Laser radar technology based method for fixed fob identification on a ramp

Abstract

This invention provides a fixed FOD (Foreign Object Demand) identification method for aprons based on lidar technology, belonging to the field of FOD identification technology. The method includes: acquiring three-dimensional point cloud data of the apron area using a solid-state lidar fixedly installed on a high-mast light or building top at a scanning frequency of ≥100Hz; sequentially performing noise filtering and ground segmentation on the acquired three-dimensional point cloud data; performing density clustering on the point cloud data of non-ground points to identify potential FOD clusters; extracting feature vectors of potential FOD clusters, determining whether they are FODs using a classification model, and performing confidence verification based on multi-frame time-series data; when a FOD that passes the confidence verification is identified, triggering an audible and visual alarm and recording its location, size, and time information. This invention, through solid-state lidar and point cloud processing algorithms, can effectively suppress noise and background interference and accurately extract small targets, and can achieve high-precision, low-false-alarm real-time detection of millimeter-level FODs in complex environments.

Description

Technical Field

[0001] This invention relates to the field of FOD identification technology, and specifically to a fixed FOD identification method for aprons based on lidar technology. Background Technology

[0002] Foreign Object Debris (FOD) on airport tarmac is a major threat to aviation safety, encompassing objects such as metal parts, rubber fragments, and gravel. According to incomplete statistics, direct losses to global civil aviation due to FOD exceed US$4 billion annually, with indirect losses reaching up to four times the direct losses.

[0003] Currently, the following technical bottlenecks exist in the field of FOD detection: First, there is a lack of millimeter-level FOD detection. Existing sensors cannot reliably identify hazardous objects in the 5 to 10 mm range (such as nuts, tire valve cores, etc.), which can cause tire blowouts.

[0004] Second, there is insufficient suppression of interference from complex backgrounds. Optical systems are affected by reflections from oil stains, and radar systems are affected by electromagnetic signals, resulting in a high false alarm rate in real-world apron scenarios.

[0005] Third, the system lacks comprehensive all-weather support capabilities. Under extreme conditions such as nighttime, heavy rain, and high temperatures (e.g., above 50°C), the reliability of existing FOD detection systems is relatively low.

[0006] Fourth, maintenance costs are too high. Existing distributed sensors require dismantling of the pavement for maintenance, while mechanical LiDAR requires regular calibration, both of which result in a heavy operational burden and high maintenance costs. Summary of the Invention

[0007] To address the problems in related technologies, this invention provides a method for identifying fixed FOD on aprons based on lidar technology.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying fixed FOD on an apron based on lidar technology includes the following steps: Step S1: Acquire three-dimensional point cloud data of the apron area using a solid-state lidar fixedly installed on the apron high-mast light or the top of a building, with a scanning frequency of ≥100Hz. The three-dimensional point cloud data includes the three-dimensional coordinates and reflection intensity information of each point. Step S2: The acquired 3D point cloud data is subjected to noise filtering and ground segmentation in sequence. The noise filtering is used to remove dynamic interference points, and the ground segmentation is used to separate ground points from non-ground points. Step S3: Perform density clustering on the point cloud data of non-ground points to identify potential FOD object clusters; Step S4: Extract the feature vectors of potential FOD object clusters, determine whether they are FODs through a classification model, and perform confidence verification based on multi-frame time series data; Step S5: When a FOD that has passed the confidence verification is identified, trigger an audible and visual alarm and record the location, size and time information.

[0009] Optionally, in step S2, the noise filtering employs a dynamic radius filtering algorithm, whose neighborhood radius... satisfy: In the formula, Based on the radius, and , The correlation coefficient of rainfall intensity and the implicit unit conversion factor are used to make The unit is , The value was determined through calibration tests (its range can be...). ), The rainfall intensity value is output in real time by the rain sensors deployed on the apron.

[0010] Optionally, in step S2, the ground segmentation employs a RANSAC plane fitting algorithm weighted by reflection intensity, with the objective function being: In the formula, For point cloud indexing, For the first The reflection intensity weight at each point , , and These are the coefficients of the ground plane equation, obtained through fitting. , and The first Unit coordinates of each point in, , For the first The reflection intensity value at each point, This represents the maximum detectable reflection intensity of the system. This represents the threshold value for the intensity of oil stain reflection.

[0011] Optionally, the oil stain reflection intensity threshold The calibration methods include: Step S2-1: Collect at least 1000 sample point clouds from typical oil-contaminated areas on the apron; Step S2-2: Calculate the histogram of reflectance intensity distribution for all sample points, and take the median value of the 85th-95th percentile interval as the median value. ; Step S2-3: Periodically update the threshold using point cloud data of manually labeled oil stain areas.

[0012] Optionally, in step S3, the density classification employs the multi-scale DBSCAN algorithm, whose core parameter is the neighborhood radius. satisfy: In the formula, Based on the neighborhood radius and , It is a negative scale correlation factor and , The estimated size of the current object cluster is obtained by calculating the minimum eigenvalue of the point cloud covariance matrix through principal component analysis.

[0013] Optionally, the multi-scale DBSCAN algorithm is configured to stably detect FOD objects with a minimum size of 5 mm, and The iterative calculation satisfies: In the formula, These are the minimum eigenvalues ​​of the point cloud covariance matrix.

[0014] Optionally, in step S4, the feature vector is defined as: in, The longest axis length of the object is calculated by performing PCA decomposition on the object cluster point cloud and taking the direction of the eigenvector corresponding to the largest eigenvalue. The surface area of ​​the object is calculated using the point cloud convex hull algorithm. Let be the standard deviation of the point reflection intensity within the cluster of objects. This represents the centroid displacement of an object across three consecutive frames of point cloud data.

[0015] Optionally, in step S4, the classification model employs a deep learning network based on the PointNet++ architecture, and its output probability is: In the formula, This is the predicted probability value for FOD. For activation function, The weights for the deep learning model were obtained through pre-training on the airport FOD dataset. For feature vectors The amount, Integer and , , , , , This is the model bias term.

[0016] Optionally, in step S4, the confidence verification must satisfy: in, There is a comprehensive confidence level for FOD. For the first FOD probability of a frame For continuous detection of frames and , It is the time decay factor and Used to reduce the weight of historical frames. For historical frame time index, The timestamp of the current frame. The sampling interval for point cloud frames.

