Multi-source perception motion compensation weeding robot and working method

Abstract

The application discloses a multi-source perception motion compensation laser weeding robot and a working method, and belongs to the technical field of intelligent agricultural laser weeding. Multi-source data is collected, weed meristems are identified and corresponding three-dimensional point clouds are segmented, local surface normal vectors of weeds are fitted, and a six-dimensional aiming state vector with position and orientation is constructed. A rigid body pose matrix from exposure to light emission interval is solved according to inertial data, a predicted aiming vector is obtained through a state vector time sequence coordinate transformation by a pose position coupling operator, a galvanometer deflection angle and laser pulse energy are simultaneously corrected according to the orientation of the target surface, and laser is vertically applied to weed meristems. After laser irradiation, algorithm parameters are adaptively updated based on a double-parameter model of burning effect. The matched robot is provided with a coplanar multi-sensing gantry support, a laser module is flexibly suspended through a steel wire rope vibration isolator to realize vibration reduction and decoupling, and the above method is jointly executed. The application has high dynamic weeding precision, strong long-term operation stability, and is suitable for automatic removal of various types of farmland weeds.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural laser weeding technology, specifically to a multi-source sensing motion compensation laser weeding robot and its operation method. Background Technology

[0002] Farmland weeds continuously compete with crops for water, fertilizer, and sunlight, significantly reducing the yield of grain and cash crops. Traditional weeding methods have obvious shortcomings: long-term use of chemical herbicides can easily lead to herbicide resistance in weeds, soil and water pollution, and excessive pesticide residues in agricultural products; mechanical weeding relies on blades for tilling, which can easily damage crop roots and disrupt the soil structure; manual weeding is labor-intensive and inefficient, making it unsuitable for large-scale, contiguous planting. Laser weeding, which uses the thermal effect of high-energy lasers to burn the meristematic tissue of weeds, has the technological advantages of zero pollution, precise targeting, and no damage to surrounding crops, and has become a core research direction in green and intelligent agriculture. Currently, laser weeding robots are generally equipped with visual perception units, inertial measurement units, lidar, and two-dimensional galvanometer scanning modules. Through multiple sensors, they acquire spatial information about weeds in the field and the motion status of the chassis. Combined with motion compensation algorithms, they eliminate the deviation in laser point of impact caused by vehicle bumps and displacement, achieving precise weeding under dynamic operation on a mobile platform. However, complex field conditions bring multiple coupling errors: there is a millisecond-level time difference between the camera exposure time and the laser emission time, vehicle pitch and roll disturbances, different curved surface shapes of weed plants causing laser incident angle shift, mismatch of sampling frequencies of multiple sensors, and mechanical vibration causing calibration parameter drift. After multiple errors are superimposed, problems such as spot shift, ablation failure, and accidental scorching of seedlings will occur, which restrict the field landing accuracy and operational stability of laser weeding equipment.

[0003] Currently, domestic universities and research institutes have conducted a large number of experimental studies on laser weeding perception and motion compensation. Representative results have been published in core agricultural equipment journals such as the "Transactions of the Chinese Society for Agricultural Machinery" and the "Transactions of the Chinese Society for Agricultural Engineering". Among them, the "Design and Experiment of Field Laser Weeding Robot Target Control System" published by Zhong Huiyu et al. systematically built a vision-galvanometer integrated target control system. Based on the two-dimensional pixel coordinates and depth information of weeds obtained by the depth camera, the geometric inverse attitude compensation model was constructed by combining the IMU to solve the chassis attitude. The camera and galvanometer coordinate system were unified through offline hand-eye calibration, and centimeter-level aiming accuracy was achieved under low-speed conditions on flat ground, providing a basic control framework for dynamic laser weeding. This study only uses a single-vision depth perception scheme and does not introduce three-dimensional lidar to collect complete point clouds of weeds. It cannot obtain the local curved surface morphology and normal characteristics of weed meristems. The laser beam is always output at a fixed vertical reference angle. When weeds are tilted, fallen, or growing close to the ground, the actual incident angle of the laser deviates significantly from the vertical state. Under the same laser power, the effective energy density received by the weed meristems is significantly reduced, which easily leads to incomplete burning and secondary germination of weeds. At the same time, its attitude compensation model only performs static geometric correction for the displacement error of the chassis rigid body and does not distinguish the time interval Δt between the camera exposure time and the laser emission time. There is no joint motion trajectory inference mechanism based on time-series tracing. It only relies on static coordinate mapping of a single frame image to complete the galvanometer control. The dynamic lag error caused by the misalignment of imaging and emission time when the vehicle is moving and bumping cannot be effectively offset. Concurrently, Liang Chao et al. optimized the weed visual recognition network in their work "Laser Weed Detection Method Based on LSPKI-YOLO Multi-Scale Feature Enhancement," improving the accuracy of key weed location under complex lighting conditions. However, their research focused solely on image recognition algorithms, completely neglecting core control logic such as laser energy and incident angle coupling control, multi-source perception six-dimensional target state vector construction, and adaptive parameter iteration correction for post-ablation effects. These academic studies generally simplify weeds to single-point spatial targets, ignoring the decisive influence of weed meristematic surface orientation on laser ablation effects. Motion compensation only performs basic corrections for chassis rigid body displacement, lacking a spatiotemporal fusion link for multi-source data from vision, IMU, and lidar. Furthermore, they failed to establish a dynamic linkage adjustment mechanism for laser galvanometer deflection angle and pulse energy based on the target surface normal, resulting in significant attenuation of weeding hit rate and eradication effect in high-speed, undulating terrain scenarios.

[0004] Domestic patents are also being developed for dynamic compensation technology in laser weeding. The patent CN117620488B, titled "A Modular Multi-Axis Laser Galvanometer Motion Controller," proposes a hardware architecture for synchronous driving of multiple galvanometers. Based on encoder feedback of the real-time position of the galvanometers, a chassis speed feedforward compensation model is established. The linear coordinate offset is used to offset the lag deviation of the light spot caused by the uniform movement of the platform. This enables synchronous scanning of multiple galvanometers to expand the working area and improves the dynamic aiming accuracy of weeds between wide rows to a certain extent. The patent focuses solely on the galvanometer hardware driver and one-dimensional lateral displacement compensation, exhibiting several core technological deficiencies: First, the sensing end only accesses a single-channel visual camera to collect two-dimensional pixel information, failing to integrate the three-dimensional point cloud data from the lidar and the high-frequency inertial data from the IMU. This prevents the calculation of the three-dimensional spatial coordinates of weeds and the local surface normal vectors of meristematic tissues, effectively treating weeds as planar pixels and failing to distinguish the target surface orientation. Furthermore, there is no adaptive adjustment logic for the laser output power and galvanometer deflection angle. Second, the motion compensation model uses only one-dimensional linear offset compensation, merely correcting the platform's forward and backward displacement errors. It does not construct a rigid body pose transformation matrix that includes rotation and translation coupling, thus failing to compensate for pitch and roll angle disturbances caused by terrain undulations. This also hinders the calculation of laser beams on slopes and uneven fields. The vertical offset of the spot is difficult to eliminate; thirdly, the perception and compensation timing are completely decoupled, and there is no algorithm that uses the camera exposure timestamp as a benchmark to reverse-engineer the platform-galvanometer joint motion trajectory from exposure to light emission throughout the entire time period. The low-frequency visual sampling and the high-frequency feedback data from the IMU and galvanometer cannot be time-aligned, and the dynamic error caused by imaging delay can only be corrected linearly afterward, and the compensation lag problem cannot be eradicated; fourthly, the entire control scheme lacks a closed-loop self-correction mechanism for operation. After the laser completes the burning, it does not iteratively update the point cloud fitting and energy compensation parameters based on the feedback of the effectiveness of weed burning. During long-term operation, the system error caused by the micro-deformation of the sensor bracket, lens dust accumulation, and galvanometer zero-point drift continues to accumulate, and the operation accuracy decreases continuously with the travel distance. In addition, the patent does not design a mechanical decoupling and shock absorption structure. The laser emission module is rigidly connected to the sensor bracket, and the chassis vibration is directly transmitted to the galvanometer and optical lens, further amplifying the perception and aiming error, making it difficult to adapt to the all-weather operation needs of the fragmented and undulating farmland in China.

