Land boundary recognition method based on fusion of laser point cloud and visible light

Abstract

The present invention relates to the field of image data processing technology, and discloses a method for identifying land boundaries by fusing laser point clouds with visible light. The method comprises: processing laser point clouds and visible light image data streams in parallel, detecting geometric elevation mutations and visual texture transitions respectively to extract modality-specific change events and their timestamps; identifying cross-modal co-occurrence instances based on a temporal weak synchronization window, and marking the spatial positions of associated events as boundary candidate points; and forming continuous boundaries through spatiotemporal filtering. The present invention fundamentally eliminates the reliance on precise spatial alignment of sensors through a coupling mechanism of independent modal change perception and temporal co-occurrence determination, thereby significantly improving the robustness and environmental adaptability of boundary recognition in complex operating environments while ensuring the high efficiency of edge computing.

Description

Technical Field

[0001] The present invention relates to a land boundary recognition method by fusing laser point cloud and visible light, belonging to the technical field of image data processing. Background Art

[0002] In the field of image data processing technology, especially in the direction of land boundary recognition based on the fusion of laser point cloud and visible light, the existing mainstream technology generally adopts a cross-modal feature alignment and fusion method, which usually includes the following core processes: first, the laser point cloud and visible light image are accurately aligned in time and space, then the geometric features and visual features are extracted, and finally the boundary is determined through feature matching or deep learning models.

[0003] Such methods can achieve a certain level of accuracy under ideal conditions. However, when applied to real-world dynamic scenarios such as agricultural inspections and topographic mapping, their inherent mechanisms expose systemic bottlenecks. Specifically, existing technologies suffer from three fundamental limitations: 1. Sensor vibration, temperature drift, and sampling timing differences lead to misalignment in cross-modal spatial registration. The subsequent fusion process relies heavily on these intentionally constructed alignment assumptions. Once the registration deviation exceeds a threshold, the matching relationship between geometric and visual features fails, and boundary recognition results deteriorate drastically. 2. To improve cross-modal feature comparability, existing methods require the construction of high-dimensional feature descriptors or the training of complex fusion models. This not only requires high-performance computing units, but also the combined computing power of feature engineering and model inference effectively hinders their application in resource-constrained scenarios such as agricultural edge devices. 3. Fixed thresholds or statically trained models make it difficult to adaptively adjust perception strategies in response to sudden changes in illumination, gradual changes in vegetation, or occlusion. For example, when image texture features degrade under strong light, the system cannot autonomously suppress the weight of the visible light modality and mechanically performs low-confidence matching, resulting in a sharp increase in boundary misjudgment rate.

[0004] While the industry has attempted to alleviate these issues by simplifying feature extraction or optimizing alignment algorithms, these improvements have always remained within the inherent framework of alignment-first, fusion-later. These optimizations can only reduce computational overhead to a limited extent, but cannot fundamentally address the cascading failures caused by misalignment. Furthermore, these simplified operations sacrifice the ability to discern complex boundaries. Therefore, the technical challenge addressed by this present invention is to construct a multimodal boundary recognition mechanism that avoids the need for precise spatial alignment, combines computational lightness with environmental adaptability, and is a promising approach. Summary of the Invention

[0005] The present invention provides a land boundary recognition method by fusing laser point cloud and visible light. Its main purpose is to solve the problems of insufficient robustness, high computational complexity and poor environmental adaptability caused by the existing fusion technology due to its reliance on cross-modal precise alignment.

[0006] To achieve the above objectives, the present invention provides a land boundary recognition method by fusing laser point cloud and visible light, the method comprising the following steps:

[0007] Step a, independently and in parallel processing the first modal data stream and the second modal data stream, including: step a1, detecting, from the first modal data stream, i.e., the laser point cloud data stream, the elevation difference between the current laser point in the laser point cloud and its adjacent laser points in the scanning sequence, or detecting the point cloud density gradient within the local neighborhood of the current laser point based on the local spatial instantaneous difference of the surface geometric elevation or point cloud density, and comparing the detection result with a first threshold to extract the laser modal geometric change event and its first timestamp; step a2, detecting, from the second modal data stream, i.e., the visible light image data stream, the absolute value of the grayscale value or color component difference between a pixel point on a predetermined line of interest in the visible light image and its one-dimensional adjacent pixels based on the local spatial instantaneous transition of the image texture, color or brightness, and comparing the detection result with a second threshold to extract the visible light modal visual change event and its second timestamp;

[0008] Step b, based on the first timestamp and the second timestamp, identifying a co-occurrence instance including both a laser modal geometry change event and a visible light modal visual change event within a predetermined time window;

[0009] Step c: identifying the spatial location of at least one event associated with the co-occurrence instance as a candidate land parcel boundary point;

[0010] Step d: performing spatiotemporal filtering on the land parcel boundary candidate point sequence to form the land parcel boundary.

[0011] Preferably, the first threshold value in step a1 is determined by multiplying the elevation difference or the statistical standard deviation of the point cloud density gradient of the recently collected laser point cloud data by a predetermined adjustment coefficient to obtain the first threshold value. ,in is the first predetermined adjustment coefficient, is the first statistical standard deviation.

[0012] Preferably, the second threshold value in step a2 is determined by multiplying the statistical standard deviation of the absolute value of the grayscale value or color component difference of the recently collected visible light image by a predetermined adjustment coefficient to obtain the second threshold value. ,in is the second predetermined adjustment coefficient, is the second statistical standard deviation.