[0017] Optionally, the real-time processing of steps S1 to S4 is achieved through GPU parallel acceleration, specifically including: The point cloud clustering task is divided into thread blocks, and the size of each thread block is... satisfy: In the formula, This represents the total number of point clouds in a single frame. This refers to the number of GPU stream processors. The system's end-to-end processing latency is ≤200ms.

[0018] Beneficial effects: 1. Through the above technical solution, this invention provides basic data using high-resolution solid-state LiDAR, and utilizes innovative point cloud processing algorithms (e.g., dynamic radius filtering, reflection intensity weighted RANSAC ground segmentation, and adaptive multi-scale DBSCAN clustering described below) to effectively suppress noise and background interference and accurately extract minute targets. Combined with high-dimensional feature extraction and a deep learning-based classifier for accurate identification, and employing rigorous multi-frame temporal confidence verification to significantly reduce false alarms, this invention ultimately achieves high-precision, low-false-alarm, real-time target detection of millimeter-level FOD in complex environments (e.g., oil stains, rainfall) and extreme conditions (nighttime, high temperatures). Simultaneously, the non-destructive fixed installation and solid-state sensor design significantly reduce the system's installation and maintenance costs.

[0019] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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.

[0021] in: Figure 1 This is a flowchart illustrating the steps of a fixed FOD identification method for aprons based on lidar technology, provided by an exemplary embodiment of the present invention. Figure 2 This is a schematic diagram of a capture guide vehicle provided in an exemplary embodiment; Figure 3 This is a schematic diagram illustrating the processing of moving vehicle parameters according to an exemplary embodiment; Figure 4 This is a schematic diagram illustrating the processing of the distance between a moving vehicle and an aircraft, provided as an exemplary embodiment. Figure 5 This is a schematic diagram illustrating the algorithm processing and matching with the guide vehicle model library provided in an exemplary implementation. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] To facilitate a clearer and more accurate understanding of the technical solution of this invention by those skilled in the art, the existing related technologies and their technical problems will be further explained below.

[0025] Foreign Object Debris (FOD) identification on airport aprons is a crucial aspect of aviation safety. The current state of technology application and management in the industry is as follows: First, traditional methods still dominate.

[0026] 1. Manual inspection: Relies on visual inspection by inspectors or handheld devices (such as magnetic sticks, metal detectors), which is inefficient (small coverage area per inspection) and greatly affected by light and weather. At the same time, manual inspection requires high-frequency inspection (such as once every 2 hours), which greatly increases labor costs and inspection risks. In addition, it is easy to miss inspections at night or in complex areas (such as runway joints, shoulders, etc.).

[0027] 2. Sweeping vehicle assistance: Mechanical sweepers are used on runways / slipways, but they have a low removal rate for small metal objects (e.g., screws) or non-metals (e.g., plastics, rubber, etc.), and they cannot achieve real-time monitoring, only able to operate during relatively fixed periods of time.

[0028] Second, technology-driven FOD monitoring systems.

[0029] 1. Radar and optics fusion technology.

[0030] The advantages of millimeter-wave radar are: it can detect metal objects as small as 2cm, and it has strong stability in low-light environments. Its disadvantages are: lower recognition rate for non-metallic materials (e.g., rubber, plastics), and susceptibility to interference from airport wireless equipment (e.g., air traffic control radar, communication systems).

[0031] For infrared + visible light cameras, they can identify FOD through image recognition algorithms, but their ability to deal with complex backgrounds on the tarmac (such as oil stains, shadows, and low-light scenes) is relatively weak.

[0032] 2. Distributed sensor networks.

[0033] It typically involves embedding pressure / vibration sensors into the pavement to monitor abnormal vibration signals. However, its deployment costs are high, and its maintenance operations are very complex (requiring the dismantling of the pavement).

[0034] In summary, the existing technologies suffer from several problems, including the lack of millimeter-level FOD detection, insufficient suppression of complex background interference, weak all-time domain protection capabilities, and excessively high maintenance costs.

[0035] In view of this, the present invention provides a novel solution: a fixed FOD identification method for aprons based on lidar technology. It aims to overcome the bottleneck of small-scale FOD detection and improve robustness in complex environments, achieving stable identification of FOD at the 5-10mm level, reducing false alarm rates in complex backgrounds, and constructing an easily maintainable detection system.

[0036] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] like Figure 1As shown, this embodiment provides a method for identifying fixed FOD on aprons based on lidar technology, including the following steps: Step S1: Acquire three-dimensional point cloud data of the apron area using a solid-state lidar fixedly installed on the high-mast lights of the apron or the top of a building, with a scanning frequency of ≥100Hz. The three-dimensional point cloud data includes the three-dimensional coordinates and reflection intensity information of each point. Step S2: The acquired 3D point cloud data is subjected to noise filtering and ground segmentation in sequence. Noise filtering is used to remove dynamic interference points, and ground segmentation is used to separate ground points from non-ground points. Step S3: Perform density clustering on the point cloud data of non-ground points to identify potential FOD object clusters; Step S4: Extract the feature vectors of potential FOD object clusters, determine whether they are FODs through a classification model, and perform confidence verification based on multi-frame time series data; Step S5: When a FOD that has passed the confidence verification is identified, trigger an audible and visual alarm and record the location, size and time information.

[0038] Through the above technical solution, this invention provides basic data using high-resolution solid-state LiDAR, and utilizes innovative point cloud processing algorithms (e.g., dynamic radius filtering, reflection intensity weighted RANSAC ground segmentation, and adaptive multi-scale DBSCAN clustering described below) to effectively suppress noise and background interference and accurately extract minute targets. Combined with high-dimensional feature extraction and a deep learning-based classifier for accurate identification, and employing rigorous multi-frame temporal confidence verification to significantly reduce false alarms, this invention ultimately achieves high-precision, low-false-alarm, real-time target detection of millimeter-level FOD in complex environments (e.g., oil stains, rainfall) and extreme conditions (nighttime, high temperatures). Simultaneously, the non-destructive fixed installation and solid-state sensor design significantly reduce the system's installation and maintenance costs.

