Airport apron fixed FOD identification method based on laser radar technology

Abstract

The invention provides an airport apron fixed FOD identification method based on a laser radar technology, and relates to the technical field of FOD identification, and the method comprises the steps: obtaining the three-dimensional point cloud data of an airport apron region at a scanning frequency greater than or equal to 100 Hz through a solid-state laser radar fixedly installed on an airport apron high-pole lamp or the top of a building; sequentially carrying out noise filtering and ground segmentation on the acquired three-dimensional point cloud data; performing density clustering on the point cloud data of the non-ground points to identify a potential FOD object cluster; feature vectors of the potential FOD object clusters are extracted, whether the potential FOD object clusters are FOD or not is judged through a classification model, and confidence verification is carried out based on multi-frame time sequence data; and when the FOD passing the confidence verification is identified, triggering a sound-light alarm and recording position, size and time information. According to the method, noise and background interference can be effectively suppressed and a tiny target can be accurately extracted through a solid-state laser radar and a point cloud processing algorithm, and high-precision and low-false-alarm real-time detection on millimeter-level FOD in a complex environment can be realized.

Description

Technical Field

[0001] The present invention relates to the field of FOD identification technology, and in particular to a method for identifying apron-fixed FOD based on laser radar technology. Background Art

[0002] Foreign Object Debris (FOD) on airport aprons is a major threat to aviation safety. It includes objects such as metal parts, rubber fragments, and gravel. According to incomplete statistics, FOD causes direct losses of over $4 billion annually to global civil aviation, with indirect losses reaching four times that amount.

[0003] At present, there are the following technical bottlenecks in the field of FOD detection: First, there is a lack of millimeter-level FOD detection. Existing sensors cannot reliably identify dangerous objects (such as nuts and tire valve cores) sized 5 to 10 mm, which can cause tire bursts.

[0004] Second, there is insufficient suppression of complex background interference. The optical system is interfered with by oil reflections, and the radar system is affected by electromagnetic signals, resulting in a high false alarm rate in real-world scenarios on the apron.

[0005] Third, the full-time support capability is weak. Under extreme conditions such as nighttime, heavy rain, and high temperatures (for example, above 50°C), the existing FOD detection system has low reliability.

[0006] Fourth, maintenance costs are too high. Existing distributed sensors require road surface demolition for maintenance, and mechanical LiDAR requires regular calibration, both of which lead to heavy operational burdens and high maintenance costs. Summary of the Invention

[0007] In order to solve the problems in the related art, the present invention provides a method for identifying apron-fixed FOD based on lidar technology.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is: A method for identifying fixed FOD on an apron based on laser radar technology includes the following steps: Step S1: Using a solid-state laser radar fixedly mounted on a high-mast light on the apron or on top of a building, three-dimensional point cloud data of the apron area is acquired at a scanning frequency of ≥100 Hz. The three-dimensional point cloud data includes the three-dimensional coordinates and reflection intensity information of each point. Step S2: performing noise filtering and ground segmentation on the acquired three-dimensional point cloud data in sequence, wherein 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: performing density clustering on the point cloud data of non-ground points to identify potential FOD object clusters; Step S4: Extract the feature vector of the potential FOD object cluster, determine whether it is FOD through the classification model, and perform confidence verification based on multi-frame time series data; Step S5: When the FOD that has passed the confidence verification is identified, an audible and visual alarm is triggered and the location, size and time information are recorded.

[0009] Optionally, in step S2, the noise filtering adopts a dynamic radius filtering algorithm, and its neighborhood radius is satisfy: Where, is the base radius, and , is the rainfall intensity correlation coefficient and the implicit unit conversion factor is used to make The unit is , Determined by calibration test (the range of its value can be ), The rainfall intensity value is output in real time by the rain sensor deployed on the apron.

[0010] Optionally, in step S2, the ground segmentation adopts a reflection intensity weighted RANSAC plane fitting algorithm, and its optimization objective function is: Where, is the point cloud index, For the The reflection intensity weight of each point, 、 、 and are the ground plane equation coefficients, obtained by fitting, 、 and Respectively The unit coordinates of the points, in, , For the The reflection intensity value of the point, is the maximum reflection intensity that the system can detect, is the oil stain reflection intensity threshold.

[0011] Optionally, the oil stain reflection intensity threshold The calibration methods include: Step S2-1: Collect at least 1,000 sample point clouds in typical oil-contaminated areas on the apron; Step S2-2: Calculate the reflection intensity distribution histogram of all sample points and take the median value of the 85th-95th percentile interval as ; Step S2-3: Regularly update the threshold using manually labeled oil stain area point cloud data.

[0012] Optionally, in step S3, the density classification adopts the multi-scale DBSCAN algorithm, whose core parameter neighborhood radius is satisfy: Where, is the base neighborhood radius and , is a negative scale-correlation factor and , is the estimated size of the current object cluster, which 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: Where, are the minimum eigenvalues of the point cloud covariance matrix respectively.

[0014] Optionally, in step S4, the feature vector is defined as: in, is the longest axis length of the object. By performing PCA decomposition on the object cluster point cloud, the eigenvector direction corresponding to the maximum eigenvalue is calculated. is the surface area of the object, calculated by the point cloud convex hull algorithm, is the standard deviation of the reflection intensity of the points within the object cluster, It is the displacement of the center of mass of the object in three consecutive frames of point cloud.

[0015] Optionally, in step S4, the classification model adopts a deep learning network based on the PointNet++ architecture, and its output probability is: Where, is the predicted value of FOD probability, is the activation function, is the deep learning model weight, obtained by pre-training on the airport FOD dataset. is the eigenvector The weight, is an integer and , , , , , is the model bias term.