[0013] Preferably, the spatiotemporal filtering processing in step d includes: clustering and connecting the plot boundary candidate points based on the proximity of the plot boundary candidate points in the estimated spatial position to form continuous boundary segments; and removing isolated candidate points that are not associated with the remaining candidate points in time or space.

[0014] Preferably, the method further includes: detecting the movement speed of the sensing platform that collects the laser point cloud data stream and the visible light image data stream; and when the movement speed of the sensing platform is lower than a predetermined speed threshold, adjusting the extraction of change events of at least one modality in step a, the adjustment including: extracting change events based on the local spatial data acquired by the modal sensor at the current moment by analyzing the structural contrast or visual feature activity in the local spatial data.

[0015] Preferably, the predetermined time window in step b is set by dynamically adjusting the online statistical characteristics of the timestamp differences actually observed between the laser modal geometric change events and the visible light modal visual change events corresponding to the co-existing instances that have recently been identified as candidate plot boundary points or have been output as final plot boundaries.

[0016] Preferably, the first threshold in step a1 and the second threshold in step a2 are both adjusted periodically or non-periodically based on the statistical characteristics of recent data of their respective modalities to adapt to changes in environmental factors such as lighting conditions, surface cover or vegetation conditions.

[0017] Preferably, the spatiotemporal filtering processing in step d also includes: based on the motion information of the sensing platform, performing consistency verification on the motion trajectory of the candidate points at the boundary of the plot and the motion trajectory of the sensing platform to identify and eliminate candidate points that are inconsistent with the motion trajectory of the sensing platform.

[0018] Preferably, the method for identifying the spatial position of the candidate points of the land boundary in step c includes: based on a predetermined approximate relative position relationship between the laser point cloud sensor and the visible light image sensor, approximately projecting the spatial position of the events associated with the inter-modal event pair.

[0019] Preferably, the method further includes: when the sensing platform moves and scans multiple adjacent plot areas, spatially splicing and fusing the plot boundary sequences identified in different time periods to form continuous plot boundary data of a larger range; the splicing and fusion are based on the spatial continuity characteristics of the plot boundaries, and combined with the positioning information of the sensing platform, by performing feature matching and consistency verification on the overlapping areas of the boundary fragments.

[0020] Compared with the background technology problems, the beneficial effects of the present invention are:

[0021] 1. In the parallel processing of laser point clouds and visible light data, change events are independently extracted based on local geometric elevation mutations and visual texture transitions, and the associated physical boundaries are then determined by co-occurrence within a preset time window. This design allows the system to rely only on the temporal approximation of events rather than precise spatial matching, effectively avoiding the fusion failure problems caused by sensor vibration, temperature drift, or sampling timing differences in traditional methods. When environmental disturbances cause spatial registration deviations, the system can still stably identify boundaries based on temporal consensus, significantly improving robustness in complex operating scenarios.

[0022] 2. A dynamic threshold adjustment mechanism based on recent data statistics is employed for the elevation differential / density gradient of laser point clouds and the grayscale / color gradient of visible light. This mechanism automatically maintains the sensitivity of event detection by sensing changes in ambient lighting and vegetation cover in real time, eliminating the burden of manual parameter adjustment. Event extraction requires only basic arithmetic operations, while co-occurrence determination relies on simple timestamp comparisons. Spatiotemporal filtering employs lightweight operations such as proximity clustering and outlier removal. The synergistic effect of these multiple steps enables the entire method to run in real time on conventional microcontrollers, providing accessible technical support for edge scenarios such as agricultural inspections. Based on the co-occurrence instances identified in step b, the timestamp differences between the two types of events are calculated online, and the window center position and width are dynamically adjusted accordingly. This mechanism autonomously adapts to sensor delay drift, boundary gradient characteristics, and hardware timing jitter, ensuring that the co-occurrence criteria always match the actual system state. When faced with blurred boundaries or environmental interference, the dynamically calibrated window parameters work in synergy with subsequent spatiotemporal filtering to form a dual fault-tolerant system, ensuring the spatial continuity and physical plausibility of boundary trajectories.

[0023] 3. When the sensing platform is moving at low speed or hovering, the event extraction logic is automatically switched through inertial units or image displacement analysis: the time-series differential detection of the laser point cloud is converted into local spatial structure contrast analysis, and the line scanning gradient detection of the visible light image is converted into two-dimensional texture activity assessment. This motion-state-based event definition conversion enables the system to capture subtle boundary features within the field of view during static observation, complementing the dynamic scanning mode. There is no need to add complex computing modules to the entire process. Only through adaptive fine-tuning of the core processing logic can a seamless transition from rapid inspection to fixed-point precision inspection be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a diagram showing the key parameters for boundary recognition using laser point cloud and visible light fusion.

[0025] Figure 2 This is a flow chart of the core processing chain for laser point cloud and visible light fusion boundary recognition in the present invention;

[0026] Figure 3 This is a multi-module collaborative processing sequence diagram for candidate point space screening and boundary optimization of the present invention.