[0005] Existing academic literature and patent solutions generally suffer from three major shortcomings: First, the sensing system is either a single sensor or simply splices data from multiple sensors, failing to simultaneously acquire the three-dimensional coordinates of weeds and the normal features of the meristematic surface, and lacking a six-dimensional aiming state vector to fully characterize the target's spatial and morphological information; second, motion compensation only performs linear correction for the one-dimensional displacement of the chassis, without constructing a platform-galvanometer coupling transformation model for exposure-light emission timing, thus failing to simultaneously offset the multiple coupling errors of vehicle rotation, translation, and timing delay; third, the laser output parameters are fixed or simply adjusted based on the size of the weeds, without dynamically matching the pulse energy to the incident angle of the target surface, and there is no adaptive correction link for ablation effect feedback, leading to continuous accumulation of errors over long-term operation. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a multi-source sensing motion-compensated laser weeding robot and its operating method; the multi-source sensing motion-compensated laser weeding operating method includes: During the movement of the mobile platform, visual images, inertial measurement data, and laser ranging point clouds are collected simultaneously. The target weed meristem is identified from the visual image and its pixel coordinates are extracted. At the same time, a three-dimensional point cluster corresponding to the target weed is segmented from the laser ranging point cloud. The local surface normal vector of the target weed meristem is obtained by fitting the three-dimensional point cluster. The rigid body pose transformation matrix of the mobile platform between the camera exposure time and the laser emission time is calculated using the inertial measurement data; The pixel coordinates and the local surface normal vector are jointly projected onto the laser emission coordinate system to construct a six-dimensional aiming state vector that includes the target position and the target surface orientation; The six-dimensional aiming state vector is transformed by the rigid body pose transformation matrix to obtain the predicted aiming state vector at the moment of laser emission. Based on the surface orientation component in the predicted aiming state vector, the deflection angle of the laser galvanometer and the pulse energy of the laser are dynamically adjusted so that the laser beam acts on the target weed meristem with an incident angle perpendicular to the surface of the target weed meristem and with an effective energy density matching the tissue ablation threshold at that incident angle.

[0007] Preferably, the local surface normal vector of the target weed meristem is obtained based on the fitting of the three-dimensional point cluster, including: The three-dimensional point cluster is subjected to voxel downsampling and statistical outlier filtering to obtain a cleaned point set; Using the three-dimensional projection point corresponding to the pixel coordinates as the query center, search for k nearest neighbors in the cleaned point set to form a local neighborhood; Principal component analysis is performed on the points in the local neighborhood, and the unit eigenvector corresponding to the minimum eigenvalue is taken as the initial normal vector. The initial normal vector is oriented to be checked for consistency using the gradient direction of the target weed meristematic region in the visual image. When the angle between the initial normal vector and the gradient direction is greater than 90 degrees, it is reversed, and the checked local surface normal vector is output.

[0008] Preferably, a six-dimensional aiming state vector is constructed, including the target position and the target surface orientation, comprising: The pixel coordinates are back-projected onto the camera coordinate system using the camera intrinsic parameter matrix and depth values ​​to obtain the three-dimensional position components of the target. Transform the verified local surface normal vector to the same camera coordinate system to obtain the target surface orientation component; The three-dimensional position component of the target and the orientation component of the target surface are concatenated into a six-dimensional column vector, and marked with a timestamp of the camera exposure time to form a time-stamped six-dimensional aiming state vector.

[0009] Preferably, the six-dimensional aiming state vector is transformed using the rigid body pose transformation matrix to obtain the predicted aiming state vector at the moment of laser emission. This is specifically achieved through an attitude position coupling transfer model, with the following formula: in, Let be the predicted aiming state vector at the moment of laser emission. This refers to the moment when the laser light is emitted. Let be the six-dimensional aiming state vector at the moment of camera exposure. For the camera exposure time, and These are the data obtained from inertial measurement at time intervals. The rotation submatrix and translation subvector obtained by inner integration This represents a custom attitude-position coupling operator, with a rotation submatrix. Simultaneously acting on Position and orientation components, translation subvectors Only the position component is superimposed, and the orientation component is normalized after transformation.

[0010] Preferably, the deflection angle of the laser galvanometer is dynamically adjusted based on the surface orientation component in the predicted aiming state vector, specifically through incident angle deviation driving, as shown in the formula: in, Let be the target deflection angle vectors of the two axes of the galvanometer. This is the basic deflection angle vector calculated solely based on the target position. To predict the surface orientation component in the aiming state vector, This is the unit vector pointing to the current output optical axis of the laser. This is the preset verticality gain coefficient. This represents the cross product operation of vectors; The incident angle deviation is converted into an angular correction amount that is consistent with the deviation direction, and then fed forward and superimposed onto the base deflection angle.

[0011] Preferably, dynamically adjusting the pulse energy of the laser includes: The cosine of the actual incident angle is calculated based on the surface orientation component in the predicted aiming state vector. Using the tissue ablation threshold energy density obtained under vertical incidence conditions as a reference, the reference energy density is divided by the cosine value to obtain the compensated energy density setting value; The compensated energy density setting value is converted into a laser drive current modulation signal, and the current setting value is updated and latched within one pulse cycle before the laser emission time.

[0012] Preferably, an adaptive correction for ablation effectiveness is performed after laser irradiation, specifically through a two-parameter attenuation feedback model, the formula of which is: in, and These represent the fitted window radius of the corrected normal vector and the energy density compensation coefficient, respectively. and The corresponding parameter values ​​before correction. This represents the average ablation effectiveness feedback value from n consecutive operations. This represents the highest ablation effectiveness feedback value among the most recent m operations. To preset the effective threshold, , , All are preset positive real number adjustment factors; when consistently below When, the fitting window radius Exponential contraction to improve the accuracy of local surface fitting, energy density compensation coefficient. The ablation margin is increased nonlinearly according to the hyperbolic tangent law.