[0016] Optionally, in step S4, the confidence verification needs to satisfy: in, is the comprehensive confidence level of FOD existence, For the FOD probability of a frame, is the number of continuous detection frames and , is the time decay factor and , used to reduce the weight of historical frames, is the historical frame time index, is the current frame timestamp, is the point cloud frame sampling interval.

[0017] Optionally, the real-time processing of steps S1 to S4 is implemented by GPU parallel acceleration, specifically including: The point cloud clustering task is divided into thread blocks, each thread block size satisfy: Where, is the total number of point clouds in a single frame, is the number of GPU streaming multiprocessors; The system end-to-end processing delay is ≤200ms.

[0018] Beneficial effects: 1. Through the above-mentioned technical solution, the present invention utilizes a high-resolution solid-state lidar to provide basic data. It 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 small targets. High-dimensional feature extraction and a deep learning-based classifier are combined for precise identification. Rigorous multi-frame temporal confidence verification significantly reduces false alarms. Ultimately, this system achieves high-precision, low-false-alarm, real-time detection of millimeter-level FOD in complex environments (e.g., oil stains, rainfall) and extreme conditions (nighttime, high temperatures). Furthermore, the non-destructive fixed mounting 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 implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments. 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 paying any creative labor.

[0021] in: Figure 1 1. It is a schematic flow chart of the steps of a method for identifying fixed FOD on an apron based on laser radar technology provided by an exemplary embodiment of the present invention; Figure 2 is a schematic diagram of a capture guide vehicle provided by an exemplary embodiment; Figure 3 is a schematic diagram of processing moving vehicle parameters provided by an exemplary embodiment; Figure 4 is a schematic diagram of processing the distance between a moving vehicle and an aircraft provided by an exemplary embodiment; Figure 5 It is a schematic diagram of matching the algorithm processing provided by an exemplary embodiment with the guide vehicle model library. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0023] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0024] In order to facilitate relevant technical personnel to have a clearer and more accurate understanding of the technical solution of the present invention, the existing related technologies and the technical problems existing therein are further explained below.

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

[0026] 1. Manual inspection: This method relies on visual inspection by inspectors or handheld devices (such as magnetic wands and metal detectors). This method is inefficient (covering a small area per inspection) and is significantly affected by lighting and weather conditions. Furthermore, manual inspections require frequent inspections (e.g., every two hours), significantly increasing labor costs and inspection risks. Furthermore, inspections are prone to missed inspections at night or in complex areas (e.g., runway joints and shoulders).

[0027] 2. Cleaning vehicle assistance: Mechanical sweepers are used for runways / taxiways, but their removal rate for small metal objects (e.g., screws) or non-metallic objects (e.g., plastic, rubber, etc.) is low. In addition, they cannot achieve real-time monitoring and can only operate at relatively fixed time periods.

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

[0029] 1. Radar and optical fusion technology.

[0030] Millimeter-wave radar has advantages in that it can detect metal objects as small as 2 cm and exhibits strong stability in low-light environments. Its disadvantages include a low recognition rate for non-metallic materials (such as rubber and plastic) and susceptibility to interference from airport radio equipment (such as air traffic control radar and communication systems).

[0031] For infrared + visible light cameras, they can identify FOD through image recognition algorithms. However, their ability to cope with complex apron backgrounds (for example, oil stains, shadows, and low-light scenes) is relatively weak.

[0032] 2. Distributed sensor network.

[0033] It generally involves burying pressure / vibration sensors in the road surface to monitor abnormal vibration signals. However, its deployment cost is high and the maintenance operation is very complicated (the road surface needs to be demolished).

[0034] In summary, the problems existing in existing related technologies include the lack of millimeter-level FOD detection, insufficient suppression of complex background interference, weak full-time domain guarantee capabilities, and excessively high maintenance costs.

[0035] In light of this, the present invention provides a novel solution: a method for identifying fixed FOD on the apron based on LiDAR technology. This method aims to overcome the bottleneck of small-scale FOD detection and improve robustness in complex environments. It achieves stable FOD detection at the 5-10mm level, reduces false alarm rates in complex environments, and builds an easily maintainable detection system.

[0036] The technical solution of the present invention is 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 an apron based on laser radar technology, including the following steps: Step S1: Using a solid-state laser radar fixedly mounted on a high-mast light or on the top of a building, three-dimensional point cloud data of the apron area is acquired at a scanning frequency of ≥100 Hz. The three-dimensional point cloud data includes the three-dimensional coordinates and reflection intensity information of each point. Step S2: performing noise filtering and ground segmentation on the acquired 3D point cloud data in sequence, wherein noise filtering is used to remove dynamic interference points, and ground segmentation is used to separate ground points from non-ground points; Step S3: performing density clustering on the point cloud data of non-ground points to identify potential FOD object clusters; Step S4: Extract the feature vector of the potential FOD object cluster, determine whether it is FOD through the classification model, and perform confidence verification based on multi-frame time series data; Step S5: When the FOD that has passed the confidence verification is identified, an audible and visual alarm is triggered and the location, size and time information are recorded.

[0038] Through the above-mentioned technical solution, the present invention utilizes a high-resolution solid-state lidar to provide basic data. It then utilizes innovative point cloud processing algorithms (e.g., dynamic radius filtering, reflectance-weighted RANSAC ground segmentation, and adaptive multi-scale DBSCAN clustering, described below) to effectively suppress noise and background interference and accurately extract tiny targets. High-dimensional feature extraction and a deep learning-based classifier are then combined for precise identification. Rigorous multi-frame temporal confidence verification significantly reduces false alarms, ultimately achieving high-precision, low-false-alarm, real-time detection of millimeter-scale FOD in complex environments (e.g., oil stains, rainfall) and extreme conditions (nighttime, high temperatures). Furthermore, the non-destructive fixed mounting and solid-state sensor design significantly reduce the system's installation and maintenance costs.