[0027] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0028] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] The present invention discloses a land boundary recognition method by fusing laser point cloud and visible light. By perceiving the independent modal changes of laser point cloud and visible light image, combined with the event fusion mechanism based on temporal co-occurrence, the boundary position corresponding to the coupled features is identified, and the boundary coherence is enhanced with the help of spatiotemporal filtering method. The whole process breaks away from the rigid mode of traditional alignment-fusion-judgment, and forms a fusion recognition strategy with fault tolerance and high adaptability. Specifically, the first step is the independent parallel processing of modal data streams: in the laser point cloud data stream, the system detects the elevation difference of each laser point with its adjacent laser points in the vertical direction based on the continuous scanning sequence of spatial position, or analyzes the elevation difference of the point in its local neighborhood (which can be set to a fixed radius or These local differences are used to identify geometric mutation areas on the surface, such as steep slopes, ridge edges, and other ground feature turning zones. If the detection value exceeds the set threshold , that is, it is determined to be a geometric change event of the laser mode, and its spatial position and the first timestamp are recorded. To ensure the sensitivity and robustness of the detection, the threshold is not fixed, but is the first statistical standard deviation of the difference or gradient value of the recent point cloud data. After the first adjustment coefficient The specific form of the correction is: ,in The recommended value is generally between 1.5 and 3. The specific value can be adjusted empirically based on the degree of surface undulation in the operation scene. In parallel, the data stream processing of the visible light modality is based on the preset scan line or area of ​​interest of the image, and the absolute value of the difference between adjacent pixels of the grayscale or color channel is used as the detection indicator of texture change. When the difference exceeds the visual change threshold, (Similar statistical mechanism is also used for dynamic setting, ,in is the standard deviation of texture differences sampled recently), that is, a visual change event is considered to have occurred, and its pixel position and the second timestamp are recorded. The change events of the two modes will be co-occurred in a dynamically adjusted time sequence weak synchronization window. Specifically, the system sets a width of The time window is used to determine whether the event contains both the laser mode and the visible light mode. Once this condition is met, it is considered an instance of cross-modal co-reality.

[0030] In order to further map the co-occurrence instances to the spatial points of the land boundary, the system performs spatial projection mapping between event pairs based on the preset relative position relationship between the laser sensor and the camera in the sensing platform. Without precise calibration, the approximate transformation (such as affine model or rigid transformation estimation) can be used to spatially overlap the event positions corresponding to the two modal data, thereby obtaining the three-dimensional coordinates of the boundary candidate points; for the obtained candidate point sequence, the system introduces a lightweight spatiotemporal filtering method to improve the continuity and stability of the boundary. This method first performs proximity clustering in the spatial dimension, and merges candidate points with similar distances to form boundary segments; then, by analyzing the sparse distribution of candidate points in the time dimension, those that are not aligned with the boundary are removed. Other points form time-correlated isolated points. In addition, considering the inertial information or GPS trajectory data of the sensing platform, the spatial motion direction of the candidate points can be further checked for consistency with the platform trajectory to filter out false boundary responses caused by platform vibration or sensor anomalies. When the sensing platform is in a low-speed movement or short-stop state, the system will automatically switch the event detection logic mode. For example, the event judgment of the laser mode is changed from being based on time series difference to analyzing the structural contrast of the local point cloud; the visible light mode is based on the texture activity (such as local variance, Gabor response) in the two-dimensional image window to identify boundary clues. This type of design ensures that under non-uniform scanning, the system can still obtain sufficient change events for subsequent fusion judgment.

[0031] In addition, the temporal co-occurrence window It is not a static setting, but a dynamic self-calibration of the statistical distribution of the time difference between the two modes based on the co-occurrence examples successfully identified in the past. For example, if the statistical median shows that the time difference between the two types of events is concentrated around 300ms, the window center will slide dynamically based on this value, and the window width will be set in combination with its variance to maximize the co-occurrence coverage of true boundary events while reducing the risk of misjudgment. For the boundary results output by the entire recognition process, if repeated scanning is performed in multiple adjacent land areas, the system also supports spatial splicing and fusion between boundary fragments to form a wide-area continuous boundary. This fusion process refers to the geometric overlap and texture consistency between the reference fragments, and uses the platform positioning information for registration and correction to ensure that the final boundary results are scalable and integrable. In specific application scenarios, the boundary recognition mechanism proposed in the present invention can adapt to lightweight mobile platforms with spatial alignment errors according to different system configurations, and can also be extended to high-precision multimodal acquisition systems with synchronous calibration capabilities. For the latter, in order to further improve the precision of boundary position recognition, the system can introduce in some implementations. Initializing the relative pose of the input sensors enhances the stability of event-space mapping. This approach serves as a compatibility expansion strategy, providing adaptive paths for different operational requirements and more effectively supporting dynamic control of temporal correlations during event matching and candidate point screening. The system's time window is not fixed but flexibly slides based on the state of real-time acquired data while maintaining temporal continuity to cover more potential co-occurrence information while mitigating the risk of missing boundaries due to data drift or synchronization delays. Furthermore, the criteria for determining changes in image texture and point cloud structural features in this method are primarily based on grayscale fluctuations within local regions and the density of spatial point distribution. While complex model calculations are not involved, both can be approximated using the statistical variation of the sampling area, providing a clear basis for perception. Furthermore, considering the resource constraints of edge computing environments, the system's perception strategy utilizes a lightweight, embeddable implementation, ensuring that the processing logic remains efficient while still being controllable. These are all extended implementations known to those skilled in the art.

