A machine vision-based unmanned laser weeding method

Abstract

The application provides a kind of unmanned laser weeding method based on machine vision, comprising: determining the weed body corresponding to the target area and the laser impact point, and the laser action time corresponding to each laser impact point;According to the laser focusing distance and the laser angle deflection amplitude corresponding to the trajectory segment between each adjacent two laser impact points, the laser emission power corresponding to the trajectory segment between each adjacent two laser impact points of laser emitter is determined;Control laser emitter to carry out laser impact on the laser burning trajectory formed by a plurality of laser impact points with the laser action time and the laser emission power corresponding to each laser impact point;After laser emitter carries out laser impact on weed body, according to the pixel area and the average gray difference of the image corresponding to weed body before and after laser impact, the laser weeding result of weed body is determined, to realize accurate verification and secondary supplementary strategy of weeding result, improve weeding accuracy and success rate.

Description

Technical Field

[0001] This invention relates to the field of agricultural weeding, and in particular to an unmanned laser weeding method based on machine vision. Background Technology

[0002] With the continuous rise in agricultural labor costs and the increasing demand for refined farmland management, weeding has become one of the key bottlenecks in agricultural production. Traditional manual weeding is inefficient and arduous, while chemical weeding, although more efficient, causes a series of problems such as soil pollution, excessive pesticide residues, and ecological damage, making it difficult to meet the requirements of green agriculture and sustainable development. Therefore, developing an efficient, environmentally friendly, and intelligent physical weeding technology has become an important direction in current agricultural automation research.

[0003] In recent years, laser weeding technology has gradually become a research hotspot in the field of physical weed control. Lasers are characterized by concentrated energy, rapid response, and precise action, capable of instantly burning weed tissue with a high-energy beam, avoiding the use of pesticides. However, existing laser weeding equipment is mostly semi-automatic or fixed systems, typically relying on tracks or manual propulsion, and cannot adapt to complex terrain or automatically identify targets. Existing technologies largely rely on two-dimensional vision and static threshold judgment, failing to accurately locate weed targets and control laser energy, and lacking post-hit feedback mechanisms, easily leading to missed or false weeds. Summary of the Invention

[0004] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows: According to one aspect of this application, a machine vision-based unmanned laser weeding method is provided, comprising: Step S100: When the laser weeding device moves to the target area, perform image analysis on the target area to determine the weeds and several laser strike points in the weeds. Step S200: Determine the laser action duration corresponding to each laser strike point based on the area of ​​the weeds, the complexity of the crop boundary around the weeds, and the laser emission distance; Step S300: Determine the laser emission power of the laser emitter of the laser weeding device for the trajectory segment between each two adjacent laser strike points based on the laser focusing distance and laser angle deflection amplitude corresponding to the trajectory segment between each two adjacent laser strike points. Step S400: Control the laser emitter to perform laser strikes on the laser burning trajectory composed of several laser strike points with the laser action duration and laser emission power corresponding to each laser strike point. Step S500: After the laser emitter finishes laser-attacking the weeds, the laser weeding result is determined based on the pixel area and average grayscale difference of the corresponding images of the weeds before and after laser-attacking.

[0005] The present invention has at least the following beneficial effects: The machine vision-based unmanned laser weeding method of this invention performs image analysis on the target area when the laser weeding device moves to the target area. Through multispectral visual perception and curvature boundary analysis algorithms, it identifies the weeds and several laser impact points within them, improving the accuracy of weed target segmentation and scorch point positioning. Then, based on the area of ​​the weeds, the complexity of the surrounding crop boundaries, and the laser emission distance, it determines the laser action duration for each laser impact point. Finally, based on the laser focusing distance and laser angle deflection amplitude corresponding to the trajectory segment between each pair of adjacent laser impact points, it determines the laser emission of the laser emitter of the laser weeding device for each pair of adjacent laser impact points. Power is used to precisely control laser beam parameters for different weed sizes, structures, and locations, dynamically adjusting the power output of each laser trajectory segment to achieve personalized strikes. After determining the laser beam parameters, the laser emitter is controlled to strike the weeds along a laser burning trajectory composed of several laser strike points with the laser action duration and laser emission power corresponding to each laser strike point. After the laser emitter finishes striking the weeds, the pixel area and average grayscale difference of the weeds before and after the laser strike are analyzed in real time based on the back-feeding thermal texture changes after weeding by introducing a backward thermal loss feedback fusion mechanism. This determines the laser weeding result of the weeds, enabling accurate verification of the weeding result and a secondary strike strategy, thereby improving the accuracy and success rate of weeding. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 A flowchart of a machine vision-based unmanned laser weeding method provided in an embodiment of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] This application proposes a machine vision-based unmanned laser weeding method, such as... Figure 1 As shown, it specifically includes: Step S100: When the laser weeding device moves to the target area, perform image analysis on the target area to determine the weeds and several laser strike points in the weeds. Furthermore, step S100 includes steps S110-S140: Step S110: Obtain the target region image corresponding to the target region; Step S120: Perform image enhancement processing, illumination normalization processing, and edge detection processing on the target region image to obtain the processed image; Step S130: Extract the image of the weeds corresponding to the weeds from the processed image using a preset image segmentation algorithm, so as to determine the weeds corresponding to the weeds in the target area. Step S140: If the area of ​​the weed body is less than or equal to the preset area threshold, then the geometric centroid corresponding to the weed body is determined as the laser strike point of the weed body; If the area of ​​the weeds is greater than the preset area threshold, then several laser strike points of the weeds are generated according to the preset edge equidistant sampling algorithm.

[0010] Traditional weeding systems struggle to accurately distinguish between weeds and crops in actual farmland, especially when crops and weeds are mixed, have similar colors, or overlap in shape, resulting in a high misjudgment rate. Therefore, the laser weeding device of this application uses a camera equipped with a multispectral lens to acquire images, performs image enhancement, illumination normalization, and edge detection processing, and employs image segmentation algorithms (such as a segmenter based on a U-Net network) to extract each individual weed. Multispectral visual perception and curvature boundary analysis algorithms are also used to improve the target segmentation of weeds and the positioning accuracy of scorch points.

[0011] Step S200: Determine the laser action duration corresponding to each laser strike point based on the area of ​​the weeds, the complexity of the crop boundary around the weeds, and the laser emission distance; Furthermore, step S200 includes steps S210-S220: Step S210: Obtain the distance between each laser strike point and the laser emitter to obtain a list of emission distances A=(A1,A2,...,Ai,...,Aj); where i=1,2,...,j; j is the number of laser strike points in the weeds; Ai is the distance between the i-th laser strike point and the laser emitter; The distance between the laser strike point and the laser emitter is the Euclidean distance (spatial slant distance) from the centroid of the weeds to the laser emission point, which is calculated through visual coordinate system transformation and depth estimation. The greater the distance, the more significant the energy attenuation, so laser time compensation is required.

[0012] Step S220: Determine the laser action duration T corresponding to the i-th laser strike point. i =(m1×log(1+S1)+m2×S2 1 / 2 )×(1+e (-Ai / n) ); Where m1 and m2 are preset energy-time scheduling weighting coefficients, obtained through extensive experimental data and regression fitting of the energy consumption minimization objective function, used to adjust the intensity of the influence of area and complexity on the strike time; log() is a preset logarithmic function; S1 is the pixel area of ​​the weed, obtained by counting the number of pixels in the region after image segmentation. The larger the area, the thicker the target, and the longer the required laser time should be; S2 is the boundary complexity obtained by integrating the Gaussian curvature of the edge contour between the weed and the surrounding crops. The boundary complexity index of the weed and the surrounding crops is obtained by integrating the Gaussian curvature of the edge contour between the weed and the surrounding crops. The more complex it is, the more irregular or varied the shape of the leaves, and the more precise the laser strike needs to be; n is the preset system thermal attenuation characteristic distance, which is an empirical value determined by the laser beam divergence characteristics and the crop evaporative heat transfer model; e () This is a preset exponential function.