[0013] A multi-source sensing motion-compensated laser weeding robot, used in the multi-source sensing motion-compensated laser weeding operation method described above, includes a mobile chassis, a sensor bracket, a laser emitting module, and a main control computing unit; The sensor bracket is a welded aluminum alloy frame, and its four bottom corners are rigidly fixed to the top support panel of the mobile chassis by bolts. An industrial camera, an inertial measurement unit, and a lidar are fixedly mounted on the top crossbeam of the sensor bracket from left to right with screws. The mounting surfaces of the three are coplanar and their optical axes or sensing axes are parallel to each other. The laser emitting module includes a fiber laser, a two-dimensional galvanometer scanner, a focusing lens group, and a module housing; the fiber laser is fixedly installed on the inner side wall of the module housing, and its output fiber passes through a sealed connector on the side wall of the housing and is coupled to the light inlet of the two-dimensional galvanometer scanner. The two-dimensional galvanometer scanner is fixed to the inner side of the top plate of the module housing by four countersunk screws, and the light outlet is vertically downward and coaxially connected to the focusing lens group; the focusing lens group is fixed to the bottom lens barrel of the module housing by screwing. A steel wire rope vibration isolator is installed at each of the four corners of the top plate of the module housing. The upper end of each steel wire rope vibration isolator is fixed to the bottom mounting plate of the sensor bracket by flange bolts, and the lower end is fixed to the outside of the top plate of the module housing by flange bolts, so that the laser emission module is suspended below the sensor bracket and forms a flexible connection structure with mechanical decoupling. The main control computing unit is fixedly installed in the internal electrical control box of the mobile chassis. It is electrically connected to the industrial camera and lidar via gigabit Ethernet cable, electrically connected to the inertial measurement unit via RS-485 bus, and electrically connected to the fiber laser and two-dimensional galvanometer scanner via analog control cable.

[0014] Preferably, a 45-degree oblique mounting hole is provided on the bottom lens barrel side wall of the module housing, a beam splitter mount is fixedly installed in the oblique mounting hole, a 45-degree beam splitter is embedded in the beam splitter mount, and the reflective surface of the beam splitter faces upward and forms a 45-degree angle with the light output axis of the focusing lens group. A coaxial monitoring camera is also fixedly installed on the outer side wall of the bottom lens barrel of the module housing by a C-type clamp. The optical axis of the coaxial monitoring camera passes horizontally through the beam splitter mount and is reflected and refracted 90 degrees by the 45-degree beam splitter, so as to be precisely coaxial with the output optical axis of the focusing lens group. The imaging sensor target surface of the coaxial monitoring camera and the focal plane of the focusing lens group satisfy the optical conjugate relationship. The coaxial monitoring camera is electrically connected to the main control computing unit via a CameraLink cable.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-source sensing motion-compensated laser weeding method.

[0016] Compared with the prior art, the technical solution of this application has the following technical effects: This invention collects three types of perception information: visual images, inertial measurement data, and laser ranging point clouds. Based on the three-dimensional point cloud processing flow, it extracts the local surface normal vector of the weed meristem and constructs a six-dimensional aiming state vector that can simultaneously represent the target position and surface orientation by combining it with pixel coordinates. This covers two key types of information: the spatial position of the weed and the surface morphology of the plant. It can completely restore the real spatial state of the weed meristem and provide comprehensive and accurate target basic data for motion compensation and laser parameter control, avoiding the one-sided aiming logic problem caused by the lack of single-dimensional perception information.

[0017] This invention obtains the rotation and translation transformation matrix within the interval from camera exposure to laser emission based on inertial data integration, completes the coordinate transformation of the six-dimensional aiming state vector using an attitude and position coupling transfer model, and derives the predicted aiming state vector corresponding to the laser emission moment. The temporal coupling transformation method uniformly handles platform translation disturbances and angle deflection disturbances, fully compensates for all motion deviations between the imaging moment and the laser action moment, realizes accurate prediction of the target state under the continuous movement of the mobile platform, and ensures that the laser aiming coordinates always match the real-time position of the weeds.

[0018] This invention synchronously adjusts the galvanometer deflection angle and laser pulse energy based on the surface orientation component within the predicted aiming state vector. It solves for the incident angle deviation through vector cross product operation and superimposes the angle correction amount, then calculates the appropriate laser energy density based on the incident angle. The laser beam can be perpendicularly incident on the surface of weed meristematic tissue, and the laser output energy matches the ablation requirements corresponding to the current incident angle. This can stably ensure that the weed tissue receives sufficient and effective thermal energy, improve the weed burning and removal effect, and reduce unnecessary energy consumption.

[0019] This invention introduces a dual-parameter attenuation feedback model for ablation effectiveness after a single laser irradiation operation. Based on the feedback results of multiple rounds of operations, the radius of the surface fitting window and the energy compensation coefficient are adjusted synchronously to continuously optimize the point cloud fitting accuracy and the degree of laser energy compensation. This gradually offsets the system cumulative error generated by long-term operations, and maintains stable accuracy in weed identification, surface fitting and laser aiming. It can maintain stable and reliable weeding operation capability under long-term field operations.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0023] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows: Figure 1 This is a schematic diagram of the overall execution process of the multi-source sensing motion-compensated laser weeding method of the present invention; Figure 2 This is a step-by-step visualization of the weed point cloud preprocessing and local surface normal vector processing of the present invention. Figure 3 This is a logic block diagram of the attitude and position coupling transfer model data flow in this invention; Figure 4 This is a block diagram of the collaborative control operation of the laser galvanometer and pulse energy in this invention; Figure 5 This is an isometric drawing of the assembly of the multi-source sensing motion compensation laser weeding robot of the present invention. Figure 6 This is a schematic cross-sectional view of the internal optical path sealing structure of the laser emission module of the present invention.

[0024] Figure label: 11. Mobile chassis; 12. Top load-bearing panel; 13. Electrical control box; 2. Sensor bracket; 21. Column; 22. Top crossbeam; 23. Bottom mounting plate; 3. Industrial camera; 4. Inertial measurement unit (IMU); 5. LiDAR; 6. Laser emitting module; 61. Module housing; 62. Fiber laser; 63. Two-dimensional galvanometer scanner; 64. Focusing lens group; 65. Bottom lens barrel; 66. Sealing joint; 7. Steel wire rope vibration isolator; 8. Main control computing unit; 101. 45° beam splitter; 102. Beam splitter mount; 103. Coaxial monitoring camera. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0026] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0027] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0028] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0029] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0030] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0031] Example 1 This embodiment mainly describes a multi-source sensing motion-compensated laser weeding operation method, such as... Figure 1 As shown, it includes: During the movement of the mobile platform, visual images, inertial measurement data, and laser ranging point clouds are collected simultaneously. The target weed meristem is identified from the visual image and its pixel coordinates are extracted. At the same time, a three-dimensional point cluster corresponding to the target weed is segmented from the laser ranging point cloud. The local A-curve normal vector of the target weed meristem is obtained by fitting the three-dimensional point cluster. The rigid body pose transformation matrix of the mobile platform between the camera exposure time and the laser emission time is calculated using the inertial measurement data; The pixel coordinates and the local surface normal vector are jointly projected onto the laser emission coordinate system to construct a six-dimensional aiming state vector that includes the target position and the target surface orientation; The six-dimensional aiming state vector is transformed by the rigid body pose transformation matrix to obtain the predicted aiming state vector at the moment of laser emission. Based on the surface orientation component in the predicted aiming state vector, the deflection angle of the laser galvanometer and the pulse energy of the laser are dynamically adjusted so that the laser beam acts on the target weed meristem with an incident angle perpendicular to the surface of the target weed meristem and with an effective energy density matching the tissue ablation threshold at that incident angle.