[0032] Example 1: This example, combined with typical agricultural inspection application scenarios, elaborates on the specific implementation process, parameter setting basis and internal logical mechanism of the key steps of the land boundary recognition method of the fusion of laser point cloud and visible light, aiming to ensure that various technical elements are traceable, operable and verifiable, so as to meet the requirements of technical personnel in this field for implementation; first, relying on a multimodal mobile acquisition platform equipped with a lidar sensor and a visible light camera, the laser point cloud data stream and the visible light image data stream are collected respectively. In the data acquisition stage, the laser point cloud data records the three-dimensional coordinates and reflection intensity information of each laser point at a fixed frequency, and the visible light image records the brightness and color component values ​​of each pixel in the image in frames. For the laser point cloud mode, the system calculates the absolute value of the elevation difference of adjacent points based on the vertical elevation values ​​of adjacent laser points in the laser point cloud scanning sequence, and compares it with the first threshold value obtained by real-time calculation. The first threshold value is determined by multiplying the standard deviation of the elevation difference of recent laser point cloud data in the current operation plot by a preset adjustment coefficient. The coefficient is typically set between 1.5 and 3, with the specific value empirically determined based on the degree of land undulation and equipment vibration levels. For plain farmland scenarios, the value is typically 2, and can be appropriately increased in complex terrain areas such as mountains or orchards. When the elevation difference exceeds this threshold, a laser modal geometry change event is identified, and the event's 3D spatial coordinates and occurrence time are recorded. For visible light image modalities, the system selects a preset scan line or region of interest within the image and sequentially calculates the absolute value of the grayscale value difference or color channel brightness difference between adjacent pixels as a visual change signature. Similarly, the system calculates the standard deviation of the adjacent pixel differences within the image region in real time and multiplies it by a preset adjustment coefficient to obtain a second threshold value, typically set between 1.2 and 2.5. The specific value is determined based on the uniformity of illumination and the complexity of vegetation types on the day of operation. A value of 1.5 is recommended for uniform illumination on sunny days and can be appropriately increased on cloudy days or under conditions of strong and uneven illumination. If the adjacent pixel difference exceeds this threshold, a visible light modal visual change event is identified, and its pixel location and event time are recorded.

[0033] Subsequently, the system sets a bidirectional timing window with the current sampling time point as the center. The window width is dynamically adjusted based on the time difference distribution of the two types of modal events with confirmed co-reality instances in the past five minutes. Specifically, the median of the time difference distribution is taken as the center position of the window, and the window width is determined based on twice the standard deviation of the time difference. If there are both laser modal geometric change events and visible light modal visual change events in the window, it is identified as a cross-modal co-reality instance, and the spatial positions of the two types of events are extracted. In view of the fact that there is a fixed installation position offset between the laser sensor and the camera, but no high-precision calibration is required, this embodiment uses an affine transformation model to approximately correct the spatial positions of the two modalities based on the approximate relative position parameters recorded when the equipment is installed. The transformation parameters are determined based on the equipment structure drawing and the sensor installation height difference. In a typical parallel forward arrangement scenario, the horizontal The offset is set to 0.2 meters, the vertical offset is set to 0.15 meters, and the rotation angle is set to zero degrees. They can be adjusted appropriately according to the actual equipment loading situation. After obtaining the candidate point sequence, the system performs spatial proximity clustering and temporal isolated point removal operations in sequence. The spatial clustering process adopts a fixed distance threshold clustering method to classify candidate points with a horizontal distance of less than 0.5 meters, a vertical distance of less than 0.3 meters, and a time interval of no more than 0.8 seconds as the same plot boundary fragment. The threshold is determined based on the actual width of the ridge boundary and the sampling accuracy of the equipment. It is recommended to be set to 0.5 meters in plain areas, and can be appropriately relaxed to 0.7 meters in complex working environments such as sloping orchards. The isolated point removal process counts the number of neighboring points of each candidate point in the time window. If there are less than 2, it is considered an occasional false detection or noise response and is deleted, thereby ensuring the spatial continuity of the final boundary result.

[0034] In addition, based on the device motion trajectory information provided by the inertial measurement unit, the motion trajectory direction of the candidate point is checked for consistency with the device's trajectory, and the angle between the two directions is calculated. If the angle is greater than 30 degrees, it is considered a non-real boundary signal and is eliminated. The angle threshold is set according to the platform's turning radius and operating speed. It is recommended to set it to 30 degrees in conventional operating scenarios, and it can be appropriately relaxed to 45 degrees in high-speed travel or frequent regional turning scenarios; when the device is in a low-speed movement (less than 0.5 meters per second) or a short stop state, the system will automatically switch the event detection logic mode, and the extraction method of the laser mode change event will be based on time. Sequential difference detection is converted into a detection method based on the gradient change of point cloud density in the local neighborhood of the laser point. Specifically, the rate of change of the number of points per unit area in the neighborhood of the laser point is calculated. If it exceeds the dynamic threshold, it is determined as a change event. The threshold is determined by multiplying the standard deviation of the point cloud density in the current neighborhood by 1.8 and is updated dynamically in real time. The visible light modality converts the adjacent pixel difference detection method into an extraction based on the texture activity index in the local image window, and uses the grayscale variance value as the texture change feature. If it exceeds the dynamic threshold, it is determined as a visual change event. The threshold is also determined by multiplying the local variance standard deviation of the current sampled image by 1.5. Finally, all candidate point sequences after filtering, clustering and consistency verification are sorted in time and space order, and adjacent points are connected in sequence to form continuous plot boundary lines. In the scenario of continuous scanning of multiple plots, the system will automatically splice and fuse the plot boundary fragments of different time periods and adjacent spatial positions based on the standards of the overlap of the beginning and end of the fragments being not less than 70% and the texture similarity being higher than 80%. The overall position of the boundary is corrected in combination with the positioning trajectory information of the equipment to ensure that the final generated wide-area continuous plot boundary has spatial coherence and consistency.