[0013] The duration of laser action corresponding to the laser strike point is used to control the duration of laser emitter output, so that heat transfer is consistent with the transpiration and dehydration of weeds.

[0014] In step S220, m1×log(1+S1) reflects that the larger the area, the longer the burning time. The log function is introduced to suppress the surge effect when the area is too large, preventing energy waste; m2×S2 1 / 2 This demonstrates that the more complex the boundary (such as a serrated blade), the longer the strike time, and the more precise the energy coverage required for complex boundaries; 1+e (-Ai / n)Considering distance factors, the laser diffusion becomes more pronounced with greater distance, requiring a moderately longer irradiation time. By integrating multi-dimensional sensing information and taking into account target size, shape complexity, and spatial distribution, the laser irradiation behavior becomes more targeted. By matching the time scheduling strategy with the thermal response model of plant leaf transpiration structure, it can achieve maximum damage with minimum heat power consumption. By introducing boundary complexity and attenuation distance factors, it ensures that each strike point is neither ineffective due to insufficient energy nor burns surrounding crops due to energy redundancy. By adjusting m1, m2, and n, it can flexibly adapt to different types of crops, different climatic environments, and different laser emitter parameters, achieving personalized and precise control of the laser irradiation time for each strike point. This effectively improves the energy efficiency ratio and target matching degree of weeding, especially in cases with complex weed morphology and growth close to crops.

[0015] Step S300: Determine the laser emission power of the laser emitter of the laser weeding device for the trajectory segment between each two adjacent laser strike points based on the laser focusing distance and laser angle deflection amplitude corresponding to the trajectory segment between each two adjacent laser strike points. Furthermore, step S300 includes steps S310-S350: Step S310: Based on spatial correlation, sort several laser strike points to obtain the rank of each laser strike point after sorting. Step S320: Based on the preset fifth-order B-spline interpolation algorithm, generate a laser burning trajectory composed of several laser impact points in ascending order of rank. Step S330: Discretize the laser burning trajectory to obtain several sub-trajectory points on the laser burning trajectory; on the laser burning trajectory, the distance between any two adjacent sub-trajectory points is equal; Step S340: Divide the laser burning trajectory into several sub-trajectory segments according to every two adjacent laser impact points on the laser burning trajectory; the two endpoints of each sub-trajectory segment are two laser impact points that are rank adjacent. Step S350: Determine the laser emission power P = e1 × ((∑) at any sub-trajectory segment. g f=2 ||B f -B f-1 || 2 ) / (g×C1 2 ))×(1+e2×C2); Where e1 is a preset laser power reference adjustment factor used to control the overall power reference value and adapt to different laser emitter models; f=2,...,g; g is the number of sub-trajectory points included in this sub-trajectory segment, representing the segmentation granularity of the trajectory segment; ||B f -B f-1|| represents the Euclidean distance between the f-th and (f-1)-th sub-trajectory points in the sub-trajectory segment, used to measure the physical length of the trajectory segment. A larger distance requires increased power to compensate for heat diffusion. C1 represents the laser focusing distance corresponding to the sub-trajectory segment, measured by a laser rangefinder from the target to the lens. A larger C1 indicates stronger laser dispersion, meaning more laser emission power is needed. C1 can be the average of several laser focusing distances on the sub-trajectory segment, or the variance of several laser focusing distances on the sub-trajectory segment, serving as a power fluctuation correction factor. C2 represents the distance between the laser emitter and the target. The laser angle deflection amplitude corresponding to the sub-trajectory segment is calculated by the change of the servo rotation angle. If the laser frequently changes direction in the trajectory, it indicates a decrease in energy transfer efficiency, so the laser emission power should be increased to compensate. e2 is a preset angle compensation amplification coefficient, which is used to control the weight of C2 in power adjustment. e2 can be set empirically (a fixed value is set in empirical testing (such as 0.5-1.5)) or dynamically adjusted by an adaptive algorithm (the system evaluates the trajectory angle variation statistics (such as variance and range) in real time, and then uses function fitting or training model to dynamically assign values).

[0016] In step S350, the laser emission power corresponding to any sub-trajectory segment is used to control the energy output level of the laser emitter during the operation of that segment. The geometric trajectory structure, along with focal length and angle changes, are introduced into the power adjustment model to achieve dynamic matching of thermal energy in continuous spatial distribution. An angle-driven and heat-complementary mechanism is adopted, that is, C2 and e2 are used to introduce the influence of laser direction changes to achieve real-time coupling of direction and thermal power. The C1 parameter can dynamically adapt to uneven ground and different target heights, significantly improving energy focusing efficiency. Practice has shown that compared with the traditional fixed power system, this method can reduce redundant laser output by about 30%, extend the life of the laser emitter, and improve the endurance of weeding operations.

[0017] The laser emission power and the aforementioned laser action duration are indirectly coupled. Unlike the traditional single-point emission strategy, the laser weeding method of this application emphasizes "coherence, thermal uniformity, and directional stability." That is, the laser emitter moves between all laser strike points along a smooth curve trajectory (such as a fifth-order B-spline). In each trajectory segment, the required average power is dynamically calculated based on the curve variation amplitude and focusing state to avoid heat loss or scorching. In a weedy area containing multiple laser strike points (scorching points), a continuous laser scorching trajectory is constructed based on the positional relationship of these points, and the power distribution is dynamically adjusted to ensure stable heat transfer. The focal length is dynamically set based on the depth change of each trajectory point from the ground (estimated by 3D model / laser ranging) to make the energy more concentrated or more diffused.

[0018] Existing technologies often use fixed power or static paths for laser weeding, which cannot precisely control laser beam parameters for different weed sizes, structures, and locations. This invention proposes a dynamic thermal beam trajectory control configurator that adjusts the power output of each laser trajectory segment based on curvature, distance, and focusing status. It assesses the complexity of the boundary between each weed and surrounding crops, primarily through contour curvature integral measurement, to obtain the three-dimensional spatial distance between the laser emitter and the weed scorching point, correcting laser transmission attenuation for personalized targeting. The mechanism describing the laser emission power in this invention innovatively introduces a fifth-order B-spline interpolation algorithm for laser... The method constructs a smooth, continuous path using a set of impact points, and dynamically calculates the average power required for each trajectory segment by combining changes in focus (C1) and angle deflection (C2). This significantly enhances the continuity, directional stability, and precision of heat control in laser ablation. Particularly during dynamic ablation, the power adjustment model, which links trajectory curvature changes with spatial depth, is significantly superior to traditional static trajectory or path selection methods. It truly achieves on-demand energy output based on the principle of balanced heat conduction. Furthermore, this method balances ablation efficiency and operational safety, effectively preventing heat accumulation or insufficient ablation of the laser in high-density weedy vegetation. In summary, the "Dynamic Thermal Beam Track Control Configuration" not only fills the gaps in existing solutions regarding path continuity and energy regulation mechanisms but also constructs an intelligent trajectory and power collaborative control framework for precision agriculture weeding operations, demonstrating outstanding substantive features and innovative technological contributions.

[0019] Step S400: Control the laser emitter to perform laser strikes on the laser burning trajectory composed of several laser strike points with the laser action duration and laser emission power corresponding to each laser strike point. Step S500: After the laser emitter finishes laser-attacking the weeds, the laser weeding result is determined based on the pixel area and average grayscale difference of the corresponding images of the weeds before and after laser-attacking.

[0020] Furthermore, step S500 includes steps S510-S520: Step S510: After the laser emitter finishes striking the weeds with laser, determine the corresponding burning success score Z = (a1 × |D1∩D2| / |D2|) + a2 × (D3 / (1 + log(1 + ∑ j i=1 T i ))); Where, a1 is the weighting coefficient for controlling the region matching weight; a2 is the weighting coefficient for controlling the image change weight; 0≤a1≤1; 0≤a2≤1; a1+a2=1; D1 is the set of pixels in the thermally damaged area extracted from the image after laser attack on the weeds, extracted by thermal texture changes and infrared response differences, usually a polygonal mask area; D2 is the set of pixels in the image before laser attack on the weeds, which is the reference area obtained in the target recognition stage; |D1∩D2| represents the thermally damaged area of ​​the weeds after laser attack and the weeds The pixel overlap area of ​​the weeds before laser strike indicates whether the heat is actually applied to the target; a higher value indicates a more accurate hit. |D2| is the pixel area in the image before laser strike, used for normalization calculation. D3 is the difference between the average pixel grayscale of the image after laser strike and the average pixel grayscale of the image before laser strike (using image registration algorithms (such as feature point matching + affine transformation) to ensure alignment of the two image regions before and after laser strike). This represents the degree to which the weeds darken, indirectly reflecting the degree of thermal damage; a larger D3 indicates more significant thermal damage.