[0039] Specifically, firstly, this invention uses a solid-state LiDAR for data acquisition, which boasts extremely high angular resolution (a fraction of a degree or even higher) and fine distance resolution (centimeter-level or millimeter-level). This enables it to detect very small objects and generate sufficiently dense point clouds to depict their outlines. Compared to traditional mechanical LiDAR, the solid-state solution offers more stable scanning and less point cloud jitter, making it more conducive to small target identification. Simultaneously, the solid-state LiDAR scans at a high frequency (≥100Hz), and the extremely high frame rate (over 100 scans per second) means the system can "see" the same tiny object multiple times within a very short time interval. This significantly increases the probability of capturing tiny, low-reflectivity (potentially only detectable at certain angles), or partially occluded FODs in consecutive frames, providing a foundation for the timing verification in subsequent step S4. In this way, the bottleneck of "lacking millimeter-level FOD detection" in existing related technologies can be specifically addressed.

[0040] Second, in step S2 of this invention, the acquired three-dimensional point cloud data is sequentially subjected to noise filtering and ground segmentation. This effectively suppresses false alarms caused by oil stain reflections, other environmental clutter (non-FOD small objects), etc. This specifically addresses the bottleneck of "insufficient suppression of complex background interference" in existing related technologies.

[0041] Third, this invention employs a solid-state lidar (an active sensor, independent of ambient light), thus effectively adapting to low-light environments. Solid-state lidar has no moving parts and its high-temperature resistance is generally superior to mechanical LiDAR. Simultaneously, noise filtering removes dynamic interference points, effectively suppressing raindrop noise interference and ensuring point cloud quality. Therefore, the method of this invention can effectively improve detection reliability under extreme conditions such as nighttime, heavy rain, and high temperatures, thus specifically addressing the bottleneck of "weak all-time domain assurance capability" in existing related technologies.

[0042] Fourth, the solid-state lidar of this invention is installed on high-mast lights or building rooftops using a non-embedded installation method, completely avoiding pavement damage during maintenance. This not only greatly simplifies the installation process but also effectively reduces initial installation costs and the high costs associated with subsequent pavement damage during maintenance. Furthermore, solid-state lidar lacks precision moving parts such as rotating motors, resulting in higher reliability (fewer mechanical failure points), longer service life (less mechanical wear), and lower maintenance requirements (no need for periodic calibration; maintenance is limited to cleaning and routine electrical checks). In addition, compared to distributed sensor networks, solid-state lidar offers simpler and more centralized data processing and transmission, effectively simplifying system architecture and wiring maintenance. This effectively addresses the bottleneck of "excessive maintenance costs" in existing related technologies.

[0043] In one embodiment of the present invention, in step S2, noise filtering employs a dynamic radius filtering algorithm, whose neighborhood radius... satisfy: In the formula, Based on the radius, and , The correlation coefficient of rainfall intensity and the implicit unit conversion factor are used to make The unit is , The value was determined through calibration tests (its range can be...). ), The rainfall intensity value is output in real time by the rain sensors deployed on the apron.

[0044] In this embodiment, firstly, by dynamically suppressing rainfall noise, the dense noise generated by raindrops in point clouds during rainy weather can be significantly reduced, preventing it from being misidentified as potential FOD. Furthermore, while suppressing noise, it ensures that real FOD point clouds with a size ≥5mm are not over-filtered.

[0045] Second, the filter intensity varies with the real-time rainfall intensity. Dynamic adjustment allows it to automatically adapt to different rainfall intensities without human intervention, effectively enabling it to adapt to environmental changes. This also directly reduces false alarms caused by meteorological interference and improves reliability under severe weather conditions.

[0046] In this embodiment, it should be noted that, First, in the algorithm formula provided by this invention ( )middle, It is the basic radius ( This is in the case of no rain or light rain. The standard filter radius used under conditions ≈0). It is calibrated based on typical apron background noise (such as slight dust, sensor background noise) and the minimum FOD size to be preserved (5mm). Smaller Designed to protect tiny FOD. Meanwhile, It quantifies how much the filter radius needs to be increased per unit increase in rainfall intensity to effectively suppress the newly added noise density under that rainfall intensity. This is determined through calibration experiments (such as testing under different known rainfall intensities to find the optimal balance between noise filtering rate and small FOD retention rate). value).

[0047] Second, the principle behind rainfall's interference with point clouds is as follows: When the laser beam emitted by the lidar encounters raindrops, it is scattered and reflected, manifesting as numerous discrete, transiently existing high-frequency noise points in the point cloud. These noise points are randomly distributed in the detection space; a single raindrop typically produces one or a few isolated points. Furthermore, the noise points appear and disappear extremely quickly (related to the raindrop's falling speed). In addition, the rainfall intensity ( The larger the radius, the more raindrops per unit volume, and the higher the noise density. Based on this, traditional fixed-radius filtering algorithms may work well in light rain (where noise is sparse), but in heavy rain, if the filtering radius is too small, it cannot effectively filter out dense noise (noise may reach the "non-noise" density threshold in its small neighborhood). If the filtering radius is too large, it may misjudge real tiny FODs (such as point cloud clusters of 5mm nuts) as noise and filter them out (because tiny FODs themselves occupy a small space, the number of points is relatively small, and the average density in a large neighborhood may be lower than the threshold).

[0048] In this invention, the dynamic radius The principle for resolving rainfall interference is as follows: When it drizzles ( (smaller) At this point, the noise is relatively sparse. In small... Within the neighborhood, the number N of other points (noise or real points) surrounding a noise point is usually below a threshold, and therefore can be effectively identified as noise and filtered out. However, real, tiny FOD point cloud clusters, even in small... Within the neighborhood, because the points are very close together and the density is high enough, N will be greater than the threshold and thus be retained.

[0049] During heavy rain or torrential rain ( (larger) The key principles behind the significant increase include: Although the noise density is high, its spatial randomness is even stronger: while the overall noise density increases dramatically during heavy rain, the randomness of raindrop falling positions results in a greater degree of spatial randomness in the expanded neighborhood. Within a sufficiently large neighborhood, the expected rate of increase in average noise density may be lower than the rate of increase in neighborhood volume. In other words, within a sufficiently large neighborhood, the spatial clustering of noise is insufficient, and its density may still be lower than that of real objects (such as FOD).

[0050] Increase the noise filtering threshold: increase This means the algorithm requires a point to find a sufficient number (N ≥ threshold) of "partner points" within a larger spatial range to avoid being considered noise. For isolated noise points or small clusters of noise generated by a single raindrop, after amplification... Within the neighborhood, the probability of finding enough other noise points decreases (because noise points are randomly distributed), making it easier to filter out those that satisfy N < threshold.