[0039] Specifically, first, the present invention uses solid-state LiDAR for data acquisition, which has very high angular resolution (a few tenths of a degree or even higher) and fine range resolution (centimeter or millimeter level). This enables it to detect very small objects and generate a sufficiently dense point cloud to depict their outlines. Compared to traditional mechanical LiDAR, solid-state solutions offer more stable scanning, less point cloud jitter, and are more conducive to small target recognition. Furthermore, solid-state LiDAR scans at a high scanning frequency (≥100Hz). The extremely high frame rate (more than 100 scans per second) means that the system can "see" the same tiny object multiple times within a very short time interval. This greatly increases the probability of capturing tiny, low-reflectivity (possibly only detectable at certain angles) or partially obscured FOD in multiple consecutive frames, providing a foundation for the subsequent timing verification in step S4. In this way, the bottleneck of "lack of millimeter-level FOD detection" in existing related technologies can be specifically addressed.

[0040] Second, in step S2 of the present invention, the acquired 3D point cloud data is sequentially subjected to noise filtering and ground segmentation. This effectively suppresses false positives caused by oil reflections and other environmental clutter (non-FOD small objects). This effectively addresses the bottleneck of "inadequate suppression of complex background interference" in existing related technologies.

[0041] Third, the present invention utilizes solid-state LiDAR (an active sensor that doesn't rely on ambient light), making it highly adaptable to low-light environments. Solid-state LiDAR has no moving parts and generally offers superior high-temperature resistance to mechanical LiDAR. Furthermore, by using noise filtering to remove dynamic interference points, the method effectively suppresses raindrop noise interference, ensuring point cloud quality. This method can effectively improve detection reliability under extreme conditions such as nighttime, heavy rain, and high temperatures, addressing the bottleneck of existing technologies: weak full-time domain support capabilities.

[0042] Fourth, the solid-state LiDAR of the present invention is installed on high-pole lamps or building tops using a non-embedded installation method, completely avoiding the need to dismantle the pavement during maintenance. This not only greatly simplifies the installation process but also effectively reduces the initial installation cost and the high cost of subsequent pavement damage due to maintenance. Furthermore, since solid-state LiDAR lacks precision moving parts such as rotating motors, it not only offers higher reliability (fewer mechanical failure points) and a longer service life (less mechanical wear), but also lower maintenance requirements (no regular calibration is required, and maintenance is limited to cleaning and routine electrical inspections). Furthermore, compared to distributed sensor networks, solid-state LiDAR's data processing and transmission are simpler and more centralized, 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 of the present invention, noise filtering adopts a dynamic radius filtering algorithm, and its neighborhood radius is satisfy: Where, is the base radius, and , is the rainfall intensity correlation coefficient and the implicit unit conversion factor is used to make The unit is , Determined by calibration test (the range of its value can be ), The rainfall intensity value is output in real time by the rain sensor deployed on the apron.

[0044] This implementation significantly reduces the dense noise generated by raindrops in the point cloud during rainfall, preventing them from being mistakenly identified as potential FOD. Furthermore, while suppressing noise, it ensures that true FOD point clouds with a size of 5mm or greater are not over-filtered.

[0045] Second, the filtering intensity varies with the real-time rainfall intensity. Dynamic adjustment can automatically adapt to different rainfall intensity conditions without human intervention, and can effectively achieve adaptive environmental changes. At the same time, it can directly reduce system false alarms caused by meteorological interference and improve reliability in severe weather.

[0046] In this embodiment, it should be noted that First, in the algorithm formula provided by the present invention ( )middle, is the base radius ( ), this is when there is no rain or light rain ( The standard filter radius used under the condition of ≈0). It is calibrated based on the typical apron background noise (such as light dust, sensor background noise) and the minimum FOD size to be retained (5mm). Designed to protect against tiny FOD. It quantifies how much the filter radius needs to be increased per unit rainfall intensity increase to effectively suppress the new noise density under that rainfall intensity. It is determined through calibration tests (e.g. testing under different known rainfall intensities to find the optimal balance between noise filtering rate and micro FOD retention rate). value).

[0047] Second, rainfall interferes with point clouds in two ways: the laser beam emitted by the lidar is scattered and reflected when it encounters raindrops, appearing as a large number of discrete, short-lived, high-frequency noise points in the point cloud. The noise points are randomly distributed in the detection space, and a single raindrop usually produces one or a few isolated points. Moreover, the noise points appear and disappear very quickly (related to the falling speed of the raindrop). In addition, the rainfall intensity ( ), the larger the radius, the more raindrops there are per unit volume, and the higher the noise density. Based on this, traditional fixed-radius filtering algorithms may work well in light rain (sparse noise), but in heavy rain, if the filter radius is too small, dense noise points cannot be effectively filtered out (noise points may reach the "non-noise" density threshold within their small neighborhood). If the filter radius is too large, actual tiny FOD (such as a point cloud cluster of 5mm nuts) may be mistakenly filtered out as noise (because tiny FODs occupy a small space and have relatively few points, and the average density within a large neighborhood may be below the threshold).

[0048] In the present invention, the dynamic radius The principle of solving rainfall interference is: When it rains lightly ( smaller), , the noise is relatively sparse. In the neighborhood, the number N of other points (noise points or real points) around the noise point is usually lower than the threshold, so it can be effectively identified as noise filtering. In the neighborhood, since the distance between its points is very close and the density is high enough, N will be greater than the threshold and thus it will be retained.

[0049] When there is heavy rain or rainstorm ( Larger), The key principles include: Although the noise density is high, the spatial randomness is stronger: although the overall noise density increases sharply during heavy rain, due to the randomness of the raindrop falling position, the noise density in the expanded neighborhood is higher. In other words, within a large enough neighborhood, the spatial aggregation of noise points is insufficient, and their density may still be lower than that of real objects (such as FOD).