[0021] In step S510, the scorching success score Z is an evaluation index for whether the weed has been successfully hit. The closer the scorching success score Z is to 1, the more successful the hit; the closer the scorching success score Z is to 0, the less accurate the hit, the less thoroughly burned, or the completely missed hit. (a1×|D1∩D2| / |D2|) is a spatial matching term, used to measure whether the actual heat damage area accurately hits the predetermined weed location, which is a reflection of the positioning accuracy. a2×(D3 / (1+log(1+∑ j i=1 T i ))) represents the texture variation term, used to measure whether burning effectively causes plant degeneration (such as blackening or collapse), log(1+∑ j i=1 T i It has an inhibitory effect, preventing prolonged exposure from causing a high "false effect" score.

[0022] The laser weeding results verification method in this application adopts a posterior visual feedback loop, which is not a "hit and go" approach, but rather a decision on whether to complete the task after verifying the results, thus constructing a truly closed-loop weeding system. Through spatial-texture dual criterion fusion, it focuses on both whether the thermal energy hits the correct area and its destructive effect, overcoming the traditional problems of false kills / false positives. It also establishes a strike energy efficiency correction mechanism, with the score correlated with the irradiation time, which can identify "redundant strikes" and provide feedback to adjust parameters, facilitating model training and dynamic parameter tuning. The burn success score can be used as one of the reward functions for reinforcement learning, constructing a long-term optimization system.

[0023] Step S520: If the burn success score corresponding to the weed is greater than the preset score threshold, then the laser weeding result corresponding to the weed is determined to be successful; otherwise, return to step S200.

[0024] Current laser weeding devices struggle to assess the effectiveness of laser weed removal after the initial application, often resulting in false positives such as "the system has judged success even though the weeds haven't been completely burned." Therefore, this invention introduces a backward thermal loss feedback fusion mechanism to analyze thermal texture changes in real time after weeding, enabling accurate verification and a secondary application strategy. Unlike simple "whether it has turned black" identification, this method integrates two dimensions: spatial overlap rate and thermal texture change. A unified scoring function outputs a success metric (i.e., a burning success score Z). If the score is below a preset threshold, it indicates poor burning effect, and the weed will be re-added to the application queue for secondary processing.

[0025] On the other hand, farmland that typically requires weeding often presents complex terrain such as sandy areas, slopes, and potholes. If the drive wheels of the laser weeding device slip or fail, the device will be unable to continue operating. Therefore, this application also proposes a method for adjusting the output torque of the steering wheel of the laser weeding device, designing an anti-imbalance torque control mechanism and a limp-walking compensation strategy to support stable escape and movement even in the event of failure of up to three drive wheels. Specifically, steps S610-S620 are included. Step S610: During the movement of the laser weeding device, the target output torque of each steering wheel is determined based on the angular velocity of each steering wheel of the laser weeding device and the hardness of the terrain. Among them, the target output torque E of the h-th steering wheel of the laser weeding device h =E0×(1-tanh(b1×G h +b2×(|H h -H0| / H0)))+F1×(∑ d c=1 I hc ×J c ) / (1+F2×L hc h = 1, 2, ..., k; k is the number of steering wheels in the laser weeding device; In the formula, E0 is the standard torque setting value of the laser weeding device's steering wheel on a stable surface, used as the calculation reference baseline; tanh() is the preset hyperbolic tangent function; b1 is the preset softness / hardness suppression weight; G hThe terrain hardness index at the current location of the h-th steering wheel is calculated by combining the change in height per unit area (the change in height per unit area is determined by extracting the terrain height difference near each drive point using lidar and inertial measurement unit, which can be calculated using point cloud difference or laser ranging gradient) and vibration change (the standard deviation of the acceleration fluctuation of the steering wheel within a micro-time window, obtained by inertial measurement unit); b2 is the preset angular velocity offset modulation coefficient; H h The angular velocity of the h-th steering wheel at the current moment is sampled in real time to determine if it deviates too much from the average value; H0 is the average angular velocity of all normally functioning steering wheels at the current moment, used for dynamic comparison to identify abnormal load conditions on a single wheel; F1 is the preset total compensation gain coefficient; c=1,2,...,d; d is the number of failed steering wheels at the current moment, automatically identified and updated by the system fault monitoring module (when a steering wheel experiences a driving abnormality (such as abnormal current or abnormal rotation), it is identified in real time and recorded in the "failed steering wheel set"); I hc The compensation weight of the h-th steering wheel to the c-th failed steering wheel is generally determined by diagonal or physical adjacency relationships: a higher diagonal weight results in a lower far-end weight; 0 ≤ I hc ≤1; J c The standard torque corresponding to the c-th failed steering wheel can be E0 or a historical value; F2 is the preset compensation attenuation index; L hc The distance between the h-th steering wheel and the c-th failed steering wheel is the physical center distance, used to control and compensate for energy decay.

[0026] In this invention, the terrain hardness index G of the steering wheel is used. h The calculation did not employ existing single-sensor feedback mechanisms or traditional terrain classification methods. Instead, it designed a multimodal softness and stiffness estimation model that integrates unit area height variation and vibration disturbance standard deviation. h It not only measures the geometric undulation of the terrain (obtained through laser ranging, structured light, or point cloud differencing to assess height variation per unit area), but also comprehensively considers the fluctuations in micro-vibration amplitude at that location within a small time window. This data can be acquired through embedded inertial measurement unit sensors. The combination of these two methods forms a joint perception of the damping characteristics, structural rigidity, and local compliance of the terrain beneath the steering wheel. In agricultural operation scenarios, many complex terrains (such as wet soft soil, gravel layers, and hard mud) may have similar geometric appearances, but their elasticity and vibration responses are drastically different. Therefore, G... h The proposed model represents a functional breakthrough over traditional terrain classification methods that primarily rely on elevation changes. More importantly, this model can be rapidly computed using real-time sampled data, requiring no offline training or feature extraction. It exhibits strong adaptability, low computational overhead, and is suitable for deployment in practical weeding platforms. In summary, G... hThe construction method is a unique sensing mechanism proposed by this invention in the field of torque control programming. It has clear inventiveness, engineering practicality and feasibility, and significantly improves the platform's adaptability and travel stability in complex farmland environments.

[0027] In the drive redundancy control strategy, the compensation weight coefficient I proposed in this invention hc Instead of originating from traditional shortest path allocation algorithms or linear optimization models, this is a "structure-driven weight mapping mechanism" built upon the physical topology, positional symmetry, and dynamic failure states of the platform's multi-steering wheel structure. Specifically, when a steering wheel is detected as failing, the system automatically assigns different compensation weights based on its spatial relationship with other steering wheels (e.g., diagonal, adjacent, or distant). Diagonal steering wheels, possessing stronger thrust vector cancellation capabilities, typically undertake higher compensation tasks, while adjacent or distant steering wheels, with longer compensation paths and weaker torque coupling, have lower I... hc The value is relatively low. Furthermore, this mechanism also considers the physical center distance L between the steering wheels. hc An exponential decay function is introduced to ensure that long-distance compensation does not cause system oscillations or torque conflicts. This dynamic weighting method, designed based on platform structural characteristics and redundant operation strategies, differs significantly from the traditional "fault masking + redistribution" control method. It not only enhances the system's local fault tolerance but also allows for maintaining directional stability and task continuity even in extreme cases of up to three rounds of failure. It is an adaptive limp-riding compensation mechanism specifically designed for agricultural unmanned platforms. The compensation weight I... hc The design features significant functional innovation, solving the problem that existing multi-wheel drive platforms cannot move stably under asymmetric fault scenarios, and has strong engineering value and patentability.