[0035] Example 2: In a typical farmland inspection scenario, the plot boundary recognition method that integrates laser point cloud and visible light imagery must ensure good spatial continuity and temporal stability in dynamic environments, complex surface morphology, and variable lighting conditions. To this end, this experiment was conducted to verify the applicability and robustness of this method under different terrain, lighting conditions, and operating speeds through actual field data collection and comparative analysis. The focus was on testing the effectiveness of the dynamic threshold adjustment strategy, the dynamic calibration mechanism of the temporal co-occurrence window, and the boundary segment splicing strategy. The experimental area was an agricultural demonstration area in Hebei Province, and four representative typical farmlands were selected, including plains, sloping land, ditch-interlaced areas, and weed-obstructed areas. The equipment platform was: Velodyne VLP-16 lidar with a point frequency of 10 Hz and a ranging error of ±2 cm; Sony IMX-290 visible light camera with a frame rate of 30 fps and a resolution of 1920×1080; Xsens MTi-300 inertial measurement unit; and Trimble GNSS positioning system. BD990, positioning error ±5cm; platform operating speed is set to three speed levels: 0.3m / s, 0.6m / s, and 1.0m / s.

[0036] The design and parameter setting of this experiment are based on the laser point cloud modal elevation difference threshold , calculate the standard deviation of local elevation difference based on the 30-second point cloud scanning data in each block , then multiplied by the adjustment coefficient OK, plain scene The value is 2.0, 2.5 for slopes, and 3.0 for occluded areas; the visible light image modal texture difference threshold 𝜃𝑉 is the standard deviation of the grayscale difference between adjacent pixels , then multiplied by the adjustment coefficient OK, the value is 1.5 under sunny conditions and 2.0 under cloudy conditions; co-occurrence time window , initially set to 0.5 seconds, and based on the 50 sets of co-occurrence instances successfully identified in the first 5 minutes, calculate the median of the time difference between the laser mode and the visible light mode events and variance , dynamically adjust the window range to The spatial clustering threshold is set to 0.5 meters horizontally and 0.3 meters vertically, which is determined comprehensively based on the width of the ridge and the sampling density of the equipment. The threshold for isolated point removal is to remove points if the number of neighboring points within the specified time window is less than 2. This experiment carried out 5 repeated measurements in four typical scenes, each lasting 8 minutes. The laser point cloud elevation difference change events, visible light image texture change events and their timestamps were recorded in real time. The dynamic threshold adjustment, dynamic correction of the co-occurrence time window, candidate point spatial projection mapping, spatial proximity clustering, isolated point removal, consistency trajectory verification and fragment splicing were performed in sequence to finally generate a continuous plot boundary line. The number of laser elevation difference events and visible light texture change events is statistically as follows:

[0037]

[0038] Under clear conditions on plain land, the standard deviation of the laser point cloud elevation difference within the first 5 minutes The initial threshold is 0.032m. , as the height of the grass stubble changes, the threshold gradually increases to 0.071m; the standard deviation of the image grayscale difference From 5.6 to 6.3, the corresponding threshold Dynamically adjusted from 8.4 to 9.45, the success rate of border segment splicing is:

[0039]

[0040] This experiment verified that the method of the present invention has good environmental adaptability and boundary recognition robustness in typical farmland operation environments. The dynamic threshold adjustment mechanism effectively adapts to lighting and terrain disturbances. The dynamic calibration mechanism of the temporal co-occurrence window ensures the reliability of event synchronization judgment. The spatial clustering and consistency verification strategy significantly improves the spatial continuity and physical rationality of the boundary line. In the four scenarios, the success rate of boundary fragment splicing exceeds 70%, meeting the application requirements of agricultural inspection and land surveying for boundary continuity. In addition, the experiment also found that in areas with severe weed obstruction, the number of event recognition decreased slightly, but the stability of boundary feature extraction was still guaranteed by switching the event extraction strategy in low-speed mode.

[0041] Example 3: This example combines Figures 1 to 3 , the implementation of the land boundary recognition method based on the fusion of laser point cloud and visible light is explained. Figure 1As shown in the figure, the setting values ​​of key parameters in the land boundary recognition method of the laser point cloud and visible light fusion of the present invention are shown. The horizontal axis is the parameter threshold, which is expressed in logarithmic coordinates. The vertical axis lists the important technical parameters involved in the extraction and screening of boundary candidate points, including low-speed movement judgment threshold, spatial clustering: horizontal distance, spatial clustering: vertical distance, spatial clustering: time interval, visible light modal adjustment coefficient (KV), laser modal adjustment coefficient (KL) and trajectory consistency verification angle. Each parameter is marked with its recommended setting value with a corresponding numerical value. Among them, the low-speed movement judgment threshold is set to 0.5 m / s, which is suitable for judging whether the sensing platform is in a low-speed or stationary state, thereby switching event extraction. Logic; spatial clustering: the horizontal distance is set to 0.5 meters, spatial clustering: the vertical distance is set to 0.3 meters, and the spatial clustering: the time interval is set to 0.8 seconds, which is used to determine the proximity between candidate points and thus form boundary segments; the visible light modal adjustment coefficient (KV) is set to 1.5, which is used to dynamically set the detection threshold of visual change events based on the standard deviation of image grayscale or color channel differences; the laser modal adjustment coefficient (KL) is set to 2, which is used to determine the detection threshold of laser modal geometric change events based on the standard deviation of point cloud elevation difference or density gradient; the trajectory consistency check angle is set to 30 degrees, which is used to determine the consistency between the candidate point movement direction and the platform travel direction, so as to eliminate misidentified non-real boundary signals.