[0032] Furthermore, during the movement of the mobile platform, spatiotemporal synchronous acquisition of multi-sensor data is performed. Visual images, inertial measurement data, and laser ranging point clouds are simultaneously acquired during the platform's movement, achieving microsecond-level synchronization based on hardware triggering or high-precision timestamps (such as the PTP protocol). Let the camera exposure time be... The IMU sampling time sequence is The LiDAR scanning time is The system needs to ensure ,in This is the maximum permissible synchronization error threshold. Simultaneously, IMU data is continuously output at a frequency far exceeding the camera frame rate to fill in the gaps. Until the laser emission time Information on high-frequency attitude changes between them.

[0033] After data acquisition, the target identification and 3D feature extraction stage begins. Target weed meristems are identified from the visual image, and pixel coordinates are extracted. Simultaneously, 3D point clusters corresponding to the target weeds are segmented from the laser ranging point cloud. Based on these 3D point clusters, the local surface normal vector of the target weed meristem is obtained. Let the first target weed identified in the visual image be... The pixel coordinates of the weed germination center are: The corresponding depth value is obtained through laser point cloud projection, denoted as . Using the camera intrinsic parameter matrix It can back-project pixel coordinates to three-dimensional points in the camera coordinate system. : This 3D point serves as the query center for subsequent normal vector fitting. To obtain high-quality geometric features from the original point cloud, preprocessing is necessary. The 3D point cluster is subjected to voxel downsampling and statistical outlier filtering to obtain a cleaned point set. Voxel downsampling is performed for each voxel grid. Calculate the centroid of all points inside it. As a representative point: ,in voxels The number of points within the point cloud. Outlier filtering is based on the global mean of the point cloud. and standard deviation Eliminate those that meet the requirements The anomalies, among which The standard deviation multiple threshold is denoted as . The cleaned point set obtained after the above processing is denoted as . It effectively preserves the local geometric structure of the weed stem surface while significantly reducing noise interference.

[0034] The cleaned point set obtained after the above processing is denoted as... It effectively preserves the local geometric structure of the weed stem surface while significantly reducing noise interference.

[0035] like Figure 2 As shown in the figure, four groups of weed samples under different field backgrounds are arranged. Each group of samples, from left to right, sequentially displays the original field weed real-world image, the weed meristem selected by visual recognition and the center query point marked, the global color 3D point cloud of weeds collected by LiDAR, the red cluster of weed targets extracted from the global point cloud, the pure weed point set after voxel downsampling and outlier filtering, and the visualization result of the local surface normal vector of weeds generated by fitting the pure point set. Furthermore, after obtaining the cleaned point set, the spatial orientation of the target surface is estimated. Using the 3D projection point corresponding to the pixel coordinates as the query center, k nearest neighbors are searched in the cleaned point set to form a local neighborhood. Let the query point be... Its k-nearest neighbor set is ,satisfy ,in To adapt the search radius, principal component analysis is performed on the points within the local neighborhood, and the unit eigenvector corresponding to the smallest eigenvalue is taken as the initial normal vector. Specifically, the covariance matrix is ​​constructed. : right Eigenvalue decomposition yields three eigenvalues. and its corresponding unit eigenvector Since the stems of weeds are approximately cylindrical or conical, the minimum eigenvalue... Corresponding feature vector That is, the initial normal vector perpendicular to the tangent plane. To quantify the quality of the fit, a flatness index is defined. : when If the value is below the preset threshold, it indicates that the local area is too flat or too noisy, and the k value needs to be increased to search again.

[0036] The normal vector obtained by PCA has directional ambiguity (i.e.) and (All are correct solutions); therefore, the gradient direction of the target weed meristematic region in the visual image is used to verify the orientation consistency of the initial normal vector. When the angle between the initial normal vector and the gradient direction is greater than 90 degrees, it is reversed, and the verified local surface normal vector is output.

[0037] Let the average gradient vector of the target region in the image be... It is rotated through the camera extrinsic rotation matrix. Transform to three-dimensional space to obtain the reference direction vector The verification rules are as follows: This ensures that the normal vector always points to the laser incident side, providing the correct geometric reference for energy compensation.

[0038] To improve the stability of the normal vector in regions with drastic curvature changes, a weighted PCA strategy is introduced, with the weight function... Defined as the inverse ratio of distance to the query point: The weighted covariance matrix becomes This strengthens the contribution of the proximal endpoint to the normal estimation. The standard deviation of the Gaussian distribution is the weighting function; Furthermore, after completing the geometric feature extraction, the scattered position and attitude information needs to be unified into state variables that can be used for motion compensation; the pixel coordinates are back-projected to the camera coordinate system through the camera intrinsic parameter matrix and depth values ​​to obtain the target's three-dimensional position components; the verified local surface normal vector is transformed to the same camera coordinate system to obtain the target surface orientation components; the target's three-dimensional position components and the target surface orientation components are concatenated into a six-dimensional column vector, and marked with the timestamp of the camera exposure time to form a time-stamped six-dimensional aiming state vector, the formula of which is: superscript Indicates the first One target. The key here is the inclusion of a time stamp. Only Valid at all times. Due to the continuous movement of mobile platforms, if directly... For Moment-dependent laser control will produce significant spatial deviations; let the platform linear velocity be... angular velocity is Time delay The uncompensated position error is approximately on the order of magnitude of The angle error is approximately .

[0039] In typical work scenarios ( The positional error can reach 30mm, far exceeding the laser spot diameter, therefore strict motion compensation is necessary. To support the transformation, a static calibration relationship between the camera coordinate system and the laser emission coordinate system must also be established. Let the homogeneous transformation matrix between the two coordinate systems be... , including rotation Peaceful relocation : The matrix is ​​obtained through pre-calibration and is treated as a constant throughout the entire operation cycle. However, in the dynamic compensation stage, the core variable is the relative pose transformation obtained by IMU integration.

[0040] Furthermore, such as Figure 3 As shown, an attitude-position coupling transfer model is constructed. The core of motion compensation lies in accurately predicting the target's state at the execution moment. The six-dimensional aiming state vector is transformed using the rigid body pose transformation matrix to obtain the predicted aiming state vector at the laser emission moment. This is specifically achieved through the attitude-position coupling transfer model, with the following formula: in, Let be the predicted aiming state vector at the moment of laser emission. This refers to the moment when the laser light is emitted. Let be the six-dimensional aiming state vector at the moment of camera exposure. For the camera exposure time, and These are the data obtained from inertial measurement at time intervals. The rotation submatrix and translation subvector obtained by inner integration This represents a custom attitude-position coupling operator, with a rotation submatrix. Simultaneously acting on Position and orientation components, translation subvectors Only the position component is superimposed, and the orientation component is normalized after transformation; By measuring the angular velocity output from the IMU gyroscope Points are awarded based on accumulated points. exist If the interior is approximately constant, then: ,in is the antisymmetric matrix of angular velocity.