[0028] In step S610, the target output torque of the steering wheel determines its propulsion / obstacle avoidance capability; b1 and b2 are fitted from empirical data and used to control the system's sensitivity to terrain and drag; F1 and F2 are used to control the propagation efficiency of the compensation capability under varying distances; (1-tanh(b1×G) h +b2×(|H h -H0| / H0))) This implements terrain and motion state perception and adjustment. If the ground is soft and the angular velocity deviation is large, the torque output automatically decreases to prevent tire spin or slippage. If the ground is hard and the resistance is low, the torque is maintained at E0 to maintain propulsion; F1×(∑ d c=1 I hc ×J c ) / (1+F2×L hcAs a compensation term, it realizes fault-driven alternative output. The original torque of the failed steering wheel is shared by the adjacent wheel sets (prioritizing diagonal). The closer the wheel set, the more compensation it provides, and the stronger the attenuation at the far end. It can support the platform to maintain stable movement even if up to three drive wheels fail completely. If only two drive steering wheels remain, it can still execute the limp strategy to escape the fault area. Therefore, it has an asymmetric dynamic adjustment mechanism, that is, the torque response of different wheel sets changes dynamically due to terrain, resistance, and failure state; and it has a failure self-compensation mechanism, that is, automatic redistribution of torque without mechanical switching or manual intervention; through the fault-tolerant limp strategy, it breaks through the limitation of traditional four-wheel drive or six-wheel drive systems stopping immediately upon failure; and it integrates multi-source perception scheduling (soft and hard perception, angular velocity detection, distance modeling) to jointly drive torque scheduling.

[0029] Step S620: Adjust the output torque of each steering wheel to the target output torque corresponding to that steering wheel.

[0030] On the other hand, this application also proposes a method for adjusting the moving speed of a laser weeding device, used to automatically adjust the platform's traveling speed based on changes in the spatial density of weeds during the operation of an unmanned weeding platform, specifically including steps S710-S720: Step S710: During the movement of the laser weeding device, the target moving speed of the laser weeding device is determined based on the weed density in the area where the laser weeding device is located. Wherein, the target's moving speed R = R0 × exp(-p1 × q1) t1 )×(1-tanh(p2×q2)) t2 ; In the formula, R0 is the preset maximum speed of the laser weeding device, usually 0.6-1.2 m / s, depending on the terrain and operation mode; exp() is the preset natural exponential function; q1 is the weed density in the area where the laser weeding device is located at the current moment; q1 is the ratio of the weed pixel area in the image of the area where the laser weeding device is located at the current moment to the actual area of ​​the image field of view (the image area is captured by the forward-looking camera of the laser weeding device, semantic segmentation is performed to extract the weed target mask, the total area of ​​the weed target within the unit field of view is counted, and the weed density of the current area is calculated); t1 is the preset density response nonlinearity order; 1.5≤t1≤2.5; q2 is the density change rate per unit time (obtained by comparing the density of the current moment with the previous moment (or frame), and smoothing is performed using a sliding window to avoid short-term fluctuations and misadjustments), used to capture the density rising / falling trend and determine whether the laser weeding device is entering or leaving a high-density area; t2 is the preset trend response control index, which determines the magnitude of the trend change on the overall speed adjustment; 1≤t2≤3.

[0031] p1 is the density suppression index coefficient, used to control the direct impact of static weed density on speed. The larger the p1 value, the more significant the deceleration; p1 = p 10 +r1×V1+r2×V2+r3×V3;p 10 r1, r2, and r3 are preset initial density suppression index coefficients; r1, r2, and r3 are preset adjustment weight factors; V1 is the noise intensity of the image of the area where the laser weeding device is located at the current moment, used to reflect visual stability; V2 is the historical fluctuation variance of weed density in the area where the laser weeding device is located at a historical moment, and p1 is increased to slow down in advance when there is high fluctuation; V3 is the preset area safety level factor corresponding to the area where the laser weeding device is located at the current moment (such as in the field, field edge, or obstacle zone).

[0032] When visual recognition is stable and density is steady, the laser weeding device is allowed to operate at high speed, and the p1 value can be reduced. When the weed distribution is complex or the laser weeding device is entering a dense area, the p1 value automatically increases to improve the system's "sensitivity" to density, thereby entering the low-speed precision operation mode in advance.

[0033] p2 is the density change rate amplification factor, used to adjust the influence of q2 in the function to ensure timely system response; p2 = p 20 +r4×V4+r5×V5+r6×V6;p 20 R4 is the preset initial density change rate amplification factor; R5, R6 are preset response weight adjustment coefficients; V4 is the acceleration of the density change of weeds in the area where the laser weeding device is located at the current moment (i.e., the "rate of change" of the trend change), and the moving speed needs to be quickly adjusted when there is a sudden change; V5 is the preset density gradient direction consistency factor, which is used to determine whether the laser weeding device is approaching a high-density area; V6 is the predicted weed density at the next moment based on the image of the area where the laser weeding device is located at the current moment.

[0034] When density changes slowly, the system (i.e., the control system of the laser weeding device) maintains a stable response to avoid excessive jitter. When density changes abruptly (such as near a boundary or obstacle), p2 automatically increases, causing the speed to slow down immediately to prevent missed detection or false hits. This dynamic adjustment of p2 mechanism, from the perspective of "trend recognition + forward prediction", improves the sensitivity of speed control to complex ecological scenarios.

[0035] To enhance the adaptability of the platform's speed regulation system to complex agricultural ecological environments, this invention introduces a dynamic adjustment mechanism for the density suppression coefficient p1 and the trend amplification coefficient p2, establishing multi-factor control functions based on terrain, noise, regional attributes, and density change trends. p1 can be dynamically adjusted according to the visual stability of the current operating environment, weed fluctuations, and regional hazard levels, achieving intelligent correction of density suppression sensitivity. p2, on the other hand, combines density acceleration, gradient direction consistency, and forward density prediction to achieve forward-looking response control to density trend changes. This design overcomes the limitations of traditional static control parameters, giving the speed control system stronger stability, responsiveness, and environmental adaptability, significantly improving the platform's ability to balance high-speed operation and precision strikes.

[0036] In step S710, exp(-p1×q1) t1 (1-tanh(p²×q²)) indicates the influence of base density. The higher the density, the faster the rate of decay, ensuring slow processing in high-density areas. t2 This represents the trend suppression term. If q2 > 0 (i.e., density is increasing), tanh approaches 1, and the overall term approaches 0, indicating that the platform should decelerate significantly. If q2 < 0 (i.e., density is decreasing), tanh is negative, and (1-tanh(...)) approaches 2, then the speed will rebound and the travel efficiency will be restored. Through exponential and double nonlinear adjustment, the speed changes more naturally and stably under different density conditions, avoiding jitter and sudden changes.

[0037] Step S720: Adjust the current moving speed of the laser weeding device to the target moving speed.

[0038] After determining the target moving speed, the moving speed of the chassis of the laser weeding device is controlled and input into the speed controller module to adjust the current moving speed of the laser weeding device to the target moving speed.