[0051] Preserving True FOD: True FOD (even small ones) has a spatially continuous and tightly clustered point cloud. When the filter radius... When it increases, as long as Not too big (this is due to) If the point cloud is enclosed within a reasonable range, the N value of the points within the cluster will still be high when the points are calculated in their neighborhoods (because the density inside the point cluster is high and continuous), and therefore will still be judged as valid points and retained. The lower limit (0.01m) and The upper limit (0.07) ensures that under extreme rainstorms, The growth will not be excessive enough to overwhelm the aggregation characteristics of the tiny FODs themselves.

[0052] Overall, this implementation directly addresses the problems of "insufficient suppression of complex background interference" (especially meteorological interference) and "weak all-time domain protection capability" (rainstorm conditions) mentioned in the background technology, providing an efficient and adaptive solution, which is one of the key guarantees for the environmental robustness of the entire FOD identification method.

[0053] In one embodiment of the present invention, in step S2, the ground segmentation employs a RANSAC plane fitting algorithm weighted by reflection intensity, and its optimization objective function is: In the formula, For point cloud indexing, For the first The reflection intensity weight at each point , , and These are the coefficients of the ground plane equation, obtained through fitting. , and The first Unit coordinates of each point in, , For the first The reflection intensity value at each point, This represents the maximum detectable reflection intensity of the system. This represents the threshold value for the intensity of oil stain reflection.

[0054] In this embodiment, the present invention introduces dynamic weights based on reflection intensity into the RANSAC plane fitting. accomplish: 1. Utilizing the high reflectivity of oil stains ( This optical characteristic distinguishes it from a clean surface; 2. Through low weight ( Actively reduce the influence of high-reflectivity points (suspected oil stains) in the least squares objective function. Simultaneously, through high weighting ( This strengthens the dominant role of low-reflectance points (clean ground points) in the fitting results. Thus, the above weights are applied to the least squares sum of squared distances (...). This allows the algorithm to preferentially fit the plane defined by clean ground points.

[0055] 3. The final fitted ground plane is unaffected by oil stains, accurately reflecting the apron geometry. Oil stains, due to their low weighting and larger weighted distance, are more easily and correctly segmented as non-ground points (although they are not FODs, this avoids the more dangerous mistake of treating them as ground points). Subsequent clustering and classification steps can better utilize other features (such as size, shape, and reflectivity uniformity). This helps distinguish between oil stains and actual FOD.

[0056] In this embodiment, it should be noted that, firstly, the core of the optimization objective function of this invention lies in using weights to control the fitting direction. Specifically, it utilizes the high reflectivity of oil stains. This salient feature, through weighting Mathematically, its influence on the ground plane fitting process is controlled. Simultaneously, oil stains are assigned low weights. This actively reduces the contribution of clean ground points to the objective function. This means that when searching for the optimal plane, the RANSAC algorithm no longer seeks to make the plane pass through these high-brightness points, or tolerates them deviating significantly from the plane. Furthermore, clean ground points are given high weights (…). This forces the algorithm to prioritize fitting planes close to these points. Thus, the final fitted ground plane is dominated by clean ground points, accurately reflecting the geometry of the apron. Highly reflective oil stains, even if geometrically located at ground level, are considered outliers by the algorithm due to their low weight and small contribution to the objective function, and are more easily segmented into non-ground points.

[0057] Second, regarding the weight of oil stains ( The design principles of this algorithm include: First, the value is much smaller than 1, which significantly weakens the influence. Second, the weight of oil stains cannot be set too small (e.g., 0.01), otherwise it may introduce numerical instability, and when a few oil stains are mixed with a large number of ground points, the algorithm may completely ignore them with little impact; however, when large areas of oil stains exist, too small a weight may cause the fitted plane to become completely uncontrolled. Therefore, after comprehensive consideration, the weight of oil stains is set to 0.3.

[0058] In one embodiment of the present invention, the oil stain reflection intensity threshold of the present invention... The calibration methods may include: Step S2-1: Collect at least 1000 sample point clouds from typical oil-contaminated areas on the apron; Step S2-2: Calculate the histogram of reflectance intensity distribution for all sample points, and take the median value of the 85th-95th percentile interval as the median value. ; Step S2-3: Periodically update the threshold using point cloud data of manually labeled oil stain areas.

[0059] In this embodiment, by setting the oil stain reflection intensity threshold Specific calibration and update methods are provided to address a fundamental issue concerning the effectiveness of the reflection intensity-weighted RANSAC algorithm: how to scientifically determine the key threshold for distinguishing oil stains from the ground. Specifically, this implementation method ensures that... It accurately reflects the typical reflection intensity characteristics of oil stains on a specific airport apron, avoiding the failure of weighting strategies due to unreasonable threshold settings; at the same time, it enables the system to adapt to the differences in reflection characteristics of different airports, different areas (e.g., oil stains may be heavier in cargo areas), and different periods (e.g., new pavement vs. old pavement); and through a regular update mechanism, it offsets the impact of pavement aging, changes in oil stain composition, and sensor performance drift on reflection intensity, ensuring the robustness of the system throughout its entire life cycle; furthermore, accurate... Weight The accuracy of ground segmentation is directly determined by the correct allocation of the ground, which can reduce false or missed FOD reports caused by misjudgment of oil stains.

[0060] In this embodiment, it should be noted that, firstly, for step S2-1, it involves obtaining typical oil stain contamination areas, such as selecting areas with high oil stain incidence, such as parking positions, taxiways, and maintenance areas, covering different types of oil stains (e.g., fresh, old, and mixed), and ensuring that the sample size is large enough (at least 1000 sample point clouds) to reflect the statistical distribution characteristics of oil stain reflection intensity. In this way, an oil stain reflection intensity database reflecting the actual situation of the airport can be established.

[0061] Second, for step S2-2, the principle is to determine the robust threshold based on statistical distribution. Specifically, it involves calculating the reflection intensity values ​​of all sample points (from all point cloud frames acquired in step S2-1). The frequency of reflections falling into different intensity ranges was summarized and counted to form a histogram of reflection intensity distribution. Then, the distribution characteristics were analyzed (true oil stains – mainly contributing to the high reflectivity; oil stain edges / mixed points – slightly lower reflection intensity; background noise / non-oil stains – even when collected from oil stain areas, they may contain a small number of splashed non-oil stains, such as points on nearby clean ground, birds, sensor noise, etc., which have lower reflection intensity). The median value of the 85th-95th percentile interval was then selected as the median value. .