[0042] like Figure 2 The figure shows a complete process, starting with the input of a first-modality laser point cloud data stream and a second-modality visible light image data stream, and sequentially going through step a, independent change detection, step b, temporal co-occurrence determination, and steps c / d, candidate point identification and spatiotemporal filtering, to ultimately output a continuous land parcel boundary (the final result). In the figure, step a, independent change detection, includes extracting geometric change events and visual change events, corresponding to the detection of elevation differences or point cloud density gradients in the laser point cloud modality and the detection of grayscale or color differences in the visible light modality, respectively. Both independently perform modality-specific change event extraction. Step b, temporal co-occurrence determination, identifies co-occurring instances of the two types of change events using a preset temporal weak synchronization window, thereby establishing an inter-modality association mechanism. Step c / d, candidate point identification and spatiotemporal filtering, marks the spatial locations of co-occurring instances as boundary candidates, and performs filtering processes such as spatial clustering, outlier removal, and trajectory consistency verification to ensure the stability and coherence of the output boundary. Ultimately, a continuous land parcel boundary (the final result) is obtained, which has good spatial integrity and environmental adaptability.

[0043] like Figure 3As shown, the candidate point manager first submits a sequence of plot boundary candidate points to the spatial clustering module, which returns the resulting continuous boundary segments. The candidate point manager then submits the continuous boundary segments to the isolated point removal module, which returns optimized segments after removing isolated points. The candidate point manager then submits the optimized boundary segments to the motion consistency verification module, which completes the consistency judgment by requesting the platform's motion trajectory and providing the platform's motion trajectory data based on the sensor platform's positioning information. After completing the consistency judgment, the motion consistency verification module returns the verified consistent boundary. Finally, the candidate point manager performs spatial stitching and fusion to generate a complete, well-structured plot boundary. Furthermore, during actual deployment, the system dynamically fine-tunes the time window, perception weights, and regional strategies involved in boundary determination by combining the statistical features of multiple observations within a recent period. Specifically, the temporal window is appropriately narrowed during periods of frequent co-occurrence of similar events to improve determination accuracy. Furthermore, when the observation area is obscured by vegetation, strong light reflections, or blurred images, the system reduces its reliance on visual modalities and instead increases its sensitivity to laser structural features. To further enhance system stability, an error tolerance band is preset during the boundary line fitting process. Its spatial range is determined based on the platform positioning accuracy and typical boundary width experience. Points exceeding this range will be judged as abnormal responses and excluded. In order to improve the hierarchy and systematicness of the processing flow, the key modules in the overall solution belong to a unified boundary construction framework. Among them, the identification of change events, extraction of co-existing instances, spatial clustering and trajectory verification steps are uniformly included in the candidate point management mechanism for collaborative scheduling. The mechanism supports dynamic adjustment of parameter strategies in various stages such as candidate point generation, screening, verification and splicing, and can automatically select the optimal processing path based on the equipment operating status and environmental characteristics to ensure that the system still has good generalization ability and reliability under diverse terrains and complex operating conditions. These are all extended implementation methods that are known to ordinary technicians in this field.

[0044] Example 4: In a typical hilly and complex agricultural area, land boundary recognition faces a series of challenges such as low multi-source heterogeneous data fusion accuracy, significant visual occlusion, and unclear object edges. To effectively improve the accuracy and stability of land boundary recognition under such complex terrain conditions, this example adopts a boundary recognition method based on the fusion of laser point clouds and visible light images. It specifically relies on the spatial structure enhancement mechanism and feature complementary alignment strategy of the present invention to achieve efficient fusion and accurate recognition of multi-source perception data. In this scenario, a ground multi-rotor drone platform first performs an aerial scan of the target land before operation to obtain high-density lidar point cloud data and high-resolution visible light image data covering the area to be identified. During the acquisition process, the lidar sensor is installed at the center of the bottom of the drone, and the visible light camera is installed in the area below and in front of it, forming a certain angle between the two in terms of visual angle. To ensure the spatial alignment of the data collected by the two types of sensors, a synchronous calibration operation is performed at the initial system startup. Known feature calibration objects are used to perform spatial posture solution, and then an accurate external parameter transformation relationship between the laser point cloud and the visible light image is established. After acquisition, the raw laser point cloud data is first subjected to ground point separation and noise removal. In this embodiment, an improved progressive morphological filtering method based on ground estimation is used. By setting the ground smoothness threshold and the maximum height change rate, non-ground vegetation points and abnormal scattered points are effectively filtered out. At the same time, the visible light image is preprocessed with multi-scale edge enhancement, and an image sharpening algorithm based on guided filtering is introduced to highlight the boundary features of the ground objects. This step has a significant effect on improving the accuracy of subsequent feature extraction and registration. After the data preprocessing stage is completed, the multimodal feature extraction module is entered. For the laser point cloud data, a uniform sparse point set is generated through voxel downsampling, and local geometric structure features are extracted based on the voxel feature encoding network. The spherical neighborhood feature aggregation module is then combined to capture the contextual structural semantic information of the point cloud at a spatial scale. For the visible light image, a lightweight convolutional neural network is used to extract multi-level texture and edge features, and a feature pyramid structure is used to effectively fuse information at different scales. After feature extraction is completed, the feature complementary alignment mechanism proposed in this invention is used to achieve fine registration of the laser point cloud structural features and image texture features in a unified space.