[0039] On the other hand, in farmland environments, the lens of the image acquisition system of a laser weeding device is easily blocked by dust, flying insects, and moisture, causing recognition failure. Therefore, this application also proposes a self-cleaning method for the image acquisition system of a laser weeding device. By linking the air knife system with an image cleanliness scoring function, an image self-maintenance mechanism is constructed to maintain the reliability of visual perception. Specifically, it includes steps S810-S820: Step S810: During the movement of the laser weeding device, the visual quality index of the laser weeding device at the current moment is determined based on the image clarity acquired by the laser weeding device. Wherein, the visual quality index of the laser weeding device at the current moment is Q=1-exp(-s1×((1-s2 / s3)). 2 +u1×s4 2+u2×s5 2 )); In the formula, s1 is the preset pollution sensitivity amplification factor (controlling the steepness of the overall change in the scoring function), which depends on the system response speed requirements; 1≤s1≤10; s2 is the image sharpness acquired by the laser weeding device at the current moment (the variance value in the image gradient map is calculated by performing gradient transformation (such as Sobel filtering) on ​​the current image frame; the higher the variance, the clearer the edges and the less blurred the image); s3 is the preset system-calibrated normal reference value for sharpness, used for normalization to make different scenes comparable; u1 and u2 are the preset occlusion term and structural change term additions. The weighting coefficients determine the contribution ratio of the three factors to the overall pollution score; u1+u2≤1; s4 is the proportion of the polluted area in the image acquired by the laser weeding device at the current moment, calculated after extracting the pollution mask through a clustering segmentation algorithm; s5 is the structural difference between the image acquired by the laser weeding device at the current moment and the image acquired at a historical moment (by comparing the structural similarity with historical reference frames and extracting the structural difference, which can be obtained by structural tensor transformation or SSIM difference map). The higher the s5, the more severe the distortion of the current image, which may be caused by dirt occlusion or lens offset.

[0040] In step S810, 1-exp(-s1×((1-s2 / s3)) 2 Used to detect potential blurring issues caused by image sharpness degradation, with normalized squared amplification effect; u1×s4 2 Used to measure the amount of noticeable obstruction on the lens, such as flying insects, dust, or water droplets; u2×s5 2 Used to capture long-term structural changes, such as image shift distortion caused by static occlusion of blemishes; the closer the visual quality index is to 1, the more severe the image contamination; the closer the visual quality index is to 0, the clearer and more stable the image, used to determine whether to trigger a cleaning operation.

[0041] Step S820: If the visual quality index of the laser weeding device at the current moment is greater than or equal to the preset visual quality threshold, then control the cleaning system of the laser weeding device to clean the image acquisition system of the laser weeding device.

[0042] When the visual quality index of the laser weeding device is greater than or equal to the preset visual quality threshold (e.g., 0.6) at the current moment, the self-cleaning action of the image acquisition system of the laser weeding device is triggered. The air knife is controlled to blow the lens surface of the image acquisition system through the solenoid valve for a preset number of seconds (e.g., 0.2-0.5 seconds). If the visual quality index of a preset number of consecutive frames exceeds the visual quality threshold, the cleaning can be extended or water spray can be added to assist cleaning. After cleaning is completed, the acquired image is updated and the next round of visual quality index determination begins.

[0043] The self-cleaning method of the image acquisition system of the laser weeding device proposed in this application ensures that the front-view and rear-view cameras of the laser weeding device are always in an optically transparent state, ensuring recognition accuracy and the reliability of weeding decisions. In outdoor operating environments, camera lenses are easily contaminated by dust, insects, water droplets, etc., causing image blurring, obstruction, and decreased contrast, which can lead to recognition failure or accidental burning of crops in severe cases. Therefore, this application introduces a scoring system based on a visual quality index to automatically determine whether the lens is currently contaminated and to activate an air knife for physical cleaning. Unlike traditional methods based on pixel brightness fluctuations, this method integrates image sharpness analysis, occlusion ratio estimation, and structural similarity deviation with historical frames to construct a nonlinear contamination scoring function. This enables accurate identification and timely linkage. A three-dimensional contamination feature fusion model is used to comprehensively calculate the score based on sharpness, occlusion ratio, and historical structural offset, avoiding misjudgment based on a single indicator. Furthermore, an exponential scoring function design is used to achieve a nonlinear enhanced response to the contamination score. The more severe the contamination, the more timely the cleaning. Through an automated cleaning linkage mechanism, no manual intervention is required. The method combines visual perception and physical execution system in a self-closed-loop response, with low hardware resource consumption. Only the original camera image and a small air pump are needed to complete contamination detection and self-cleaning, without the need for complex sensor arrays.

[0044] On the other hand, laser weeding devices consume a lot of energy during weeding operations. If the electric-hydraulic drive system cannot be intelligently managed, the continuous operating time will be severely limited. (During automatic weeding operations, the demand for electricity and fuel changes drastically at different stages (such as continuous identification, laser strikes, and obstacle crossing). Without a unified energy scheduling mechanism, the drive system of the laser weeding device will either over-discharge or waste fuel, seriously affecting the endurance and operating efficiency. At the same time, energy switching (such as from electric drive to electric drive or hybrid mode) itself involves energy consumption and time delay.) Therefore, this application also proposes a method for adjusting the drive system of a laser weeding device, designing an energy resonance distribution control system, and optimizing energy use efficiency through an energy efficiency scoring function to extend the endurance. Specifically, it includes steps S911-S912: Step S911: When the laser weeding device is in working condition, determine the unit energy efficiency score of the laser weeding device when using the electric drive system, the unit energy efficiency score of the laser weeding device when using the fuel system, and the unit energy efficiency score of the laser weeding device when using the hybrid energy system, based on the system power requirements of the laser weeding device and the duration of the current task phase. Among them, the unit energy efficiency score of the laser weeding device using the electric drive system is W1=(M1×M2) / (y1×M2+M3); The energy efficiency score per unit of fuel system used by the laser weeding device is W2 = (M1 × M2) / (y2 × M2 + M4); The energy efficiency score of the laser weeding device using a hybrid energy system is W3 = (M1 × M2) / (y3 × M2 + M5); In the formula, M1 is the total system power requirement of the laser weeding device at the current moment (the sum of the real-time read power of the laser emitter, the drive system, and the vision and computing modules), which is a measure of the current load intensity of the entire platform; M2 is the expected duration of the laser weeding device in the current task phase, calculated by the scheduler predictor, such as the time required to travel a section or perform one round of laser cleaning; y1 is the energy consumption coefficient of the electric drive system of the laser weeding device, representing how much battery energy is consumed per second, including the internal losses of the electronic control system; M3 is the energy loss cost of the laser weeding device switching from the current drive system to the electric drive system, including the conversion delay, response inertia, and control signal loss during the switching process; y2 is the energy consumption coefficient of the fuel system of the laser weeding device, representing the energy conversion coefficient of the fuel-driven equipment per unit time; M4 is the energy loss cost of the laser weeding device switching from the current drive system to the fuel system; y3 is the energy consumption coefficient of the hybrid energy system of the laser weeding device; and M5 is the energy loss cost of the laser weeding device switching from the current drive system to the hybrid energy system.

[0045] In step S911, the unit energy efficiency score of the laser weeding device in the current task stage (the higher the value, the more economical the energy use) is used to determine whether the current energy mode is optimal or whether an energy switch is needed.

[0046] The battery discharge rate and discharge efficiency of the electric drive system are monitored by the electronic control system; the fuel flow and oil engine conversion efficiency of the fuel system are monitored by the oil drive controller; the energy switching costs (M3, M4, M5) are estimated by the scheduling system, including response delay and conversion loss.

[0047] Step S912: Schedule the drive system of the laser weeding device to the drive system corresponding to MAX(W1,W2,W3); where MAX() is a preset maximum value determination function.

[0048] The adjustment method of the drive system of the laser weeding device in this application dynamically calculates the unit energy efficiency score of each drive system by sampling the current task load, battery status and generator status. This is used to guide whether to use electric energy, fuel energy or hybrid energy supply in the current stage. The goal is to ensure the reliability of task execution while minimizing energy expenditure. In order to ensure the optimal energy efficiency and stability of multi-energy scheduling, this application designs a parallel scoring mechanism based on mode evaluation. By calculating the unit energy efficiency score in electric drive, fuel drive and hybrid drive modes respectively, the highest score is selected as the basis for energy mode selection in the current stage. Meanwhile, to avoid system oscillations and additional energy consumption caused by frequent switching, a switching penalty term (M3, M4, M5) is introduced into the scoring function. This value is dynamically estimated based on the switching path between the current energy mode and the candidate mode, and is included in the calculation of the unit energy efficiency score of the candidate mode, thereby constraining the energy consumption of the "switching behavior" itself. This structure achieves a balance between energy use efficiency and control stability, establishes a unified energy use evaluation system, integrates electricity, oil, and switching costs to form a single scoring function, quantifies the cost-effectiveness of system operation, considers the coupling of task continuity and energy efficiency ratio, not only considers the current energy consumption, but also considers the impact of task duration on energy efficiency, and adopts an adaptive energy switching strategy to dynamically switch energy sources based on the unit energy efficiency score, reducing reliance on manual configuration, building a system-level energy consumption perception architecture, establishing a closed-loop mechanism for energy scheduling and task execution, and effectively improving the operating endurance of the laser weeding device.