[0041] For greater accuracy, quaternion updates or Runge-Kutta integration can be used. Then based on accelerometer data By double integration and combining it with gravity compensation, we obtain: ,in It is the gravitational acceleration vector. Let be the attitude matrix at the intermediate time step; Custom Operators Although the position and normal vector belong to the same rigid body, their transformation rules are different. The specific calculation rules are as follows: Special emphasis is placed on normalization here because numerical integration and floating-point operations may cause the normal vector's magnitude to deviate from 1, while trigonometric function calculations require a unit vector as input. The normalization operation is defined as follows: ,in To prevent small values ​​from being divided by zero, the model fully considers the rigid body motion characteristics of the mobile platform in a short period of time, avoiding the shortcomings of traditional methods that only compensate for position while ignoring attitude changes. It is especially suitable for operation scenarios on uneven ground where the platform experiences pitch and roll vibrations.

[0042] Furthermore, such as Figure 4 As shown, after obtaining the predicted state, it needs to be converted into control commands for the galvanometer. The deflection angle of the laser galvanometer is dynamically adjusted according to the surface orientation component in the predicted aiming state vector. Specifically, this is achieved through the incident angle deviation driving equation, the formula of which is: in, Let be the target deflection angle vectors of the two axes of the galvanometer. This is the basic deflection angle vector calculated solely based on the target position. To predict the surface orientation component in the aiming state vector, This is the unit vector pointing to the current output optical axis of the laser. This is the preset verticality gain coefficient. This represents the vector cross product operation; it converts the incident angle deviation into an angular correction amount in the same direction as the deviation, and then adds it to the base deflection angle in a feedforward manner. The term calculates the current beam direction. With ideal normal direction Angle deviation between That is, the complementary angle of the actual incident angle; cross product term The shortest rotation axis direction that redirects the beam to the normal direction is given. Multiply the two by the gain. This constitutes a proportional feedforward correction; the basic deflection angle Solve using inverse kinematics: ,in The inverse kinematic mapping function of the galvanometer system; General instructions This ensures the beam reaches the target position while actively correcting the incident angle to approximate perpendicular incidence. A low-pass filter can also be introduced to smooth the control signal. ,in This is the filtering time constant to prevent high-frequency jitter from damaging the galvanometer motor.

[0043] Considering the response delay of the galvanometer itself Advance compensation can be added to the command: Further improve dynamic tracking performance.

[0044] Furthermore, the laser ablation effect is highly dependent on the incident angle. The cosine value of the actual incident angle is calculated based on the surface orientation component in the predicted aiming state vector. Using the tissue ablation threshold energy density obtained under vertical incident conditions as a reference, the reference energy density is divided by the cosine value to obtain the compensated energy density setting value. The compensated energy density setting value is converted into a laser drive current modulation signal, and the current setting value is updated and latched within one pulse cycle before the laser emission time.

[0045] This strategy is based on Lambert's cosine law: at oblique incidence, the energy density received per unit area is proportional to the cosine of the incident angle. Let the ablation threshold for perpendicular incidence be... The actual angle of incidence is The required compensation energy density for: in The corrected target beam direction (ideally equal to) To prevent the energy demand from approaching infinity at grazing angles, a maximum incident angle limit is set. : in This represents the upper limit for laser safety. Energy density and drive current. The relationship is usually non-linear, which can be addressed by calibrating the curve. The inverse function is obtained as follows: The real-time energy adjustment is ensured by updating the latch within one pulse cycle, preventing the current pulse from using old parameters due to communication delays. Furthermore, considering the spatial distribution of the laser beam (e.g., Gaussian distribution), the actual ablation effect is also related to the beam overlap rate, which can be addressed by introducing an overlap factor. Make corrections: in For scanning speed, The repetition frequency, Let be the waist radius.

[0046] Furthermore, open-loop compensation is insufficient to cope with long-term environmental drift, therefore closed-loop optimization is required. An adaptive correction of ablation effectiveness is performed after laser irradiation, specifically through a two-parameter attenuation feedback model, with the following formula: in, and These represent the fitted window radius of the corrected normal vector and the energy density compensation coefficient, respectively. and The corresponding parameter values ​​before correction. This represents the average ablation effectiveness feedback value from n consecutive operations. This represents the highest ablation effectiveness feedback value among the most recent m operations. To preset the effective threshold, , , All are preset positive real number adjustment factors; when consistently below When, the fitting window radius Exponential contraction to improve the accuracy of local surface fitting, energy density compensation coefficient. The ablation margin is increased nonlinearly according to the hyperbolic tangent law.

[0047] This model embodies two complementary adaptive mechanisms. For geometry awareness: when recent effects... Significantly lower than the historical best hour, The exponent term is less than 1, making Shrink. Smaller This means a more refined local fit, better capturing details such as stem bending or leaf wrinkling, and reducing incident angle errors caused by inaccurate normal estimation. Energy execution: When hour, By combining the overall logic of the formula, and relying on the saturation characteristics of the hyperbolic tangent function to constrain the range of parameter changes, we can prevent the system from oscillating due to excessive single adjustment of parameters.

[0048] Ablation effectiveness feedback quantity It can be obtained in various ways, such as color difference analysis based on post-irradiation images, thermal imaging temperature rise detection, or acoustic signal monitoring. Let the feedback quantity be... Its moving average is: The recent peak value is recorded to reflect the upper limit of the system's performance under ideal conditions: The coordinated adjustment of the two parameters enables online self-calibration of the "sensing-execution" dual-channel system, allowing the system to maintain stable weeding performance during long-term operation without manual intervention.

[0049] In summary, this technical solution constructs a complete dynamic laser weeding closed-loop control system through four key components: multi-source sensor fusion, rigid body kinematics prediction, geometry-energy joint optimization, and adaptive feedback. The formulas are interconnected, forming a rigorous logical chain from raw data to final execution, significantly improving the robustness and efficiency of the mobile platform in unstructured environments.

[0050] This embodiment implements a scenario of uniform straight-line weeding on leveled farmland. The mobile chassis maintains a constant travel speed, and the industrial camera, inertial measurement unit, and lidar synchronously collect data with a unified clock. The system identifies weed meristems and performs point cloud noise reduction and surface normal vector fitting to construct a six-dimensional aiming state vector. The rigid body pose matrix from exposure to light emission is obtained by integrating the inertial data. The aiming vector is then predicted by solving the coupled transfer model, and the galvanometer angle and laser energy are simultaneously corrected. The laser acts vertically on the weed meristems. A coaxial monitoring camera collects scorch images to complete dual-parameter adaptive correction, and a steel wire rope vibration isolator isolates minor road bumps. The entire process is suitable for continuous inter-row weeding operations on leveled plots.