[0062] Third, regarding steps S2-3, by periodically updating the thresholds, the following can be achieved: adapting to environmental changes (compensating for reflection characteristic drift caused by pavement aging and darkening, sensor performance degradation / contamination, and the introduction of new oil products), optimizing regional differences (if specific areas, such as freight areas, have significantly different oil stain characteristics, the thresholds can be individually calibrated and updated for those areas), and improving long-term accuracy (ensuring the system maintains high accuracy throughout its entire service life). It can always accurately reflect the current boundary of "oil stain vs. ground" reflection intensity.

[0063] In this implementation, it should be noted that, firstly, the 85th-95th percentile interval is chosen for the following reasons: 1. To avoid tail noise (below the 85th percentile, i.e., below P85): The lower end of the distribution (below P85) may contain a large number of non-oil stain points (noise, background ground points mixed in, weak reflection points at the edge of oil stains). Directly taking the minimum value or a low percentile (such as the P50 median) will be severely interfered with by these points, resulting in an excessively low threshold. 2. To avoid extreme outliers (above P95): The top of the distribution (above P95) may contain a very small number of abnormally bright points (such as specular reflection peaks, sensor saturation points, metallic reflections), which do not represent typical oil stains. Directly taking the maximum value or a high percentile (such as P99) will be interfered with by these points, resulting in an excessively high threshold. 3. To focus on the core oil stain distribution (P85-P95): This interval likely contains the most typical and concentrated oil stain point reflection intensity values. It removes the interference noise at the bottom and the extreme outliers at the top.

[0064] Secondly, the median is chosen because it is robust to potential asymmetric distributions or small fluctuations within the interval. Compared to the average, the median is less affected by extreme values ​​within the interval. Thus, the final result... It is a stable statistical quantity that can represent the reflection intensity level of the main body of a typical oil stain at the airport.

[0065] In one embodiment of the present invention, in step S3, the density classification employs the multi-scale DBSCAN algorithm, whose core parameter is the neighborhood radius. satisfy: In the formula, Based on the neighborhood radius and , It is a negative scale correlation factor and , The estimated size of the current object cluster is obtained by calculating the minimum eigenvalue of the point cloud covariance matrix through principal component analysis.

[0066] In this embodiment, density clustering is improved by introducing the multi-scale DBSCAN algorithm and its core parameter, neighborhood radius. An adaptive formula is developed to address the clustering failure issue caused by large differences in FOD size. Specifically, it can stably detect both millimeter-scale (≥5mm) small FODs (such as nuts) and centimeter / decimeter-scale large FODs (such as tool bags and gravel) in a single clustering operation, avoiding the failure of traditional single-scale clustering algorithms at extreme size values. Furthermore, it ensures that the point cloud clusters of small FODs are not "swallowed" or "ignored" by large objects or background noise, significantly improving the detection rate of 5mm-scale targets. In addition, it prevents the point cloud of large FODs from being overly segmented into multiple scattered small clusters, ensuring object integrity and facilitating subsequent feature extraction and classification. This improves robustness to FODs of different sizes and reduces missed detections due to size factors.

[0067] In this implementation, it should be noted that, firstly, the DBSCAN algorithm is a density-based clustering algorithm, and its core parameter is the neighborhood radius. And the minimum number of points, MinPts. It groups points in the point cloud that meet the following conditions into one class: core points (centered on this point, ...). The neighborhood of the radius contains at least MinPts points, and the boundary points (at the core point) Within its neighborhood (but with fewer than MinPts) and noisy points (neither core nor boundary points). Current DBSCAN algorithms are generally fixed. Detecting tiny FOD requires small To focus on localized high-density areas; detecting large-scale FOD requires large [data / capacity]. To connect relatively sparse interior points. Fixed. It is impossible to satisfy both of these extreme demands at the same time.

[0068] Based on this, the present invention proposes a dynamically adaptable neighborhood radius. To solve the above problems, the DBSCAN algorithm of this invention can be described as follows: 1. Initial segmentation / pre-clustering: A very loose (large) fixed group can be used first. Alternatively, preliminary grouping can be performed based on spatial location, dividing the point cloud into multiple candidate regions or supervoxels. Each candidate region contains one or more potential objects or a portion of the background.

[0069] 2. Perform the following steps for each candidate region: A. Calculation Perform PCA on all points within this region to calculate... ; B. Calculation Dynamics By substituting into the formula Computational Dynamics ; C. Apply DBSCAN: Use the calculated dynamics only within the candidate region. It performs standard DBSCAN clustering with the preset MinPts. The clustering result is a more refined cluster of objects identified at this region scale.

[0070] 3. Merge / Output: Integrate the clustering results of all regions to form the final segmentation.

[0071] Second, the following describes the present invention through adaptive... How to solve the size dilemma (as mentioned in point one above)? This is explained by the fact that it is impossible to satisfy both of these extreme needs at the same time.

[0072] When dealing with tiny FODs, Smaller → It is a negative number with a small absolute value → → At this point, within this small area, use a smaller Run DBSCAN. (Smaller) It is perfectly suited for focusing on local high-density point clouds of tiny objects. Background noise points or points that are slightly distant cannot be included in the neighborhood, but the points of tiny FODs can effectively meet the core point conditions, thus being correctly identified as an independent cluster, which can effectively ensure that tiny FODs are not overwhelmed.

[0073] When dealing with large FODs, Larger → It is a negative number with a large absolute value → Significantly less than 1 → At this point, within this large area, a relatively smaller one is used. ( Run DBSCAN. Within the initially segmented candidate regions for large objects ( Larger), using a more refined scale ( Clustering is performed. This not only more accurately identifies possible structural boundaries or voids within an object (although they may eventually be merged into a single object); most importantly, it effectively avoids incorrectly merging other independent objects or background noise adjacent to the large FOD. Follow By increasing the size of the boundary and actively decreasing it, the accuracy of boundary segmentation for large objects is improved, preventing the generation of "super-clusters." Thus, for the large object itself, as long as the spacing between its internal points is smaller than this smaller value... (Typically, the points on the surface of large objects are relatively uniform in spacing and smaller than their size), and they can still aggregate into a complete cluster through the connection relationship between the core points and the boundary points, thus effectively preventing large FODs from being fragmented.