[0049] On the other hand, traditional laser weeding devices are prone to damaging crop seedlings when operating close to the edge, lacking an active avoidance strategy. Therefore, this application also proposes a method for correcting the travel path of a laser weeding device, proposing a seedling body pressure avoidance track corrector that integrates crop height, flexibility, and wheel pressure models to dynamically adjust the travel edge distance to avoid crop damage. This is used to prevent crushing damage to crops caused by improper paths during farmland operations, specifically including steps S921-S923: Step S921: During the movement of the laser weeding device, determine the required safety margin of the laser weeding device at the current moment based on the average height of the crops on both sides of the moving path of the laser weeding device at the current moment. Wherein, the safety margin required by the laser weeding device at the current moment is Y = Y0 × (1 + x1 × (Y1 / (Y2 × Y3))). x2 ); In the formula, Y0 is the preset basic safety margin, which is usually preset according to the width of the laser weeding device and the tire radius (e.g., 100-150 mm); x1 is the preset margin expansion factor, which determines the influence of the seedling crushing risk weight on the determined safety margin; 0.5≤x1≤2; Y2 is the average height of the crops on both sides of the laser weeding device's current movement path (using a forward-looking camera and depth sensor to detect the position coordinates and height of the crops on the left / right edges of the laser weeding device), reflecting the growth stage of the crops. Taller seedlings are more easily crushed, while shorter seedlings can have their margin appropriately reduced; Y3 is the preset crop leaf flexibility factor of the crops on both sides of the laser weeding device's current movement path, representing the deformation capacity that the crops can withstand under unit pressure, which can be set by crop type; x2 is the preset nonlinear expansion index, which controls the response curvature of the determined safety margin to Y1 and Y2; 1.2≤x2≤2.5.

[0050] Y1 is the ground pressure per unit area of ​​the laser weeding device; Y1=(Y4×Y5) / Y6; Y4 is the load mass of the laser weeding device at the current moment; Y5 is the gravitational acceleration; Y6 is the contact area of ​​the tires of the laser weeding device; the larger Y1 is, the higher the risk of crushing seedlings.

[0051] In step S921, the required safety margin for the laser weeding device at the current moment is the minimum distance that must be maintained between the original navigation path and the nearest crop edge to avoid crushing the seedlings. When the ground pressure is low and the crop is short (Y1 is small, Y2 is small), the safety margin Y is approximately equal to Y0, indicating that it is safe to travel close to the edge. When the crop height increases or the pressure rises, the safety margin Y increases, and the path planning will automatically deviate outward to prevent the wheels from encroaching on the seedling crushing boundary. Because the formula uses a power function form (Y1 / (Y2×Y3)), x2 It has the ability to "suppress and amplify" in high-pressure and high-seedling scenarios, enabling the system to automatically increase the pressure avoidance radius.

[0052] Step S922: Obtain the distance N between the center of the vehicle body of the laser weeding device at the current moment and the baseline of the crop that is closest to the lateral side of the vehicle body of the laser weeding device; The crop boundary line is obtained by detecting the forward / downward view camera and multispectral image of the laser weeding device. The platform self-positioning system of the laser weeding device (such as inertial measurement unit + encoder + SLAM real-time localization and mapping system) combined with the transformation matrix can project the vehicle body coordinates onto the crop boundary coordinate system to obtain the platform center projection position of the laser weeding device. The difference between the lateral crop boundary position and the platform center projection position is determined as the distance N between the vehicle body center of the laser weeding device at the current moment and the baseline of the crop that is closest to the lateral side of the laser weeding device.

[0053] Step S923: If N < Y, control the laser weeding device to move laterally so that the center coordinates of the laser weeding device are adjusted to the target correction coordinates. Wherein, the target correction coordinates U = U0 + (YN); U0 is the center coordinate of the vehicle body of the laser weeding device at the current moment.

[0054] If N≥Y, it means the laser weeding device does not need correction; if N<Y, it means the laser weeding device needs to be laterally offset. YN represents the lateral distance compensation amount that the laser weeding device needs to offset outward to ensure that the safety margin meets the standard. The target correction coordinate U is the corrected lateral target position of the laser weeding device, which is the "expected lateral center coordinate" corrected by the platform navigation system of the laser weeding device. It is used to guide the platform to offset towards the safe area. The target correction coordinate U will be passed as input to the path planner or motion controller to drive the platform to perform lateral fine-tuning or path reconstruction in real time, call the local path replanner to generate short-term obstacle avoidance trajectory (such as Z-shaped turn, reversing back), or request "standby reassessment state" to prevent seedling crushing accidents.

[0055] Since the laser weeding device adopts the "edge-following walking + visual guidance" mechanism, it is difficult to avoid walking close to crops during the process, especially in the seedling stage or at the edge of complex fields. If dynamic path correction is not performed, it is very easy to cause seedlings to be crushed, broken, or even uprooted. Therefore, this application uses crop target detection results, plant height estimation, tire pressure model, and seedling flexibility prediction model as inputs to construct a physical avoidance behavior control mechanism. In the path planning layer, a seedling crushing risk prediction function (i.e., the safety margin determination function in step S921) is introduced. The current movement trajectory is adjusted through a nonlinear safety margin expansion function to ensure that the laser weeding device avoids the seedling crushing risk area under the premise of walkability.

[0056] The driving path correction method of the laser weeding device in this application adopts biophysical perception modeling, introduces a cross model of seedling height, flexibility and contact pressure, and constructs a dynamic margin adjustment system based on "seedling crushing risk". It adopts a nonlinear expansion control function, which is different from the traditional fixed safety margin setting, to achieve environmental adaptation and intelligent self-correction of the path. It also integrates the collaborative mechanism of control and vision. The identification information is not only used for classification, but also for navigation control logic, forming a perception-behavior closed loop, which supports stable operation in complex boundary scenarios. It is suitable for high-risk seedling crushing areas such as field edges, dense planting belts, and curved plots, so as to improve the reliability of operation and crop protection rate.

[0057] The proposed method for correcting the travel path of a laser weeding device addresses a frequently overlooked problem in actual agricultural operations—the risk of physical conflict between the robot's path and seedlings. It provides a complete closed-loop logic from perception and modeling to behavior correction, demonstrating significant innovation. Traditional weeding platforms often employ a "visual recognition + fixed-margin edge-walking" strategy, which is tolerable under ideal field conditions or during crop growth. However, in tall seedling stages, on sloping land, slippery ground, or areas with dense seedlings, edge-walking can easily cause seedling damage, breakage, or even accidental crop injury.

[0058] In contrast, this invention introduces an intelligent path correction mechanism based on physical contact modeling and crop flexibility feature modeling. It combines real-time crop height, robot load, tire contact area and unit ground pressure to dynamically construct a crop pressure prediction model, and uses a nonlinear seedling damage risk scoring function Y=Y0×(1+x1×(Y1 / (Y2×Y3)). x2 The system can perform real-time expansion correction on the edge path to ensure that tall seedling areas are not trampled, balancing weeding accuracy and crop protection. In addition, the nonlinear power function structure (x2 control) in the algorithm enhances the avoidance effect in high-risk scenarios. That is, in the state of "high pressure + tall seedlings", the safety margin Y expands significantly, and the system automatically adjusts the path outward to effectively avoid physical damage. This mechanism also considers the actual output of navigation control (YN path correction) and the judgment of the operation boundary to ensure that the path deviation does not exceed the boundary and cause the operation to fail.