[0050] Increase the noise filtering threshold: Increase This means that the algorithm requires a point to find enough (N ≥ threshold) "partner points" in a larger spatial range in order not to be considered as noise. For isolated noise points or small clusters of noise generated by a single raindrop, after the expansion In the neighborhood, the probability of finding enough other noise points is reduced (because the noise points are randomly distributed), so it is easier to meet N < threshold and be filtered out.

[0051] Protecting real FOD: The point cloud of real FOD (even if it is tiny) is spatially continuous and tightly clustered. When increasing, as long as Not too big (this is determined by The reasonable range of the guarantee) covers the entire tiny FOD point cloud cluster. Then, when the points in the cluster are calculated in each other's neighborhood, their N values will still be very high (because the density inside the point cluster is high and continuous), so they will still be judged as valid points and retained. The lower limit (0.01m) and The upper limit (0.07) ensures that in extreme rainstorms, The growth will not be excessive enough to overwhelm the aggregation characteristics of the tiny FOD themselves.

[0052] In general, this implementation directly addresses the problems of "insufficient suppression of complex background interference" (especially meteorological interference) and "weak full-time domain guarantee capability" (rainstorm conditions) mentioned in the background technology, and provides an efficient and adaptive solution. It 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 of the present invention, ground segmentation adopts the reflection intensity weighted RANSAC plane fitting algorithm, and its optimization objective function is: Where, is the point cloud index, For the The reflection intensity weight of each point, 、 、 and are the ground plane equation coefficients, obtained by fitting, 、 and Respectively The unit coordinates of the points, in, , For the The reflection intensity value of the point, is the maximum reflection intensity that the system can detect, is the oil stain reflection intensity threshold.

[0054] In this embodiment, the present invention introduces dynamic weights based on reflection intensity into RANSAC plane fitting. accomplish: 1. Use the high reflective intensity of oil stains ( ) This optical feature distinguishes it from a clean ground; 2. Through low weight ( ) actively weakens the influence of high reflection points (suspected oil spots) in the least squares objective function. At the same time, through high weight ( ) strengthens the dominant role of low and medium reflection points (clean ground points) on the fitting results. In this way, 3 applies the above weights to the least squares distance sum of squares ( ), the algorithm can prioritize fitting the plane defined by clean ground points.

[0055] 3. The final fitted ground plane is not disturbed by oil stains and truly reflects the apron geometry; oil stains are more likely to be correctly segmented as non-ground points due to their low weights and large weighted distances (although they are not FOD, they avoid the more dangerous mistake of mistaking them for ground points). Subsequent clustering and classification steps can better utilize other features (such as size, shape, reflection uniformity, etc.) ) to distinguish the oily area from the real FOD.

[0056] In this embodiment, it should be noted that, first, the core of the optimization objective function of the present invention is to use weights to control the fitting direction. Specifically, the oil stain point has a high reflection intensity. ) This significant feature is achieved through weight Mathematically control its influence on the ground plane fitting process. At the same time, give low weight to oil spots ( ), actively reducing its contribution to the objective function. This makes the RANSAC algorithm no longer pursue the plane passing through these highlights or tolerate them deviating far from the plane when searching for the optimal plane. In addition, high weights are given to clean ground points ( ), forcing the algorithm to prioritize fitting the plane close to these points. This way, the final fitted ground plane is dominated by clean ground points and accurately reflects the apron's geometric surface. Highly reflective oil spots, even if located at ground level, are considered outliers by the algorithm due to their low weight and minimal contribution to the objective function, making them more likely to be segmented as non-ground points.

[0057] Second, for the weight of oil stain points ( ), its design principles include the following: First, this value is much smaller than 1, which significantly reduces its influence. At the same time, the oil stain weight cannot be set too small (for example, 0.01), otherwise it may introduce numerical instability. Moreover, when a small number of oil stains are mixed with a large number of ground points, the algorithm may completely ignore them, making their impact minimal. However, when large oil stains are present, too small a weight may cause the fitted plane to become completely uncontrollable. Therefore, after comprehensive consideration, the oil stain weight is set to 0.3.

[0058] In one embodiment of the present invention, the oil stain reflection intensity threshold of the present invention is Calibration methods can include: Step S2-1: Collect at least 1,000 sample point clouds in typical oil-contaminated areas on the apron; Step S2-2: Calculate the reflection intensity distribution histogram of all sample points and take the median value of the 85th-95th percentile interval as ; Step S2-3: Regularly update the threshold using manually labeled oil stain area point cloud data.

[0059] In this embodiment, the oil stain reflection intensity threshold is set. A specific calibration and update method is provided to solve the premise problem of the effectiveness of the reflection intensity weighted RANSAC algorithm, that is, how to scientifically determine the key threshold for distinguishing oil stains from the ground. Specifically, through this implementation, it is possible to ensure Accurately reflects the typical reflection intensity characteristics of oil stains on the apron of a specific airport, avoiding the failure of the weight strategy due to unreasonable threshold setting; at the same time, the system can adapt to the differences in reflection characteristics of different airports, different areas (such as cargo areas where oil stains may be heavier), and different periods (such as new pavement vs. old pavement); and, through a regular update mechanism, offsets the impact of factors such as pavement aging, changes in oil stain composition, and sensor performance drift on reflection intensity, ensuring the robustness of the system throughout its life cycle; in addition, accurate is the weight The correct allocation basis directly determines the accuracy of ground segmentation, thereby reducing FOD false alarms or omissions caused by misjudgment of oil stains.