[0074] In other words, use smaller sizes for tiny objects. This allows for focused localized high density, protecting tiny FODs from being overwhelmed by noise or missed detection. Smaller detectors are used for larger objects. This allows for the use of a finer scale within the target area, preventing large FODs from excessively merging with neighboring objects or the background, while ensuring their own integrity is not disrupted.

[0075] Third, in this implementation method Setting it to a negative value makes Able to follow The size increases and decreases accordingly to achieve "small objects using smaller..." Focus, large objects use smaller ones "Prevent overflow." At the same time... This ensures that the scaling is both effective and not excessive.

[0076] In one embodiment of the present invention, the multi-scale DBSCAN algorithm of the present invention is configured to stably detect FOD objects with a minimum size of 5 mm, and The iterative calculation satisfies: In the formula, These are the minimum eigenvalues ​​of the point cloud covariance matrix.

[0077] In this embodiment, the present invention limits... The calculation method is described, emphasizing its connection with millimeter-level FOD detection capabilities, to address the key input problem in adaptive neighborhood radius calculation: how to scientifically and stably quantify the spatial scale of object clusters. Specifically, it can provide not only estimates of the physical size of objects in their most compact dimension (…). ), which is the dynamic neighborhood radius of the multi-scale DBSCAN ( The computation provides reliable input and ensures that the algorithm has a stable size perception capability for tiny objects ≥5mm, which is the mathematical basis for its correct clustering and detection. In addition, the results depend only on the spatial distribution of the point cloud and are independent of the orientation of the object in the radar coordinate system.

[0078] In this implementation, it should be explained that, firstly, why is it necessary? And how to define it. This is due to the dynamic neighborhood radius. Calculation formula ( The core of ) lies in It can accurately reflect the physical size of the currently processed point cloud cluster, and it is the core input for realizing "size-aware" adaptive clustering. However, since a point cloud cluster is a collection of discrete points, how to use a scalar ( How can we effectively characterize the spatial scale of FOD, especially for small, irregular FODs? Then, we need... It must meet the following four requirements: clear physical meaning (can intuitively correspond to the geometric dimensions of the object, such as length, width, thickness, etc.), stability (insensitive to point cloud sampling density, noise, slight movement or rotation of the object), orientation independence (the result does not depend on the orientation of the object relative to the radar), and computational efficiency (can meet real-time requirements).

[0079] The following section will first explain the implementation process of principal component analysis (PCA) and the process of determining the minimum eigenvalue.

[0080] 1. Construct the point cloud covariance matrix: Assume the point cloud cluster being processed contains N points, and the coordinates of each point are... Calculate the 3D centroid of a point cloud : , , Calculate the 3×3 covariance matrix C: Its physical significance lies in the fact that the covariance matrix C quantifies the degree of dispersion (variance) of the point cloud around its centroid in three-dimensional space and the correlation of changes in different coordinate axis directions. Its diagonal elements represent the variance (dispersion) of the point in the x, y, and z directions, respectively, and the off-diagonal elements represent the covariance (correlation) between different directions.

[0081] 2. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C: , Three eigenvalues ​​were obtained , and ( ) and the corresponding feature vector , and ; Among them, for eigenvalues , and In this context, it represents the variance of the point cloud along the direction of the corresponding feature vector. Maximum indicates the highest degree of point cloud dispersion along the longest axis; Minimum (i.e., ), indicating that the point cloud has the lowest degree of dispersion in the most compact direction.

[0082] Thus, in a geometric sense, This can be approximated as the "radius" or "thickness" of the point cloud in its most compact dimension. Furthermore, the eigenvalues / vectors calculated by PCA are based on the statistical distribution of the point cloud itself and are independent of the object's rotation angle in the radar coordinate system. Regardless of the orientation of the FOD, the calculated... and All of them can reflect their inherent minimum dimension size.

[0083] The following explains why we chose... Instead Alternatively, an average value can be used for explanation.

[0084] first, It directly characterizes the size of an object in the smallest scale direction. This is crucial for identifying the cross-section of sheet-like objects (such as rubber sheets, which are thin but have a large area) or elongated objects (such as the diameter of a screw). If using (Maximum size) A thin wire may be mistakenly considered to have a large size, causing its neighborhood radius ϵ to be incorrectly reduced, which is not conducive to the complete clustering of its own point cloud.

[0085] Secondly, the point cloud of small FOD is usually closer to isotropic (spherical) or significantly compressed in one dimension. It is most sensitive to this compression dimension. Compared to the maximum size ( The average size, or the average size, is more stable in reflecting the key scale that limits the detection of tiny FODs.

[0086] Finally, the goal of the dynamic neighborhood radius ϵ is to protect small FODs (requiring small ϵ) and prevent large FODs from excessively merging (requiring even smaller ϵ). Effectively distinguishes tiny ( Small) and large ( Large objects, and provide the scale information of the thinnest part of large objects, which is the key input needed to prevent over-merging (the neighborhood radius should be set with reference to its thinnest part to avoid crossing into neighboring objects).

[0087] In one embodiment of the present invention, in step S4, the feature vector is defined as: in, The longest axis length of the object is calculated by performing PCA decomposition on the object cluster point cloud and taking the direction of the eigenvector corresponding to the largest eigenvalue. The surface area of ​​the object is calculated using the point cloud convex hull algorithm. Let be the standard deviation of the point reflection intensity within the cluster of objects. This represents the centroid displacement of an object across three consecutive frames of point cloud data.

[0088] In this embodiment, the present invention proposes a multidimensional feature vector by defining key aspects of the feature extraction step. The focus is on solving the core problems of accurate FOD identification and interference differentiation. Specifically, this involves considering the maximum size of the object (…). ), spatial range ( ), reflection uniformity ( ), short-term motion ( Four key discrimination features, integrating multi-dimensional information on geometry, material, and motion, significantly improve the ability to distinguish FOD from common interference objects (oil stains, reflective markings, fallen leaves, and birds). To address interference from oil stains / water accumulation, its uniform reflectivity is utilized ( The extremely low (FOD) characteristic, compared to true FOD (even with uniform material, small size can lead to FOD due to discrete points). (not extremely low) and differences in complex formation; For dynamic interference objects. Utilizing the absolute fixation of the real FOD (Focus on Distance). Based on its characteristics, it can reliably distinguish between birds, fallen leaves, plastic bags, and other objects that temporarily land on the ground. ).