[0059] Compared with existing laser weeding technologies, this invention offers significant systemic improvements and functional expansions. Firstly, regarding target recognition and scorching strategies, existing technologies largely rely on RGB images for planar recognition. Recognition accuracy is greatly affected by factors such as lighting and background interference, making it difficult to distinguish between neighboring crops and weeds, leading to accidental damage or missed detections. In contrast, this invention employs multispectral image analysis and complexity control mechanisms, enabling not only weed identification but also automatic generation of scorching points and irradiation durations, achieving customized targeting. Secondly, most existing solutions use single-point laser control with fixed power and rigid paths, lacking dynamic power allocation and smooth path transition mechanisms, easily resulting in energy waste or intermittent targeting. This invention introduces a thermal beam trajectory control and dynamic focal length adjustment model, effectively improving laser utilization efficiency and coverage quality.

[0060] Furthermore, this invention surpasses traditional solutions in motion control, feedback loop, and system robustness. Existing technologies fail to effectively address issues such as platform drive failure and complex terrain changes, lacking steering wheel torque regulation and limp-out self-rescue mechanisms, while this invention possesses anti-balance drive capabilities. Traditional systems generally lack post-burn effect verification mechanisms, making it impossible to determine whether the attack was successful; this invention introduces a back-view feedback scoring model to achieve feedback-driven precision weed removal. Simultaneously, this invention adds an image contamination self-maintenance mechanism and an energy scheduling scoring strategy, which can significantly extend the system's continuous operating time and improve robustness. In summary, this invention far surpasses existing best-in-class technologies in recognition accuracy, control strategy, autonomy, and system intelligence, demonstrating substantial technological advancements at the system level.

[0061] Through the various control methods described above for the laser weeding device, key bottlenecks in traditional laser weeding equipment, such as recognition accuracy, strike control, terrain adaptation, energy consumption management, behavior feedback, and crop protection, have been solved. Each innovative module has independent technological creativity and can form a closed-loop collaborative mechanism in the system. It has outstanding substantive features and significant progress in terms of overall structure, control logic, and actual effectiveness.

[0062] The machine vision-based unmanned laser weeding method of this invention performs image analysis on the target area when the laser weeding device moves to the target area. Through multispectral visual perception and curvature boundary analysis algorithms, it identifies the weeds and several laser impact points within them, improving the accuracy of weed target segmentation and scorch point positioning. Then, based on the area of ​​the weeds, the complexity of the surrounding crop boundaries, and the laser emission distance, it determines the laser action duration for each laser impact point. Finally, based on the laser focusing distance and laser angle deflection amplitude corresponding to the trajectory segment between each pair of adjacent laser impact points, it determines the laser emission of the laser emitter of the laser weeding device for each pair of adjacent laser impact points. Power is used to precisely control laser beam parameters for different weed sizes, structures, and locations, dynamically adjusting the power output of each laser trajectory segment to achieve personalized strikes. After determining the laser beam parameters, the laser emitter is controlled to strike the weeds along a laser burning trajectory composed of several laser strike points with the laser action duration and laser emission power corresponding to each laser strike point. After the laser emitter finishes striking the weeds, the pixel area and average grayscale difference of the weeds before and after the laser strike are analyzed in real time based on the back-feeding thermal texture changes after weeding by introducing a backward thermal loss feedback fusion mechanism. This determines the laser weeding result of the weeds, enabling accurate verification of the weeding result and a secondary strike strategy, thereby improving the accuracy and success rate of weeding.

[0063] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine vision-based unmanned laser weeding method, characterized in that, include: Step S100: When the laser weeding device moves to the target area, perform image analysis on the target area to determine the weeds and several laser strike points in the weeds. Step S200: Determine the laser action duration corresponding to each laser strike point based on the area of ​​the weeds, the complexity of the crop boundary surrounding the weeds, and the laser emission distance; Step S300: Based on the laser focusing distance and laser angle deflection amplitude corresponding to the trajectory segment between each two adjacent laser strike points, determine the laser emission power of the laser emitter of the laser weeding device for the trajectory segment between each two adjacent laser strike points; Step S400: Control the laser emitter to perform laser strikes on the laser burning trajectory composed of several laser strike points with the laser action duration and laser emission power corresponding to each laser strike point; Step S500: After the laser emitter finishes laser-attacking the weeds, the laser weeding result is determined based on the pixel area and average grayscale difference of the corresponding images of the weeds before and after the laser-attacking. Wherein, step S100 includes steps S110-S140: Step S110: Obtain the target region image corresponding to the target region; Step S120: Perform image enhancement processing, illumination normalization processing, and edge detection processing on the target region image to obtain the processed image; Step S130: Extract the image of the weeds corresponding to the weeds from the processed image using a preset image segmentation algorithm, so as to determine the weeds corresponding to the image of the weeds from the target region; Step S140: If the area of ​​the weed body is less than or equal to a preset area threshold, then the geometric centroid corresponding to the weed body is determined as the laser strike point of the weed body; If the area of ​​the weeds is greater than a preset area threshold, then a number of laser strike points for the weeds are generated according to a preset edge equidistant sampling algorithm. Wherein, step S300 includes steps S310-S350: Step S310: Based on spatial correlation, sort the several laser strike points to obtain the rank of each laser strike point after sorting. Step S320: Based on the preset fifth-order B-spline interpolation algorithm, generate a laser burning trajectory composed of several laser impact points in ascending order of rank. Step S330: Discretize the laser burning trajectory to obtain several sub-trajectory points on the laser burning trajectory; on the laser burning trajectory, the distance between any two adjacent sub-trajectory points is equal; Step S340: Divide the laser burning trajectory into several sub-trajectory segments according to every two adjacent laser impact points on the laser burning trajectory; the two endpoints of each sub-trajectory segment are two laser impact points that are rank adjacent; Step S350: Determine the laser emission power P = e1 × ((∑) at any of the said sub-trajectory segments. g f=2 ||B f -B f-1 || 2 ) / (g×C1 2 ))×(1+e2×C2); Where e1 is a preset laser power reference adjustment factor; f=2,...,g; g is the number of sub-trajectory points included in the sub-trajectory segment; ||B f -B f-1 || represents the Euclidean distance between the f-th sub-trajectory point and the (f-1)-th sub-trajectory point in the sub-trajectory segment; C1 represents the laser focusing distance corresponding to the sub-trajectory segment; e2 represents the preset angle compensation amplification coefficient; and C2 represents the laser angle deflection amplitude of the laser emitter in the sub-trajectory segment.

2. The method according to claim 1, characterized in that, Step S200 includes: Step S210: Obtain the distance between each laser strike point and the laser emitter to obtain a emission distance list A=(A1,A2,...,Ai,...,Aj); where i=1,2,...,j; j is the number of laser strike points in the weed body; Ai is the distance between the i-th laser strike point and the laser emitter; Step S220: Determine the laser action duration T corresponding to the i-th laser strike point. i =(m1×log(1+S1)+m2×S2 1 / 2 )×(1+e (-Ai / n) ); Where m1 and m2 are preset energy time scheduling weighting coefficients; log() is a preset logarithmic function; S1 is the pixel area of ​​the weed; S2 is the boundary complexity obtained by performing a Gaussian curvature integral on the edge contour between the weed and its surrounding crops; n is a preset system thermal decay characteristic distance; e () This is a preset exponential function.

3. The method according to claim 2, characterized in that, Step S500 includes: Step S510: After the laser emitter finishes laser-attacking the weeds, determine the burning success score Z for the weeds: Z = (a1 × |D1∩D2| / |D2|) + a2 × (D3 / (1 + log(1 + ∑)) j i=1 T i ))); Where, a1 is the weighting coefficient for controlling the matching weight of the control region; a2 is the weighting coefficient for controlling the weight of image changes; 0≤a1≤1; 0≤a2≤1; a1+a2=1; D1 is the set of pixels in the thermally damaged area extracted from the image after laser attack on the herbaceous plant; D2 is the set of pixels in the image before laser attack on the herbaceous plant; |D1∩D2| is the overlapping area of ​​pixels in the thermally damaged area after laser attack on the herbaceous plant and the herbaceous plant before laser attack; |D2| is the pixel area in the image before laser attack on the herbaceous plant; D3 is the difference between the average pixel grayscale of the image after laser attack on the herbaceous plant and the average pixel grayscale of the image before laser attack on the herbaceous plant. Step S520: If the burn success score corresponding to the weed is greater than the preset score threshold, then the laser weeding result corresponding to the weed is determined to be successful; otherwise, return to step S200.