[0051] Example 2 describes in detail a multi-source sensing motion-compensated laser weeding robot, specifically including a mobile chassis 11, a sensor bracket 2, a laser emitting module 6, and a main control computing unit 8. like Figure 5 As shown, the chassis adopts a portal frame integrated structure, divided into a left-side load-bearing box and a right-side load-bearing box. A hollow cavity is reserved between the left and right boxes, which is defined as the laser operation through window. Each of the left and right boxes is equipped with rubber anti-slip off-road wheels at the bottom, and the wheels are equipped with built-in drive motors to drive the whole machine to move continuously and at a constant speed along the farmland rows. The upper surfaces of both left and right boxes are machined with evenly distributed threaded countersunk holes, which serve as the mounting reference surface for the top load-bearing panel 12. The hollow through window is completely unobstructed from above and below, and the laser beam output by the laser emitting module 6 can pass vertically through the window and directly hit the weeds on the ground. There is no problem of structural obstruction of light spots, which is suitable for dynamic continuous weeding between rows and between plants.

[0052] The top support panel 12 has two independent metal plates, which are rigidly locked to the upper surfaces of the left and right boxes of the mobile chassis 11 by internal hex bolts. Each support panel has two sets of vertical mounting flange seats with positioning pin holes and threaded holes in the center of the flange seats for bottom positioning and locking of the sensor bracket column 21, realizing a rigid mechanical connection between the sensor bracket and the walking chassis without relative sliding. When the whole machine moves or bumps, the bracket and the chassis maintain a completely synchronized six-degree-of-freedom motion posture.

[0053] The electrical control box 13 is integrated into the outer closed cavity of the left and right boxes of the mobile chassis 11, and houses the DC power supply module, signal relay terminal block, and shielded cable storage trough.

[0054] The sensor bracket 2 is a gantry-type integrated rigid welded frame. The core components include two vertical columns 21, a top crossbeam 22, and a bottom mounting plate 23. The three are fixed together as an overall frame structure through a full welding process.

[0055] There is a vertical square profile column 21 on each side. The lower end of the column 21 is machined with a flange mounting plate. The flange plate is completely fitted and locked to the flange seat of the top support panel 12 below by positioning pins and locking bolts. The two columns 21 extend vertically upwards on the left and right sides, and the height is higher than the height of the weeds on the ground, so that the sensing equipment above will not be blocked by the stems and leaves of the crops. The inside of the column 21 is reserved with a wiring groove. The power supply line and synchronous trigger signal line of the industrial camera 3, inertial measurement unit 4, and lidar 5 are all stored inside the column wiring groove, realizing neat and protected wiring.

[0056] The top beam 22 is horizontally mounted at the very top of the two columns 21. The top surface of the beam is uniformly milled to form a plane mounting reference. Three types of multi-source synchronous sensing devices are positioned and assembled sequentially from left to right along the transverse direction of the beam 22. The installation order is industrial camera 3, inertial measurement unit 4, and lidar 5. The three share the same plane mounting reference to ensure that their spatial relative positions are fixed. The spatial extrinsic parameter calibration between the camera, IMU, and lidar is completed in one go at the factory stage, and no relative offset will occur during operation.

[0057] The industrial camera 3 is bolted to the left end of the crossbeam 22, with the lens pointing vertically downward toward the laser operation window to collect a color two-dimensional image of the weed meristem within the window. The camera has a built-in hardware synchronous trigger interface. The inertial measurement unit 4 (IMU) is fixed to the right side of the industrial camera 3 and is completely coplanar with the camera mounting plane. It has a built-in three-axis gyroscope and a three-axis accelerometer. The hardware synchronization pin is directly connected to the industrial camera 3 to achieve nanosecond-level synchronous sampling. The lidar 5 is mounted on the far right end of the crossbeam 22. The radar scanning head is vertically downward and outputs three-dimensional spatial point cloud data of weeds. The synchronous trigger line is connected in parallel with the camera and IMU, and the three share the same hardware synchronous clock signal issued by the main control.

[0058] The images, inertial attitude data, and raw 3D point cloud data collected by the three types of sensing devices are transmitted to the main control computing unit 8 via shielded data cables passing through the column cable tray and electrical control box 13.

[0059] The bottom mounting plate 23 is horizontally fixed to the lower middle section of the two columns 21, located directly below the top crossbeam 22 and above the laser operation window. The mounting plate is a square flat plate with through holes machined at the four corners. Each through hole corresponds to a set of wire rope vibration isolators 7. The upper flange of the wire rope vibration isolator 7 is bolted to the lower surface of the bottom mounting plate 23, and the lower flange of the wire rope vibration isolator 7 is used to hang and fix the laser emitting module 6. The bottom mounting plate 23 and the gantry support are a rigid integrated structure. The vibration generated by the chassis movement and bumps will be completely transmitted to the mounting plate, and then the vibration is attenuated by the wire rope vibration isolator 7 to achieve flexible buffering, thereby decoupling the mechanical vibration between the laser emitting module and the rigid support.

[0060] Four identical sets of wire rope vibration isolators 7 are set up in a rectangular arrangement at the four corners between the bottom mounting plate 23 and the laser emission module 6, forming a four-point flexible suspension vibration reduction structure.

[0061] Each set of vibration isolators consists of two metal flanges, one above the other, and multiple strands of cross-wound stainless steel wire rope. The upper flange is rigidly locked to the underside of the bottom mounting plate 23, and the lower flange is locked to the four corners of the upper surface of the laser emission module housing 61 by long bolts. The entire laser emission module is suspended by only four sets of wire rope vibration isolators, without any rigid connection or contact with the support.

[0062] When the mobile chassis travels over field ridges and potholes, causing bumps and vibrations, the vibrations are rigidly transmitted along the mobile chassis 11, the top load-bearing panel 12, the column 21, and the bottom mounting plate 23. The wire rope vibration isolator 7 absorbs the high-frequency vibration impact by relying on the elastic deformation of the wire rope itself, which greatly reduces the vibration amplitude transmitted to the precision optical components such as the galvanometer, lens, and beam splitter inside the laser emission module, thus avoiding slight deviations in the optical path and failure of calibration parameters.

[0063] like Figure 6 As shown, the laser emitting module 6 is the core of the entire laser execution system. It is externally packaged with a housing 61. The four corner flanges of the housing are fixed to the lower ends of four sets of steel wire rope vibration isolators 7. The interior of the housing is divided into three independent sealed cavities: a laser emitting cavity, an optical path scanning cavity, and a coaxial monitoring cavity. The assembly of each component and the optical path are described below: The module housing 61 is a fully enclosed metal sealed housing. Locking flanges are reserved at the four corners of the top surface of the housing for connecting the vibration isolator 7. Cable through holes are opened on the left side wall of the housing, and sealing connectors 66 are installed through the holes. The bottom of the housing is integrally connected to the bottom lens barrel 65, which opens downward to form a laser light output channel. A 45° lateral mounting port is opened on the side wall of the housing for mounting the beam splitter mount 102 and the coaxial monitoring camera 103. All assembly joints are equipped with dustproof and waterproof sealing rings to isolate field dust and water vapor from corroding the internal optical components.