[0060] In this embodiment, it should be noted that, first, for step S2-1, typical oil stain contamination areas are obtained, for example, parking spaces, taxiway vicinity, maintenance areas and other oil stain-prone areas are selected, covering different types of oil stains (for example, fresh, old, mixed) and ensuring that the sample size is large enough (at least 1000 sample point clouds) to reflect the statistical distribution characteristics of the oil stain reflection intensity. In this way, an oil stain reflection intensity database reflecting the actual conditions 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, the reflection intensity values of all sample points (from all point cloud frames collected in step S2-1) are Summarize and count the frequencies falling into different intensity intervals to form a reflection intensity distribution histogram; then analyze its distribution characteristics (real oil stains - mainly contribute to high reflection, oil stain edges / mixed points - slightly lower reflection intensity, background noise / non-oil stain points - even if collected in an oil stain area, it may contain a small amount of splashed non-oil stain points, such as points on the clean ground next to it, flying birds, sensor noise, etc., whose reflection intensity is low), and then select the median value of the 85th-95th percentile interval as .

[0062] Third, for step S2-3, by regularly updating the threshold, it is possible to: adapt to environmental changes (compensate for reflectance drift caused by aging and darkening of the pavement, sensor performance degradation / contamination, and the introduction of new oils), optimize regional differences (if a specific area, such as a freight area, has different oil stain characteristics, the threshold can be calibrated and updated separately for that area), and improve long-term accuracy (to ensure that the system remains accurate throughout its service life). Always accurately reflects the current "oil vs. ground" reflection intensity boundary).

[0063] In this embodiment, it should be noted that, first, the 85th-95th percentile interval is selected for the following reasons: 1. It avoids tail noise (less than the 85th percentile, i.e., less than P85): The low end of the distribution (less than P85) may contain a large number of non-oil stain points (noise, background ground points, and weak reflection points at the edge of the oil stain). Directly taking the minimum value or a low percentile (such as the P50 median) will be severely affected by these points, resulting in an excessively low threshold. 2. It avoids extreme outliers (greater than P95): The top end of the distribution (greater than P95) may contain a very small number of unusually bright points (such as specular reflection peaks, sensor saturation points, and metal reflections) that do not represent typical oil stains. Directly taking the maximum value or a high percentile (such as P99) will be affected by these points, resulting in an excessively high threshold. 3. It focuses on the core oil stain distribution (P85-P95): This interval is likely to contain the most typical and concentrated reflection intensity values of oil stain points. It removes interfering noise at the bottom and extreme outliers at the top.

[0064] Second, the reason for choosing the median of the interval is that the median is robust to possible asymmetric distribution or small fluctuations within the interval. Compared with the average, the median is less affected by extreme values in the interval. It is a stable statistic that can represent the reflection intensity level of the main part of the typical oil stain at the airport.

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

[0066] In this implementation, density clustering is improved by introducing the multi-scale DBSCAN algorithm and its core parameter neighborhood radius The adaptive formula focuses on solving the clustering failure problem caused by large differences in FOD size. Specifically, it can not only stably detect small FOD (such as nuts) at the millimeter level (≥5mm) and large FOD (such as tool kits and gravel) at the centimeter / decimeter level in a single cluster, avoiding the failure of traditional single-scale clustering algorithms at the extreme ends of size, but also ensure that the point cloud clusters of tiny FOD are not "swallowed" or "ignored" by large objects or background noise, significantly improving the detection rate of 5mm-level targets. In addition, it can also prevent the point cloud of large FOD from being over-segmented into multiple scattered small clusters, ensuring the integrity of the object and facilitating subsequent feature extraction and classification. This improves the robustness to FOD of different sizes and reduces missed detections due to size factors.

[0067] In this embodiment, it should be noted that, first, for the DBSCAN algorithm, it is a density-based clustering algorithm, and the core parameter is the neighborhood radius. And the minimum number of points MinPts. It classifies the points in the point cloud that meet the following conditions into one category: core points (with this point as the center, The neighborhood of the radius contains at least MinPts points), boundary points (in the core point Neighborhood, but the number of points in its own neighborhood is less than MinPts) and noise points (neither core points nor boundary points). The current DBSCAN algorithm is generally fixed , while detecting tiny FOD requires small to focus on local high-density areas; detecting large FOD requires large To connect relatively sparse internal points. It is impossible to satisfy both extreme demands simultaneously.

[0068] Based on this, the present invention proposes a dynamic adaptive neighborhood radius To solve the above problem. Specifically, the process of the DBSCAN algorithm of the present invention can be: 1. Preliminary segmentation / preclustering: You can first use a very loose (large) fixed Alternatively, the point cloud can be divided into multiple candidate regions or over-segmented blocks (Supervoxels) by performing preliminary grouping based on spatial location. Each candidate region contains one or more potential objects or a part of the background.

[0069] 2. For each candidate area: A. Calculation :Perform PCA on all points in the region and calculate ; B. Calculation Dynamics :By substituting into the formula Computational Dynamics ; C. Apply DBSCAN: Only within the candidate area, use the calculated dynamic Perform standard DBSCAN clustering with the preset MinPts. The clustering results are finer object clusters identified at this regional scale.

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

[0071] Second, the present invention is adaptively How to solve the size dilemma (fixed as described in the first point above It is impossible to meet both extreme requirements at the same time).

[0072] When dealing with tiny FOD, Smaller → is a negative number with a small absolute value → → , at this time, in this small area, use a smaller Run DBSCAN. Smaller It is just right for focusing on local high-density point clouds of tiny objects. Background noise points or slightly distant points cannot be included in the neighborhood. The points of tiny FOD can effectively meet the core point conditions and are correctly identified as an independent cluster, which can effectively ensure that tiny FOD is not submerged.