[0089] This implementation method represents the core innovation of this invention at the feature engineering level, providing a systematic solution to the high false alarm rate problem caused by insufficient suppression of complex background interference in the prior art. The designed... and By focusing on the characteristics of materials and their optical and kinematic properties—two dimensions often overlooked by traditional methods—these features can significantly improve robustness and reliability in real, complex apron environments, and are a key element in achieving high-precision, low-false-alarm FOD detection.

[0090] In one embodiment of the present invention, in step S4, the classification model employs a deep learning network based on the PointNet++ architecture, and its output probability is: In the formula, This is the predicted probability value for FOD. For activation function, The weights for the deep learning model were obtained through pre-training on the airport FOD dataset. For feature vectors The amount, Integer and , , , , , This is the model bias term.

[0091] In this implementation, the classification model is specified, employing a deep learning network based on the PointNet++ architecture, and its output probabilities are provided. The calculation formula is proposed to address the insufficient generalization ability of traditional classification methods in complex FOD discrimination. Specifically, PointNet++, a network designed specifically for point clouds, is used. It possesses permutation invariance and hierarchical feature extraction capabilities, enabling efficient processing of point clouds or their extracted features F. Furthermore, the network learns weights... Automatically assign different features to the feature vector ( It adaptively determines importance and captures complex nonlinear combinatorial relationships, and utilizes the Sigmoid activation function. Output intuitive FOD probability values This supports subsequent confidence verification. In addition, the mechanism of pre-training based on the airport FOD dataset enables the model to deeply internalize the discrimination knowledge in specific scenarios, significantly improving the discrimination accuracy and generalization ability of complex interference (especially oil stains and afterimages of moving objects).

[0092] In this implementation, the output probability calculation formula ( In terms of ), The term is a weighted linear combination of the four features, and can be viewed as the projection of the optimal linear discriminant direction learned by the network onto the feature space. Model bias term. These are also learned parameters. They represent the parameters obtained when all features... The base logit value when it is 0 (or the mean) affects the overall bias of the classifier (e.g., the proportion of FOD samples in the dataset). Activation function The term is the result of a weighted sum mapped to the interval [0,1], which can be intuitively interpreted as "the probability that the object is a FOD". .

[0093] In this way, the output has a clear probabilistic meaning, which facilitates the subsequent setting of confidence thresholds. Furthermore, it provides a soft metric for decision-making, which reflects uncertainty better than hard labels (0 / 1), facilitating time-series fusion. In addition, the saturation characteristic of the Sigmoid (small gradients at both ends) provides some robustness to anomalous inputs.

[0094] This implementation describes the operation of the last layer (or decision layer) of the PointNet++ network. The preceding layers (multiple SA layers, FP layers, MLP layers) perform highly non-linear transformations and abstractions on the input feature vector F, learning richer feature representations. Finally, these high-level features are flattened or pooled into a vector, which is then mapped to a single node (Logit value) through a fully connected layer. This Logit value is the value in the formula. (Note: Here) In reality, these are the abstract features ultimately learned by the network, not the original input features. However, to simplify and emphasize the importance of features, this implementation uses the original feature symbols to represent the weight relationships. Finally, an activation function is applied to obtain the probabilities. The key point lies in the weighting. and It is learned end-to-end from the data, rather than being manually set.

[0095] In one embodiment of the present invention, in step S4, the confidence verification must satisfy: in, There is a comprehensive confidence level for FOD. For the first FOD probability of a frame For continuous detection of frames and , It is the time decay factor and Used to reduce the weight of historical frames. For historical frame time index, The timestamp of the current frame. The sampling interval for point cloud frames.

[0096] In this embodiment, the present invention precisely defines the confidence verification step and proposes a multi-frame temporal confidence verification formula to focus on solving the false alarm problem caused by single-frame detection uncertainty. Specifically, the present invention utilizes ≥3 frames of continuous scan data ( This integrates time-series information to achieve temporal information fusion; and, through exponentially decaying weights... ( This approach assigns higher weights to the current frame and neighboring frames to ensure the system is sensitive to the latest state, achieving attenuation weighting to emphasize recent events. Simultaneously, a high threshold (>0.8) is set for the weighted average confidence level, requiring the target to consistently and stably exhibit high FOD probability characteristics across multiple frames. Overall, this implementation leverages the advantage of LiDAR's ≥100Hz high-frequency scanning, combined with the probabilistic results output by a deep learning classifier. Through rigorous temporal statistical verification, it fundamentally eliminates the vast majority of false alarms caused by transient interference and short-lived objects, effectively solving the high false alarm rate problem caused by insufficient suppression of complex background interference in the background technology. Its parameter design ( , The optimal balance between real-time performance (≤200ms) and reliability is the core guarantee for the system to meet the stringent security requirements of airports.

[0097] In this embodiment, it should be noted that, for the confidence verification formula, where is the number of consecutive detection frames and That is, at least 3 frames of data are required for timing verification, balancing real-time performance and reliability. Exponential decay weighting term ( ) for different historical frames Assigning different weights; weighted average and normalization term ( First use The FOD probabilities of N consecutive frames are time-decayed and weighted, then normalized by dividing by N to ensure the output confidence falls within a reasonable probability range for comparison with a threshold of 0.8. In this way, confidence comprehensively reflects the degree of credibility that the target "persistently exhibits FOD" across recent frames; it is not merely a simple average but emphasizes the importance of recent evidence.

[0098] In one embodiment of the present invention, the real-time processing of steps S1 to S4 is implemented through GPU parallel acceleration, specifically including: The point cloud clustering task is divided into thread blocks, and the size of each thread block is... satisfy: In the formula, This represents the total number of point clouds in a single frame. This refers to the number of GPU stream processors. The system's end-to-end processing latency is ≤200ms.

[0099] In this implementation, key real-time performance guarantees are imposed on the overall processing flow from steps S1 to S4. By introducing a GPU parallel acceleration architecture and its thread block configuration formula, the core bottlenecks of the FOD detection system in terms of processing latency and maintenance costs are addressed. Specifically, the GPU is used as the core computing power carrier, leveraging its massively parallel architecture and high memory bandwidth to drastically accelerate computationally intensive tasks; simultaneously, a load balancing formula ( This ensures that tasks are evenly distributed across all GPU stream processors (SMs), and each SM has a load of approximately 128 threads (close to the hardware concurrency limit), maximizing resource utilization. Furthermore, through a combination of algorithm optimization, parallel computing, pipelined processing, asynchronous I / O, and other methods, the end-to-end latency can be compressed to ≤200ms, meeting the real-time requirements of safety-critical systems.