[0064] The fiber laser 62 is horizontally locked and fixed inside the left cavity of the module housing 61. The laser is the source of laser energy. The output fiber optic cable of the laser passes through the sealing joint 66 and extends into the housing. The end of the fiber optic cable is inserted into the laser incident port of the two-dimensional galvanometer scanner 63. The sealing joint 66 is filled with waterproof and dustproof sealing filler to prevent external moisture and dust from entering the housing through the gaps in the fiber optic cable and contaminating the optical path. The power modulation control line of the fiber laser 62 is led out through the sealing joint 66, connected to the electrical control box 13, and then connected to the main control computing unit 8 to receive the energy compensation drive signal issued by the main control unit.

[0065] The two-dimensional galvanometer scanner 63 is fixed in the optical path cavity in the middle of the housing and is located downstream of the output optical path of the fiber laser 62. The galvanometer is equipped with two sets of high-speed deflecting reflective mirrors, X and Y. The mirrors receive the galvanometer control electrical signals sent by the main controller and synchronously complete the two-axis angle deflection, changing the horizontal and vertical propagation paths of the laser beam. The lower end of the galvanometer is vertically connected to the focusing lens group 64. The laser beam after being deflected by the galvanometer is vertically injected into the focusing lens group 64.

[0066] The focusing lens group 64 is composed of multiple stacked optical convex lenses, which are nested inside the bottom lens tube 65. The bottom lens tube 65 is a cylindrical structure, with its upper end sealed and connected to the galvanometer scanner 63, and its lower end vertically facing down towards the ground laser operation window. After being deflected by the galvanometer, the laser beam passes through the focusing lens group 64, and the lens gathers the diverging laser into a high-energy fine spot, which is then projected downward along the vertical laser beam axis onto the weed meristem in the field.

[0067] A 45° installation window is obliquely opened on the side wall of the bottom lens tube 65. A beam splitter mount 102 is fixed on the outside of the window, and a 45° beam splitter 101 is installed inside the beam splitter mount 102. The beam splitter is obliquely arranged on the vertical main laser emission axis. The beam splitter can transmit the downward propagating weeding laser and reflect the environmental reflected light after the weeds are burned upward, forming an independent monitoring optical path and realizing the coaxial arrangement of laser emission and image synchronous monitoring.

[0068] The coaxial monitoring camera 103 is horizontally locked and fixed to the outside of the beam splitter mount 102, with the camera lens facing the 45° beam splitter 101. The reflected light generated after the weeds are burned by the laser propagates upward along the vertical main optical path, and after being horizontally reflected by the beam splitter 101, it enters the lens of the coaxial monitoring camera. The camera collects images of the burned area in real time. The image transmission cable of the camera is led outward, passes through the sealed connector of the module housing, and connects to the electrical control box 13. Finally, the ablation effect image data is transmitted back to the main control computing unit 8, providing the original image basis for the main control to perform dual-parameter adaptive feedback correction.

[0069] Two-dimensional galvanometer scanner 63 control output port: After the main controller completes the calculation of platform-galvanometer joint motion compensation and incident angle deviation correction, it generates two-axis deflection control electrical signals, which are transmitted to the galvanometer via shielded cable to dynamically adjust the laser spot aiming coordinates; Fiber laser 62 power modulation output port: The main controller calculates the adaptive compensation energy density based on Lambert's law, converts it into a drive current control signal and transmits it to the fiber laser to dynamically adjust the laser output energy. Mobile chassis walking drive output port: The main controller synchronously outputs the walking speed control signal to the chassis walking motor to maintain the machine's uniform and continuous movement, matching the dynamic operation rhythm of laser weeding.

[0070] The mobile chassis 11 drives the gantry-type sensor support 2 forward at a constant speed. The industrial camera 3, IMU 4, and lidar 5 mounted on the top beam 22 synchronously collect multi-source sensing data under the control of a unified hardware clock. All data is sent to the main control computing unit 8 in real time. The main control unit runs the multi-source sensing motion compensation laser weeding operation method of this application, completes point cloud preprocessing, surface normal vector fitting, six-dimensional target state construction, and platform joint motion compensation deduction from exposure to light emission period, and solves the galvanometer deflection command and laser energy parameters adapted to the dynamic movement conditions. The control signal is sent to the two-dimensional galvanometer scanner 63 and fiber laser 62 inside the laser emission module 6, and the focused laser is output to accurately hit the weed meristem during the movement. The laser emission module 6 relies on the four corner steel wire rope vibration isolators 7 to achieve flexible vibration reduction and decoupling from the support, which greatly reduces the interference of chassis bumps and vibrations on internal optical components and ensures long-term stability of optical path calibration. The coaxial monitoring camera 103 synchronously acquires images of weed burning and transmits them back to the main control. The main control calculates the feedback amount of ablation effectiveness based on the images, iteratively updates the surface fitting search radius and energy compensation coefficient online, and completes the closed-loop adaptive correction of the entire process. The entire mechanical structure, optical path structure, and electrical signal transmission link fully support the weeding operation method of this application, which features time-series feedforward motion compensation, multi-source fusion sensing, and closed-loop self-correction.

[0071] This embodiment is adapted to complex farmland conditions with uneven terrain and fluctuating travel speed. Multiple sensors synchronously collect dynamic attitude and weed geometry data at high frequency, expanding the search radius of the point cloud neighborhood to improve the robustness of surface fitting. Based on the inertial high-frequency sampling data, multiple short-time integral intervals are subdivided to accurately calculate the composite offset caused by platform pitch and roll disturbances, optimizing the attitude and position coupling transfer model. The galvanometer advance compensation coefficient is increased, and the laser energy adjustment range is simultaneously widened. The algorithm parameters are continuously corrected by coaxial monitoring closed loop. The flexible vibration reduction structure significantly reduces the interference of severe vibration on optical path calibration, adapting to the dynamic weeding needs of unstructured undulating farmland.

[0072] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A multi-source sensing motion-compensated laser weeding method, characterized in that, include: During the movement of the mobile platform, visual images, inertial measurement data, and laser ranging point clouds are collected simultaneously. The target weed meristem is identified from the visual image and its pixel coordinates are extracted. At the same time, a three-dimensional point cluster corresponding to the target weed is segmented from the laser ranging point cloud. The local surface normal vector of the target weed meristem is obtained by fitting the three-dimensional point cluster. The rigid body pose transformation matrix of the mobile platform between the camera exposure time and the laser emission time is calculated using the inertial measurement data; The pixel coordinates and the local surface normal vector are jointly projected onto the laser emission coordinate system to construct a six-dimensional aiming state vector that includes the target position and the target surface orientation; The six-dimensional aiming state vector is transformed by the rigid body pose transformation matrix to obtain the predicted aiming state vector at the moment of laser emission. Based on the surface orientation component in the predicted aiming state vector, the deflection angle of the laser galvanometer and the pulse energy of the laser are dynamically adjusted so that the laser beam acts on the target weed meristem with an incident angle perpendicular to the surface of the target weed meristem and with an effective energy density matching the tissue ablation threshold at that incident angle.