[0073] When dealing with large FOD, Larger → is a negative number with a large absolute value → Significantly less than 1 → , at this time, in this large area, use a relatively smaller ( ) Run DBSCAN. In the candidate area of the large object that has been preliminarily segmented ( larger), use a finer scale ( This not only allows for more accurate identification of possible structural boundaries or holes within an object (although they may still be merged into a single object), but also, most importantly, effectively avoids the erroneous merging of other independent objects or background noise adjacent to the large FOD. Follow Increase and actively reduce, improve the accuracy of large object boundary segmentation, and prevent the generation of "super large clusters". In this way, for large objects themselves, as long as the internal point spacing is smaller than this smaller (Usually, the surface point spacing of large objects is relatively uniform and smaller than their size), they can still be aggregated into a complete cluster through the connection relationship between core points and boundary points, which can effectively prevent large FOD from being split.

[0074] That is, for small objects use smaller , to focus on local high density and protect tiny FOD from being drowned by noise or missed. For large objects, a smaller , so as to use a finer scale within the target area, preventing large FOD from excessively merging adjacent objects or backgrounds while ensuring its own integrity is not fragmented.

[0075] Third, in this embodiment, Set to a negative value so that Able to follow The increase of the Focusing, large objects use smaller Prevent overflow. At the same time, It ensures that this scaling is effective but 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: Where, are the minimum eigenvalues of the point cloud covariance matrix respectively.

[0077] In this embodiment, the present invention defines The calculation method of the adaptive neighborhood radius is proposed, and its connection with the millimeter-level FOD detection capability is emphasized to solve the key input problem of the adaptive neighborhood radius calculation, that is, how to scientifically and stably quantify the spatial scale of the object cluster. Specifically, it can not only provide an estimate of the physical size of the object in the most compact dimension ( ), is the dynamic neighborhood radius of multi-scale DBSCAN ( ) calculations, and also ensures that the algorithm has stable size perception capabilities for tiny objects ≥5mm, which is the mathematical basis for their 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 embodiment, it is necessary to explain why, first, And how to define it. This is due to the dynamic neighborhood radius Calculation formula ( ) is the core of It can accurately reflect the physical size of the current point cloud cluster, which is the core input for implementing "size-aware" adaptive clustering. However, since the point cloud cluster is a collection of discrete points, how to use a scalar ( ) to effectively characterize its spatial scale, especially for small and irregular FOD? Then, we need The following four requirements must be met: clear physical meaning (can intuitively correspond to the geometric dimensions of the object, such as length, width, thickness, etc.), stable (insensitive to point cloud sampling density, noise, slight movement or rotation of the object), direction-independence (the result does not depend on the orientation of the object relative to the radar), and computationally efficient (able to meet real-time requirements).

[0079] The following first explains 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 that the point cloud cluster currently 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 center of mass in three-dimensional space and the correlation between changes in different coordinate axis directions. Its diagonal elements represent the variance (dispersion) of the points in the x, y, and z directions, respectively, while the off-diagonal elements represent the covariance (correlation) between different directions.

[0081] 2. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C: , Get three eigenvalues 、 and ( ) and the corresponding eigenvector 、 and ; Among them, for the eigenvalue 、 and In terms of , it represents the variance of the point cloud in the direction of the corresponding eigenvector. Maximum, indicating that the point cloud has the highest degree of dispersion in the direction of the longest axis; Minimum (i.e. ), indicating that the point cloud has the lowest degree of discreteness in the most compact direction.

[0082] Thus, in a geometric sense, It can be roughly understood as the "radius" or "thickness" of the point cloud in its most compact dimension. At the same time, the eigenvalues / vectors calculated by PCA are based on the statistical distribution of the point cloud itself and have nothing to do with the rotation angle of the object in the radar coordinate system. No matter how the FOD is placed, the calculated and Both can reflect their inherent minimum dimensions.

[0083] The following are the reasons why you choose instead of Or average value for explanation.

[0084] first, Directly characterizes the size of an object in the smallest dimension. This is crucial for identifying thin objects (such as rubber, which has a small thickness but a large area) or the cross-section of a slender object (such as the diameter of a screw). (maximum size), a thin and long wire may be mistakenly regarded as having a large size, resulting in its neighborhood radius ϵ being incorrectly reduced, which is not conducive to the complete clustering of its own point cloud.

[0085] Second, point clouds with small FOD are usually more isotropic (spherical) or significantly compressed in one dimension. is most sensitive to this compression dimension. Compared to the largest size ( ) or average size, which can more stably reflect the key scale that limits the detection of tiny FOD.

[0086] Finally, the goal of the dynamic neighborhood radius ϵ is to protect tiny FODs (need a small ϵ) and prevent large FODs from over-merging (need an even smaller ϵ). Can effectively distinguish tiny ( Small) and Large ( Large) objects and provide large objects with the scale information of their thinnest point, which is the key input needed to prevent over-merging (the neighborhood radius should be set with reference to its thinnest point to avoid crossing into neighboring objects).

[0087] In one embodiment of the present invention, in step S4 of the present invention, the feature vector is defined as: in, is the longest axis length of the object. By performing PCA decomposition on the object cluster point cloud, the eigenvector direction corresponding to the maximum eigenvalue is calculated. is the surface area of the object, calculated by the point cloud convex hull algorithm, is the standard deviation of the reflection intensity of the points within the object cluster, It is the displacement of the center of mass of the object in three consecutive frames of point cloud.

[0088] In this embodiment, the present invention provides a multi-dimensional feature vector by defining the key features extraction step. Focus on solving the core problem of accurate FOD identification and interference discrimination. Specifically, the maximum size of the comprehensive object ( ), spatial range ( ), reflection uniformity ( ), short-term exercise ( ) four key discriminant features, integrating multi-dimensional information of geometry, material, and motion, significantly improving the ability to distinguish FOD from common interference objects (oil stains, reflective marks, fallen leaves, flying birds), among which, In order to deal with the interference of oil stains / water accumulation, the reflection is highly uniform ( Very low) characteristics, and the real FOD (even if the material is uniform and the size is small, it may be caused by discrete points will not be extremely low) and differences in complex formation; For dynamic interference. Using real FOD absolute fixation ( ) characteristics, reliably distinguish between short-term objects such as flying birds, fallen leaves, and plastic bags ( ).

