smooth trajectory tracking; the flight speed of the unmanned aerial vehicle between point positions is set to 0.5 m / s to 2 m / s, which is dynamically adjusted according to the complexity of the environment: high-speed mode is adopted in open areas, and the speed is reduced to below 0.8 m / s in narrow channels or near obstacle areas; at the same time, the unmanned aerial vehicle continuously runs a real-time obstacle avoidance algorithm based on a laser radar, and automatically generates a fly-around path after detecting an obstacle; the real-time obstacle avoidance algorithm uses a data structure combining a grid map and an octree, updates the obstacle occupancy probability in real time, and uses an improved A * algorithm for path planning, which minimizes the detour distance under the premise of safety; when the unmanned aerial vehicle enters a range of 3 m from the target correction point, a barcode label recognition preparation state is triggered, a depth camera is switched to a high-gain mode, and a laser radar is focused on a 2 m cone-shaped area in front.

5. The unmanned aerial vehicle positioning method for inspection of urban rail transit vehicles according to claim 4, characterized in that: The step 4 comprises: After the barcode label is collected, adaptive histogram equalization and Gaussian filtering are firstly performed to suppress uneven illumination and noise; then a sliding window and a Haar-like feature classifier are used to coarsely position the barcode label, and a morphological closing operation is used to connect the broken edges caused by light, shadow or partial occlusion in the coarsely positioned region to accurately outline the boundary of the barcode label; then a target tracking algorithm based on kernel correlation filtering constructs a circulant matrix in the frequency domain, uses discrete Fourier transform to convert convolution operation into point-by-point multiplication to obtain the center position of the barcode label; finally, a decoding module is used to binarize, denoise and measure the width of the center position of the barcode label, parse the unique ID code, and compare and verify it with a pre-stored database; after verification, the three-dimensional coordinates of the barcode center point in the camera coordinate system are calculated through back projection of the pinhole imaging model, and compared with the preset theoretical position to generate a position deviation vector.

6. The unmanned aerial vehicle positioning method for urban rail transit vehicle inspection according to claim 5, characterized in that: The step 5 comprises: The position deviation vector is decomposed into four control channels, including roll angle control, pitch angle control, vertical thrust control and yaw angle control; a proportional-integral-derivative servo control algorithm is independently run in each channel, and the mathematical expression is: ； In the formula, For the first Output of each control cycle For the current deviation, The sampling period is This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients, and the control period is 50ms. This is a proportional term used to correct for deviations; This is the integral term, used to eliminate steady-state error; This is a differential term used to suppress overshoot and oscillation; The normalized position and scale factor of the barcode label in the image are continuously monitored, and when the normalized position deviation is less than 0.05 and the scale factor fluctuation is less than 5%, it is determined that accurate positioning is completed, and a high-definition image acquisition instruction is triggered; at this time, the depth camera switches to the highest resolution mode, and the laser radar performs local high-density scanning to obtain multi-modal data of the key components, ensuring that the shooting angle and focal length meet the geometric requirements of defect detection.

7. The unmanned aerial vehicle positioning method for urban rail transit vehicle inspection according to claim 6, characterized in that: The step 6 comprises: The ground control station receives the actual positioning coordinates of each point in real time, and compares them with the theoretical path globally; if systematic deviation is found, path offset compensation parameters can be dynamically issued to update the theoretical coordinates of subsequent points; The real-time position, flight trajectory, sensor data and barcode recognition state of the unmanned aerial vehicle are visually displayed, and remote update of the inspection path parameters and emergency intervention control are supported.

8. The unmanned aerial vehicle positioning method for urban rail transit vehicle inspection according to claim 7, characterized in that: The step 7 comprises: The multi-modal data is encoded and compressed using H.265, supplemented by forward error correction codes to ensure transmission reliability in complex electromagnetic environments; After the inspection is completed, the unmanned aerial vehicle receives a return instruction, plans a shortest and safest path to return to the charging nest based on the last known inspection positioning point and the local map constructed by the laser radar SLAM. The charger nest top is provided with high-contrast visual marks for final docking of the unmanned aerial vehicle; the charger nest has automatic docking function, the charging interface is designed in a magnetic suction type, and is integrated with an environmental monitoring sensor.

9. The unmanned aerial vehicle positioning method for urban rail transit vehicle inspection according to claim 1, characterized in that: Further comprising an analysis module based on historical inspection data; the analysis module performs defect prediction on the collected multi-modal data of the key components through a machine learning algorithm, and generates an inspection report and maintenance recommendations; the analysis module is deployed in a ground control station, uses a convolutional neural network model, the input is multi-modal data of key components, and the output is probability distribution of crack, corrosion and loosening defects, used to assist operation and maintenance personnel in decision-making.

10. A UAV positioning system for urban rail transit vehicle inspection, characterized in that: The application applies a positioning method for unmanned aerial vehicle for urban rail transit vehicle inspection according to any one of claims 1-9, and the system comprises an identification layout module, a flight platform module, a navigation control module, a visual recognition module, a servo positioning module and a ground control module. The identification layout module is used for laying bar code labels with unique codes on the surfaces of key components of vehicle bogies, doors and pantographs to form a modified point network covering the whole vehicle surface. The flight platform module is used for carrying sensor equipment and performing autonomous flight tasks, including system self-checking, loading inspection path and completing multi-sensor calibration. The navigation control module is used for generating flight instructions based on the preset path and sensor fusion data to realize point-to-point autonomous navigation and real-time obstacle avoidance. The visual recognition module is used for collecting bar code labels and decoding and positioning, and calculating the position deviation vector of the actual position and the theoretical coordinates. The servo positioning module is used for receiving the position deviation vector and adjusting the unmanned aerial vehicle attitude through a multi-channel PID control algorithm to realize centimeter-level accurate hovering of the inspection point. The ground control module comprises an analysis module for task planning, data receiving and storage, path correction decision-making, and controls the unmanned aerial vehicle to return and data verification.

Images