[0045] During the feature alignment process, the system introduces a feature fusion strategy based on spatial attention weights. By calculating the mutual information between point cloud features and image features within the same spatial region, the system dynamically adjusts the weights of various features in the fusion process. Specifically, in areas with vegetation occlusion or sudden terrain changes, the system assigns higher fusion weights to the laser point cloud to enhance the representation of 3D structural information. In areas with rich edge texture and slow structural changes, the system appropriately increases the weight of image features to enhance the perception of boundary details. Once feature fusion is complete, the system enters the parcel boundary extraction module. This module comprehensively considers the spatial distribution, texture continuity, and geometric consistency of the fused features to construct a boundary confidence map. Combined with a boundary fitting algorithm based on graph cut optimization, the module achieves high-precision closed-loop extraction of parcel boundaries. During this process, the system designs adaptive boundary tracking rules for typical complex terrain scenarios, such as slope turns and the contiguous borders of fragmented parcels. Based on the local mutation amplitude and directional consistency of the fused features, the system adaptively adjusts the fitting window size and fitting curve elasticity coefficient to ensure the continuity and geometric rationality of the boundary curve.In the entire processing flow, the system further introduces a feedback mechanism for self-verification of fusion accuracy: based on the initially extracted land boundary, the boundary results are mapped back to the original point cloud space and image space respectively through the back-projection mechanism to evaluate the overlap and consistency index in the two types of data; if the overlap is found to be lower than the preset threshold (such as the boundary offset exceeds a certain range), the system will automatically go back to the feature fusion stage and adjust the feature weight distribution until the overlap meets the accuracy requirements and the final boundary confirmation is completed. This feedback mechanism has significantly improved the stability of the fusion algorithm under multi-source uncertainty conditions in practical applications. At the same time, in order to take into account the system processing capabilities and recognition accuracy requirements under different operating environments, in scenarios with high-performance computing power support, the system can also select some deep feature extraction methods in the structure enhancement module and feature alignment strategy to further improve the boundary expression capabilities under complex terrain conditions; such strategies are scalable paths for this method and are consistent with the aforementioned lightweight process in the core event perception mechanism and co-occurrence judgment logic. The processing module can be flexibly enabled according to the platform computing power conditions, thereby achieving a balanced switch between resource constraints and high-precision requirements. The extraction of change events is the premise of co-occurrence judgment. The setting of the time window is based on the dynamic adjustment of the time distribution of previously successfully identified events, and is based on real-time statistics of the current data stream rather than relying on static templates. In order to enhance the actual perception effect, the mentioned local spatial changes and modality-specific events refer to elevation mutations or density gradient mutations in laser point cloud data, and significant transitions in brightness or texture distribution in image information. The system obtains the judgment basis by comparing the internal structure distribution or grayscale changes in adjacent areas; among them, structural contrast can be understood as the density of spatial points in the local area, and texture activity reflects the richness of brightness detail changes in the image window. Both use the local differences of the current data frame as the evaluation basis, avoiding dependence on complex mathematical models, ensuring that the method is operable in different computing environments, and are all extended implementation methods known to ordinary technicians in this field.

[0046] Example 5: In another specific application of the present invention, for example, an operation scenario for automated management of hilly vineyards, agricultural machinery equipped with multimodal sensors needs to achieve autonomous navigation in a plot with undulating micro-topography and uneven lighting. In this specific operation scenario, a key technical challenge is that the boundaries between vine rows and inter-row operation channels are composed of irregular grass edges, exposed soil, and vine shadows. Their physical boundaries are fuzzy and their visual characteristics are changeable. Existing technical solutions are easily affected by slight changes in sensor posture or sudden changes in ambient lighting, and boundary recognition is often interrupted or misjudged, which leads to navigation operation failure. The method disclosed in the present invention effectively addresses this challenge with its event-driven mechanism, temporal co-occurrence judgment logic, and a set of embedded deterministic calibration and adaptive procedures.

[0047] Specifically, before the agricultural machinery is deployed to the vineyard for the first time, for example, a pre-deployment deterministic calibration procedure can be performed to systematically set the core parameters. First, a ten-meter-long calibration section that contains the overall boundary characteristics that can represent the plot is selected in the operating area. The operator drives the agricultural machinery along this section at a low speed, synchronously collecting the laser point cloud data stream and the visible light image data stream, and manually marking the ground truth boundary trajectory of the section through a high-precision differential positioning system. This serves as the benchmark truth value for subsequent parameter optimization to determine the laser mode adjustment coefficient. The system first collects the elevation difference values ​​of adjacent laser points in all non-boundary areas on the calibration section to form a background noise sample set and calculates its statistical standard deviation. Then, the system iteratively tests each Candidate value. For any candidate value, the system uses the formula Calculate the detection threshold of the geometric change event, and based on the aforementioned benchmark true value, evaluate the true positive rate and false positive rate under the threshold, and finally select the one that maximizes the difference between the true positive rate and the false positive rate. value, as the optimal setting for this operation scenario, and accordingly, the visible light modal adjustment coefficient Determine, follow and The calibration procedure is exactly the same, except that the evaluation indicator is replaced by the absolute value of the pixel grayscale value or color component difference instead of the elevation difference. The horizontal distance, vertical distance, and time interval thresholds required for spatial clustering are determined by performing a grid search on the boundary candidate point set generated by the calibration section. The set of threshold combinations that minimize the average Hausdorff distance between the boundary segments formed after clustering and the benchmark true value trajectory is selected.