[0089] This implementation is the core innovation of the present invention at the feature engineering level, and provides a systematic solution to the high false alarm rate problem caused by "insufficient suppression of complex background interference" in the background technology. and Features, respectively from the optical properties of materials and motion characteristics, two dimensions that are easily overlooked by traditional methods, can significantly improve the robustness and reliability in real complex apron environments, and are the key to achieving high-precision, low-false-alarm FOD detection.

[0090] In one embodiment of the present invention, in step S4 of the present invention, the classification model adopts a deep learning network based on the PointNet++ architecture, and its output probability is: Where, is the predicted value of FOD probability, is the activation function, is the deep learning model weight, obtained by pre-training on the airport FOD dataset. is the eigenvector The weight, is an integer and , , , , , is the model bias term.

[0091] In this implementation, the classification model is defined, a deep learning network based on the PointNet++ architecture is used, and its output probability is given. The calculation formula is used to solve the problem of insufficient generalization ability of traditional classification methods in complex FOD discrimination. Specifically, PointNet++, a network designed for point clouds, is used. It has permutation invariance and hierarchical feature extraction capabilities, and can efficiently process point clouds or their extracted features F. At the same time, the network learns weights , automatically assign different features in the feature vector ( ) Adaptive discriminant importance, and capture its complex nonlinear combination relationship, and through the Sigmoid activation function Output intuitive FOD probability value , supporting subsequent confidence verification; in addition, based on the pre-training mechanism of the airport FOD dataset, the model deeply internalizes the discrimination knowledge in specific scenarios, significantly improving the discrimination accuracy and generalization ability of complex interference (especially oil stains and dynamic object afterimages).

[0092] In this embodiment, for the output probability calculation formula ( ), among which, The term is a weighted linear combination of the four features, which can be regarded as the projection of the best linear discriminant direction learned by the network on the feature space. Model bias term It is also a learned parameter. It represents the The basic logit value when it is 0 (or mean) affects the overall tendency of the classifier (such as the proportion of FOD samples in the dataset). The term is the result of the weighted sum mapped to the interval [0,1], which is intuitively interpreted as "the probability that the object is FOD" .

[0093] This gives the output a clear probabilistic meaning, facilitating subsequent confidence threshold setting. Furthermore, it provides a soft metric for decision making, which better reflects uncertainty than hard labels (0 / 1) and facilitates time series fusion. Furthermore, the saturation property of the Sigmoid algorithm (small gradients at both ends) provides a degree of robustness to anomalous inputs.

[0094] The formula for this implementation describes the operation of the last layer (or decision layer) of the PointNet++ network. The previous layers of the network (multiple SA layers, FP layers, and MLP layers) perform highly nonlinear transformations and abstractions on the input feature vector F, learning a richer feature representation. Ultimately, these high-level features are compressed (flattened) or pooled (pooled) into a vector, and then mapped to a single node (Logit value) through a fully connected layer (Fully Connected Layer). This Logit value is the value in the formula (Note: Here In fact, it is the abstract features that the network finally learns, 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 relationship. Finally, the activation function is applied to obtain the probability The key point is the weight and It is learned end-to-end from the data rather than manually set.

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

[0096] In this embodiment, the present invention precisely defines the confidence verification steps and proposes a multi-frame time sequence confidence verification formula to focus on solving the false alarm problem caused by the uncertainty of single frame detection. Specifically, the present invention uses ≥3 frames of continuous scanning data ( ), integrate the time dimension information to achieve time series information fusion; and through the exponential decay weight ( ) gives higher weights to the current frame and adjacent frames to ensure that the system is sensitive to the latest state, so as to achieve attenuated weighted emphasis on the recent period; at the same time, a high threshold of >0.8 is set for the weighted average confidence Confidence, requiring the target to continuously and stably exhibit high FOD probability characteristics in multiple frames. In general, this implementation method takes advantage of the high-frequency scanning of the lidar ≥100Hz, combined with the probabilistic results output by the deep learning classifier, and through rigorous time series statistical verification, fundamentally eliminates the vast majority of false alarms caused by instantaneous interference and short-staying objects, and can effectively solve the problem of high false alarm rate caused by "insufficient suppression of complex background interference" in background technology. Its parameter design ( 、 and thresholds) achieves the optimal balance between real-time performance (≤200ms) and reliability, which is the core guarantee for the system to meet the stringent safety 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 to perform timing verification, balancing real-time performance and reliability. Exponential decay weighting term ( ) for different historical frames Assign different weights; weighted average and normalization term ( ) First use The FOD probabilities of N consecutive frames are weighted and summed with time decay, then normalized by dividing by N to ensure that the output confidence score (Confidence) falls within a reasonable probability range for comparison with the threshold of 0.8. This way, Confidence comprehensively reflects the credibility of the target's "continuous FOD" across multiple recent frames. It is not just a simple average, but rather emphasizes the importance of recent evidence.

[0098] In one embodiment of the present invention, the real-time processing of steps S1 to S4 of the present invention is implemented by GPU parallel acceleration, specifically including: The point cloud clustering task is divided into thread blocks, each thread block size satisfy: Where, is the total number of point clouds in a single frame, is the number of GPU streaming multiprocessors; The system end-to-end processing delay is ≤200ms.