2. The multi-source sensing motion-compensated laser weeding method according to claim 1, characterized in that, The local surface normal vector of the target weed meristem is obtained based on the fitting of the three-dimensional point clusters, including: The three-dimensional point cluster is subjected to voxel downsampling and statistical outlier filtering to obtain a cleaned point set; Using the three-dimensional projection point corresponding to the pixel coordinates as the query center, search for k nearest neighbors in the cleaned point set to form a local neighborhood; Principal component analysis is performed on the points in the local neighborhood, and the unit eigenvector corresponding to the minimum eigenvalue is taken as the initial normal vector. The initial normal vector is oriented to be checked for consistency using the gradient direction of the target weed meristematic region in the visual image. When the angle between the initial normal vector and the gradient direction is greater than 90 degrees, it is reversed, and the checked local surface normal vector is output.

3. The multi-source sensing motion-compensated laser weeding method according to claim 2, characterized in that, Construct a six-dimensional aiming state vector that includes the target position and the target surface orientation, including: The pixel coordinates are back-projected onto the camera coordinate system using the camera intrinsic parameter matrix and depth values ​​to obtain the three-dimensional position components of the target. Transform the verified local surface normal vector to the same camera coordinate system to obtain the target surface orientation component; The three-dimensional position component of the target and the orientation component of the target surface are concatenated into a six-dimensional column vector, and marked with a timestamp of the camera exposure time to form a time-stamped six-dimensional aiming state vector.

4. The multi-source sensing motion-compensated laser weeding method according to claim 3, characterized in that, The six-dimensional aiming state vector is transformed using the rigid body pose transformation matrix to obtain the predicted aiming state vector at the moment of laser emission. This is specifically achieved through an attitude-position coupling transfer model, as shown in the formula: in, Let be the predicted aiming state vector at the moment of laser emission. This refers to the moment when the laser light is emitted. Let be the six-dimensional aiming state vector at the moment of camera exposure. For the camera exposure time, and These are the data obtained from inertial measurement at time intervals. The rotation submatrix and translation subvector obtained by inner integration This represents a custom attitude-position coupling operator, with a rotation submatrix. Simultaneously acting on Position and orientation components, translation subvectors Only the position component is superimposed, and the orientation component is normalized after transformation.

5. The multi-source sensing motion-compensated laser weeding method according to claim 4, characterized in that, The deflection angle of the laser galvanometer is dynamically adjusted based on the surface orientation component in the predicted aiming state vector, specifically through incident angle deviation driving, as shown in the formula: in, Let be the target deflection angle vectors of the two axes of the galvanometer. This is the basic deflection angle vector calculated solely based on the target position. To predict the surface orientation component in the aiming state vector, This is the unit vector pointing to the current output optical axis of the laser. This is the preset verticality gain coefficient. This represents the cross product operation of vectors; The incident angle deviation is converted into an angular correction amount that is consistent with the deviation direction, and then fed forward and superimposed onto the base deflection angle.

6. The multi-source sensing motion-compensated laser weeding method according to claim 5, characterized in that, Dynamically adjusting the pulse energy of the laser, including: The cosine of the actual incident angle is calculated based on the surface orientation component in the predicted aiming state vector. Using the tissue ablation threshold energy density obtained under vertical incidence conditions as a reference, the reference energy density is divided by the cosine value to obtain the compensated energy density setting value; The compensated energy density setting value is converted into a laser drive current modulation signal, and the current setting value is updated and latched within one pulse cycle before the laser emission time.

7. The multi-source sensing motion-compensated laser weeding method according to claim 6, characterized in that, After laser irradiation, an adaptive correction for ablation effectiveness is performed, specifically through a two-parameter attenuation feedback model, as shown in the formula: in, and These represent the fitted window radius of the corrected normal vector and the energy density compensation coefficient, respectively. and The corresponding parameter values ​​before correction. This represents the average ablation effectiveness feedback value from n consecutive operations. This represents the highest ablation effectiveness feedback value among the most recent m operations. To preset the effective threshold, , , All are preset positive real number adjustment factors; when consistently below When, the fitting window radius Exponential contraction to improve the accuracy of local surface fitting, energy density compensation coefficient. The ablation margin is increased nonlinearly according to the hyperbolic tangent law.

8. A multi-source sensing motion-compensated laser weeding robot, used in the multi-source sensing motion-compensated laser weeding operation method according to any one of claims 1 to 7, characterized in that, Includes a mobile chassis, sensor bracket, laser emission module, and main control computing unit; The sensor bracket is a welded aluminum alloy frame, and its four bottom corners are rigidly fixed to the top support panel of the mobile chassis by bolts. An industrial camera, an inertial measurement unit, and a lidar are fixedly mounted on the top crossbeam of the sensor bracket from left to right using screws. The mounting surfaces of the industrial camera, the inertial measurement unit, and the lidar are coplanar and their optical axes or sensing axes are parallel to each other. The laser emitting module includes a fiber laser, a two-dimensional galvanometer scanner, a focusing lens group, and a module housing; the fiber laser is fixedly installed on the inner side wall of the module housing, and the output fiber passes through a sealed connector on the side wall of the housing and is coupled into the light inlet of the two-dimensional galvanometer scanner. The two-dimensional galvanometer scanner is fixed to the inner side of the top plate of the module housing by four countersunk screws, and the light outlet is vertically downward and coaxially connected to the focusing lens group; the focusing lens group is fixed to the bottom lens barrel of the module housing by screwing. A steel wire rope vibration isolator is installed at each of the four corners of the top plate of the module housing. The upper end of each steel wire rope vibration isolator is fixed to the bottom mounting plate of the sensor bracket by flange bolts, and the lower end is fixed to the outside of the top plate of the module housing by flange bolts, so that the laser emission module is suspended below the sensor bracket and forms a flexible connection structure with mechanical decoupling. The main control computing unit is fixedly installed in the internal electrical control box of the mobile chassis. It is electrically connected to the industrial camera and lidar via gigabit Ethernet cable, electrically connected to the inertial measurement unit via RS-485 bus, and electrically connected to the fiber laser and two-dimensional galvanometer scanner via analog control cable.

9. The multi-source sensing motion-compensated laser weeding robot according to claim 8, characterized in that, An oblique mounting hole is provided on the bottom lens barrel side wall of the module housing. A beam splitter mount is fixedly installed in the oblique mounting hole. A beam splitter is embedded in the beam splitter mount. The reflective surface of the beam splitter faces the light output direction of the focusing lens group and is arranged at a preset folding angle with the light output optical axis. A coaxial monitoring camera is also fixedly installed on the outer side wall of the bottom lens barrel of the module housing by fasteners. The optical axis of the coaxial monitoring camera passes through the beam splitter mount and is reflected and refracted by the beam splitter, and is precisely coaxial with the output optical axis of the focusing lens group. The imaging sensor target surface of the coaxial monitoring camera and the focal plane of the focusing lens group satisfy the optical conjugate relationship. The coaxial monitoring camera is electrically connected to the main control computing unit via a high-speed image transmission cable.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-source sensing motion-compensated laser weeding method as described in any one of claims 1 to 7.

Images