4. The method according to claim 3, characterized in that, Also includes: Step S610: During the movement of the laser weeding device, the target output torque of each steering wheel is determined based on the angular velocity of each steering wheel of the laser weeding device and the hardness of the terrain. Wherein, the target output torque E of the h-th steering wheel of the laser weeding device h =E0×(1-tanh(b1×G h +b2×(|H h -H0| / H0)))+F1×(∑ d c=1 I hc ×J c ) / (1+F2×L hc h = 1, 2, ..., k; k is the number of steering wheels in the laser weeding device; In the formula, E0 is the standard torque setting value of the rudder wheel of the laser weeding device on a stable ground; tanh() is the preset hyperbolic tangent function; b1 is the preset softness / hardness suppression weight; G h b1 represents the terrain hardness index at the current location of the h-th steering wheel; b2 represents the preset angular velocity offset modulation coefficient; H h H0 is the angular velocity of the h-th steering wheel at the current moment; F1 is the average angular velocity of all normally functioning steering wheels at the current moment; c = 1, 2, ..., d; d is the number of failed steering wheels at the current moment; I hc Let h be the compensation weight of the h-th steering wheel for the c-th failed steering wheel; 0≤I hc ≤1; J c F1 is the standard torque corresponding to the c-th failed steering wheel; F2 is the preset compensation attenuation index; L hc Let h be the physical center distance between the h-th steering wheel and the c-th failed steering wheel; Step S620: Adjust the output torque of each steering wheel to the target output torque corresponding to that steering wheel.

5. The method according to claim 4, characterized in that, Also includes: Step S710: During the movement of the laser weeding device, the target moving speed of the laser weeding device is determined according to the weed density in the area where the laser weeding device is located. Wherein, the target moving speed R = R0 × exp(-p1 × q1) t1 )×(1-tanh(p2×q2)) t2 ; In the formula, R0 is the preset maximum speed of the laser weeding device; exp() is the preset natural exponential function; q1 is the weed density in the area where the laser weeding device is located at the current moment; q1 is the ratio of the weed pixel area to the actual area of ​​the image field of view in the area where the laser weeding device is located at the current moment; t1 is the preset density response nonlinearity order; 1.5≤t1≤2.5; q2 is the rate of change of density per unit time; t2 is the preset trend response control index; 1≤t2≤3; p1 is the density suppression index coefficient; p1 = p 10 +r1×V1+r2×V2+r3×V3;p 10 r1, r2, and r3 are preset initial density suppression index coefficients; r1, r2, and r3 are preset adjustment weight factors; V1 is the noise intensity of the image of the area where the laser weeding device is located at the current moment; V2 is the historical fluctuation variance of the weed density in the area where the laser weeding device is located at a historical moment; V3 is the preset area security level factor corresponding to the area where the laser weeding device is located at the current moment. p2 is the density change rate amplification factor; p2 = p 20 +r4×V4+r5×V5+r6×V6;p 20 r4 is the preset initial density change rate amplification factor; r4, r5, and r6 are preset response weight adjustment coefficients; V4 is the acceleration of the density change of the weed density in the area where the laser weeding device is located at the current moment; V5 is the preset density gradient direction consistency factor; V6 is the predicted weed density at the next moment based on the image of the area where the laser weeding device is located at the current moment. Step S720: Adjust the current moving speed of the laser weeding device to the target moving speed.

6. The method according to claim 5, characterized in that, Also includes: Step S810: During the movement of the laser weeding device, the visual quality index of the laser weeding device at the current moment is determined based on the image clarity acquired by the laser weeding device. Wherein, the visual quality index Q of the laser weeding device at the current moment is Q=1-exp(-s1×((1-s2 / s3)). 2 +u1×s4 2 +u2×s5 2 )); In the formula, s1 is the preset pollution sensitivity amplification factor; 1≤s1≤10; s2 is the clarity of the image acquired by the laser weeding device at the current moment; s3 is the preset normal reference value for clarity calibrated by the system; u1 and u2 are the preset weighting coefficients for occlusion and structural change terms; u1+u2≤1; s4 is the proportion of the polluted area in the image acquired by the laser weeding device at the current moment; s5 is the structural difference between the image acquired by the laser weeding device at the current moment and the image acquired at a historical moment. Step S820: If the visual quality index of the laser weeding device at the current moment is greater than or equal to a preset visual quality threshold, then control the cleaning system of the laser weeding device to clean the image acquisition system of the laser weeding device.

7. The method according to claim 6, characterized in that, Also includes: Step S911: When the laser weeding device is in working condition, determine the unit energy efficiency score of the laser weeding device using the electric drive system, the unit energy efficiency score of the laser weeding device using the fuel system, and the unit energy efficiency score of the laser weeding device using the hybrid energy system based on the system power requirements of the laser weeding device and the duration of the current task phase. Wherein, the unit energy efficiency score of the laser weeding device using the electric drive system is W1=(M1×M2) / (y1×M2+M3); The laser weeding device uses a fuel system with a unit energy efficiency score of W2=(M1×M2) / (y2×M2+M4); The laser weeding device uses a hybrid energy system with a unit energy efficiency score of W3 = (M1 × M2) / (y3 × M2 + M5); In the formula, M1 is the total system power requirement of the laser weeding device at the current moment; M2 is the expected duration of the laser weeding device in the current task phase; y1 is the energy consumption coefficient of the electric drive system of the laser weeding device; M3 is the energy loss cost of the laser weeding device switching from the current drive system to the electric drive system; y2 is the energy consumption coefficient of the fuel system of the laser weeding device; M4 is the energy loss cost of the laser weeding device switching from the current drive system to the fuel system; y3 is the energy consumption coefficient of the hybrid energy system of the laser weeding device; and M5 is the energy loss cost of the laser weeding device switching from the current drive system to the hybrid energy system. Step S912: Schedule the drive system of the laser weeding device to the drive system corresponding to MAX(W1,W2,W3); where MAX() is a preset maximum value determination function.

8. The method according to claim 7, characterized in that, Also includes: Step S921: During the movement of the laser weeding device, the required safety margin of the laser weeding device at the current moment is determined based on the average height of the crops on both sides of the moving path of the laser weeding device at the current moment. Wherein, the safety margin required by the laser weeding device at the current moment is Y = Y0 × (1 + x1 × (Y1 / (Y2 × Y3))). x2 ); In the formula, Y0 is the preset basic safety margin; x1 is the preset margin expansion factor; 0.5≤x1≤2; Y2 is the average height of crops on both sides of the moving path of the laser weeding device at the current moment; Y3 is the preset crop leaf flexibility factor of crops on both sides of the moving path of the laser weeding device at the current moment; x2 is the preset nonlinear expansion index; 1.2≤x2≤2.5; Y1 is the ground pressure per unit area of ​​the laser weeding device; Y1=(Y4×Y5) / Y6; Y4 is the load mass of the laser weeding device at the current moment; Y5 is the gravitational acceleration; Y6 is the contact area of ​​the tires of the laser weeding device. Step S922: Obtain the distance N between the center of the vehicle body of the laser weeding device at the current moment and the baseline of the crop that is closest to the vehicle body of the laser weeding device laterally; Step S923: If N < Y, then control the laser weeding device to move laterally so that the center coordinates of the laser weeding device are adjusted to the target correction coordinates. Wherein, the target correction coordinates U = U0 + (YN); U0 is the center coordinate of the vehicle body of the laser weeding device at the current moment.

Images