[0099] In this implementation, the overall processing flow from step S1 to step S4 is critically limited to real-time performance. 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 delay and maintenance cost are resolved. Specifically, the GPU is used as the core computing power carrier, leveraging its massively parallel architecture and high memory bandwidth to violently accelerate computationally intensive tasks. At the same time, the load balancing formula ( ) ensures that tasks are evenly distributed across all GPU stream processors (SMs), and that each SM is loaded with ≈128 threads (close to the hardware concurrency limit), maximizing resource utilization. Furthermore, through a combination of algorithm optimization, parallel computing, pipeline processing, asynchronous I / O, and other means, the end-to-end latency can be compressed to ≤200ms, meeting the real-time requirements of safety-critical systems.

[0100] In this embodiment, it should be noted that for the load balancing formula ( ), it aims to maximize GPU utilization: 1. Evenly divide the single-frame point cloud processing task into all SMs; 2. Ensure that each SM is assigned a load of about 128 threads (close to its maximum parallel capability).

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

[0102] The following briefly describes the process of scanning and identifying a guide vehicle using the method of the present invention in conjunction with an exemplary embodiment.

[0103] See also Figures 2 to 5 The guide vehicle scanning and identification process includes: 1. If Figure 2 As shown, the method of the present invention is applied to capture a moving vehicle.

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

[0105] 3. If Figure 4 As shown, the algorithm processes the distance between the moving vehicle and the aircraft.

[0106] 4. If Figure 5 As shown, the algorithm processing is matched with the guide vehicle model library.

[0107] 5. Identified as a guide vehicle, the berth guidance system guides the aircraft into position normally, and the system operates according to the guidance process.

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

[0109] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. The method for identifying fixed FOD on the apron based on laser radar technology is characterized by: The steps include: Step S1: Using a solid-state laser radar fixedly mounted on a high-mast light on the apron or on top of a building, three-dimensional point cloud data of the apron area is acquired at a scanning frequency of ≥100 Hz. The three-dimensional point cloud data includes the three-dimensional coordinates and reflection intensity information of each point. Step S2: performing noise filtering and ground segmentation on the acquired three-dimensional point cloud data in sequence, wherein 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: performing density clustering on the point cloud data of non-ground points to identify potential FOD object clusters; Step S4: Extract the feature vector of the potential FOD object cluster, determine whether it is FOD through the classification model, and perform confidence verification based on multi-frame time series data; Step S5: When the FOD that has passed the confidence verification is identified, an audible and visual alarm is triggered and the location, size and time information are recorded.

2. The method for identifying apron-fixed FOD based on laser radar technology according to claim 1 is characterized in that: In step S2, the noise filtering adopts a dynamic radius filtering algorithm, and its neighborhood radius is satisfy: Where, is the base radius, and , is the rainfall intensity correlation coefficient and the implicit unit conversion factor is used to make The unit is , Determined by calibration test (the range of its value can be ), The rainfall intensity value is output in real time by the rain sensor deployed on the apron.

3. The method for identifying apron-fixed FOD based on laser radar technology according to claim 1 is characterized in that: In step S2, the ground segmentation adopts the reflection intensity weighted RANSAC plane fitting algorithm, and its optimization objective function is: Where, is the point cloud index, For the The reflection intensity weight of each point, 、 、 and are the ground plane equation coefficients, obtained by fitting, 、 and Respectively The unit coordinates of the points, in, , For the The reflection intensity value of the point, is the maximum reflection intensity that the system can detect, is the oil stain reflection intensity threshold.

4. The method for identifying apron-fixed FOD based on laser radar technology according to claim 3 is characterized in that: The oil stain reflection intensity threshold The calibration methods include: Step S2-1: Collect at least 1,000 sample point clouds in typical oil-contaminated areas on the apron; Step S2-2: Calculate the reflection intensity distribution histogram of all sample points and take the median value of the 85th-95th percentile interval as ; Step S2-3: Regularly update the threshold using manually labeled oil stain area point cloud data.

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

6. The method for identifying apron-fixed FOD based on laser radar technology according to claim 5 is characterized in that: The multi-scale DBSCAN algorithm is configured to stably detect FOD objects with a minimum size of 5 mm, and The iterative calculation satisfies: Where, are the minimum eigenvalues of the point cloud covariance matrix respectively.

7. The method for identifying fixed FOD on the apron based on laser radar technology according to claim 1 is characterized in that: In step S4, the feature vector is defined as: in, is the longest axis length of the object. By performing PCA decomposition on the object cluster point cloud, the eigenvector direction corresponding to the maximum eigenvalue is calculated. is the surface area of the object, calculated by the point cloud convex hull algorithm, is the standard deviation of the reflection intensity of the points within the object cluster, It is the displacement of the center of mass of the object in three consecutive frames of point cloud.

8. The method for identifying fixed FOD on the apron based on laser radar technology according to claim 7 is characterized in that: In step S4, the classification model adopts a deep learning network based on the PointNet++ architecture, and its output probability is: Where, is the predicted value of FOD probability, is the activation function, is the deep learning model weight, obtained by pre-training on the airport FOD dataset. is the eigenvector The weight, is an integer and , , , , , is the model bias term.

9. The method for identifying fixed FOD on the apron based on laser radar technology according to claim 1 is characterized in that: In step S4, the confidence verification must satisfy: in, is the comprehensive confidence level of FOD existence, For the FOD probability of a frame, is the number of continuous detection frames and , is the time decay factor and , used to reduce the weight of historical frames, is the historical frame time index, is the current frame timestamp, is the point cloud frame sampling interval.

10. The method for identifying fixed FOD on the apron based on laser radar technology according to any one of claims 1 to 9, 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, each thread block size satisfy: Where, is the total number of point clouds in a single frame, is the number of GPU streaming multiprocessors; The system end-to-end processing delay is ≤200ms.

Images