[0048] In real-time workflow, timing weak synchronization window The management follows a set of full life cycle adaptive logic. During the cold start phase of the system, there is no historical common case. The synchronization window is set to an initial value of 0.8 seconds. Once the system successfully identifies and confirms the first five instances of cross-modal co-reality, its dynamic adjustment mechanism is activated. This mechanism continuously maintains a sliding queue of timestamp difference samples from the most recent fifty valid co-reality instances. Whenever a new co-reality instance is confirmed, the system recalculates μ and the variance σ of the timestamp differences in this queue and immediately updates the effective range of the synchronization window to [μ - 1.5σ, μ + 1.5σ]. This allows the synchronization window to automatically track and compensate for minor inter-sensor delay drift caused by long-term equipment operation or environmental factors, ensuring the continued effectiveness of co-occurrence determination. When the agricultural machinery reaches the end of the field and its movement speed drops below the preset low-speed motion threshold of 0.5 meters per second, the event extraction logic seamlessly switches, shifting from temporal difference detection to local spatial structure contrast analysis of the current single-frame point cloud. Specifically, the system defines a cylindrical neighborhood with a radius of 0.4 meters and a height of 0.2 meters with each laser point as the center, and calculates the ratio of the point cloud density of the upper half of the space to the lower half of the space in the neighborhood. When this ratio exceeds a specific threshold, it is determined to be a geometric change event. The threshold is set to the 95th percentile of the density ratio set calculated from all laser points in the current data frame, thereby achieving adaptive capture of significant surface structure changes. At the same time, the extraction of visual change events in the visible light modality is also transformed from one-dimensional adjacent pixel comparison to texture activity evaluation within 32×32 pixel blocks in the image. The activity is quantified as the standard deviation of the grayscale values ​​of all pixels in the block.

[0049] As a result, through this complete technical solution of calibration procedures, full life cycle adaptive logic and multi-modal event extraction, agricultural machinery deployed in hilly vineyards can stably and continuously identify fuzzy and changeable row boundaries. The problems of discontinuous boundary recognition or position offset caused by terrain and lighting changes are effectively solved. Agricultural machinery then achieves high-precision autonomous tracking operations. By eliminating the reliance on precise spatial alignment of sensors and relying instead on temporal consensus and dynamic adaptive strategies, the environmental adaptability of plot boundary recognition in complex operating environments is significantly improved.

[0050] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A land boundary recognition method based on the fusion of laser point cloud and visible light, characterized in that: The method comprises the following steps: Step a, independently and in parallel processing the first modal data stream and the second modal data stream, including: step a1, detecting, from the first modal data stream, i.e., the laser point cloud data stream, the elevation difference between the current laser point in the laser point cloud and its adjacent laser points in the scanning sequence, or detecting the point cloud density gradient within the local neighborhood of the current laser point based on the local spatial instantaneous difference of the surface geometric elevation or point cloud density, and comparing the detection result with a first threshold to extract the laser modal geometric change event and its first timestamp; step a2, detecting, from the second modal data stream, i.e., the visible light image data stream, the absolute value of the grayscale value or color component difference between a pixel point on a predetermined line of interest in the visible light image and its one-dimensional adjacent pixels based on the local spatial instantaneous transition of the image texture, color or brightness, and comparing the detection result with a second threshold to extract the visible light modal visual change event and its second timestamp; Step b, based on the first timestamp and the second timestamp, identifying a co-occurrence instance including both a laser modal geometry change event and a visible light modal visual change event within a predetermined time window; Step c: identifying the spatial location of at least one event associated with the co-occurrence instance as a candidate land parcel boundary point; Step d: performing spatiotemporal filtering on the land parcel boundary candidate point sequence to form the land parcel boundary.

2. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The first threshold value in step a1 is determined by multiplying the elevation difference or the statistical standard deviation of the point cloud density gradient of the recently collected laser point cloud data by a predetermined adjustment coefficient to obtain the first threshold value. ,in is the first predetermined adjustment coefficient, is the first statistical standard deviation.

3. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The second threshold value in step a2 is determined by multiplying the statistical standard deviation of the absolute value of the grayscale value or color component difference of the recently collected visible light image by a predetermined adjustment coefficient to obtain the second threshold value. ,in is the second predetermined adjustment coefficient, is the second statistical standard deviation.

4. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The spatiotemporal filtering process in step d includes: clustering and connecting the plot boundary candidate points based on the proximity of the estimated spatial positions of the plot boundary candidate points to form continuous boundary segments; and removing isolated candidate points that are not temporally or spatially associated with the remaining candidate points.

5. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The method further includes: detecting the movement speed of a sensing platform that collects laser point cloud data streams and visible light image data streams; and when the movement speed of the sensing platform is lower than a predetermined speed threshold, adjusting the extraction of change events of at least one modality in step a, the adjustment including: extracting change events by analyzing structural contrast or visual feature activity within the local spatial data based on the local spatial data acquired by the modality sensor at the current moment.

6. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The predetermined time window in step b is set by dynamically adjusting the online statistical characteristics of the timestamp differences actually observed between the laser modal geometric change events and the visible light modal visual change events corresponding to the co-existing instances that have recently been identified as candidate plot boundary points or have been output as final plot boundaries.

7. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The first threshold in step a1 and the second threshold in step a2 are both adjusted periodically or non-periodically based on the statistical characteristics of the recent data of their respective modalities.

8. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The spatiotemporal filtering process in step d further includes: performing consistency check between the motion trajectory of the candidate points at the land boundary and the motion trajectory of the sensing platform based on the motion information of the sensing platform.

9. The method for identifying land parcel boundaries by fusing laser point cloud and visible light according to claim 1, characterized in that: The method for identifying the spatial positions of candidate land boundary points in step c includes: approximating the spatial positions of events associated with the inter-modal event pair based on a predetermined approximate relative position relationship between the laser point cloud sensor and the visible light image sensor.

Images