[0100] In this implementation, it should be noted that for the load balancing formula ( In this regard, it aims to maximize GPU utilization: 1. Distribute the single-frame point cloud processing task evenly across all SMs; 2. Ensure that each SM is allocated a load of approximately 128 threads (close to its maximum parallel capability).

[0101] 3. The total number of threads is proportional to the number of points, enabling dynamic task allocation.

[0102] The following is a brief description of the process of scanning and recognizing a guide vehicle using the method of the present invention, with reference to an exemplary embodiment.

[0103] Please see Figures 2 to 5 The guided vehicle scanning and recognition process includes: 1. For example Figure 2 As shown, the method of the present invention is used to capture a moving vehicle.

[0104] 2. For example Figure 3 As shown, the algorithm is used to process the parameters of the moving vehicle.

[0105] 3. For example Figure 4 As shown, the algorithm processes the distance between moving vehicles and aircraft.

[0106] 4. For example Figure 5 As shown, the algorithm processes and matches the guide vehicle model library.

[0107] 5. Once identified as a guide vehicle, the berth guidance system normally guides the aircraft into its designated position, and the system operates according to the guidance procedure.

[0108] 6. If a foreign object is detected, the berth guidance system will issue a foreign object alarm.

[0109] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions 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 method for identifying fixed FOD on aprons based on lidar technology, characterized in that, Includes the following steps: Step S1: Acquire three-dimensional point cloud data of the apron area using a solid-state lidar fixedly installed on the apron high-mast light or the top of a building, with a scanning frequency of ≥100Hz. The three-dimensional point cloud data includes the three-dimensional coordinates and reflection intensity information of each point. Step S2: The acquired 3D point cloud data is subjected to noise filtering and ground segmentation in sequence. The noise filtering is used to remove dynamic interference points, and the ground segmentation is used to separate ground points from non-ground points. Step S3: Perform density clustering on the point cloud data of non-ground points to identify potential FOD object clusters; Step S4: Extract the feature vectors of potential FOD object clusters, determine whether they are FODs through a classification model, and perform confidence verification based on multi-frame time series data; Step S5: When a FOD that has passed the confidence verification is identified, trigger an audible and visual alarm and record the location, size, and time information; In step S2, the ground segmentation employs a RANSAC plane fitting algorithm weighted by reflection intensity, and its objective function is: In the formula, For point cloud indexing, For the first The reflection intensity weight at each point , , and These are the coefficients of the ground plane equation, obtained through fitting. , and The first Unit coordinates of each point in, , For the first The reflection intensity value at each point, This represents the maximum detectable reflection intensity of the system. This represents the threshold value for the intensity of oil stain reflection.

2. The method for identifying fixed FOD on aprons based on lidar technology according to claim 1, characterized in that, In step S2, the noise filtering employs a dynamic radius filtering algorithm, whose neighborhood radius... satisfy: In the formula, Based on the radius, and , The correlation coefficient of rainfall intensity and the implicit unit conversion factor are used to make The unit is , The range of its values ​​was determined through calibration experiments. , The rainfall intensity value is output in real time by the rain sensors deployed on the apron.

3. The method for identifying fixed FOD on aprons based on lidar technology according to claim 1, characterized in that, The oil stain reflection intensity threshold The calibration methods include: Step S2-1: Collect at least 1000 sample point clouds from typical oil-contaminated areas on the apron; Step S2-2: Calculate the histogram of reflectance intensity distribution for all sample points, and take the median value of the 85th-95th percentile interval as the median value. ; Step S2-3: Periodically update the threshold using point cloud data of manually labeled oil stain areas.

4. The method for identifying fixed FOD on aprons based on lidar technology according to claim 1, characterized in that, In step S3, density classification employs the multi-scale DBSCAN algorithm, whose core parameter is the neighborhood radius. satisfy: In the formula, Based on the neighborhood radius and , It is a negative scale correlation factor and , The estimated size of the current object cluster is obtained by calculating the minimum eigenvalue of the point cloud covariance matrix through principal component analysis.

5. The method for identifying fixed FOD on aprons based on lidar technology according to claim 4, characterized in that, The multi-scale DBSCAN algorithm is configured to stably detect FOD objects with a minimum size of 5mm, and The iterative calculation satisfies: In the formula, These are the minimum eigenvalues ​​of the point cloud covariance matrix.

6. The method for identifying fixed FOD on aprons based on lidar technology according to claim 1, characterized in that, In step S4, the feature vector is defined as: ;in, The longest axis length of the object is calculated by performing PCA decomposition on the object cluster point cloud and taking the direction of the eigenvector corresponding to the largest eigenvalue. The surface area of ​​the object is calculated using the point cloud convex hull algorithm. Let be the standard deviation of the point reflection intensity within the cluster of objects. This represents the centroid displacement of an object across three consecutive frames of point cloud data.

7. The method for identifying fixed FOD on aprons based on lidar technology according to claim 6, characterized in that, In step S4, the classification model employs a deep learning network based on the PointNet++ architecture, and its output probability is: In the formula, This is the predicted probability value for FOD. For activation function, The weights for the deep learning model were obtained through pre-training on the airport FOD dataset. For feature vectors The amount, Integer and , , , , , This is the model bias term.

8. The method for identifying fixed FOD on aprons based on lidar technology according to claim 1, characterized in that, In step S4, the confidence verification must satisfy: ;in, There is a comprehensive confidence level for FOD. For the first FOD probability of a frame For continuous detection of frames and , It is the time decay factor and Used to reduce the weight of historical frames. For historical frame time index, The timestamp of the current frame. The sampling interval for point cloud frames.

9. The method for identifying fixed FOD on an apron based on lidar technology according to any one of claims 1-8, characterized in that, The real-time processing of steps S1 to S4 is achieved through GPU parallel acceleration, specifically including: The point cloud clustering task is divided into thread blocks, and the size of each thread block is... satisfy: In the formula, This represents the total number of point clouds in a single frame. This refers to the number of GPU stream processors. The system's end-to-end processing latency is ≤200ms.

Images