A method for fabricating solid-state particles in a single-line lidar

Abstract

The application discloses a kind of single-line laser radar solid particle simulation method, belong to involve laser radar data processing technical field, including the following steps: step S1, single-line laser radar obtains original point cloud data according to flight time, according to the feature model set, extract feature information, send to particle feature analysis module and analyze, including: reflection intensity analysis;Point cloud density analysis;Motion trajectory analysis;Step S2, particle determination: based on the characteristic value of step S1, calculate particle probability P by weighted fusion, and according to P value range mark particle group or real obstacle;Step S3, dynamic parameter adjustment: according to the particle concentration of environment, self-adapting adjustment motion threshold in motion trajectory analysis;Step S4, false alarm suppression: remove the point cloud marked as particle group, output the real obstacle data after purification.By analyzing target reflection intensity, motion mode, spatial distribution characteristics, dynamically distinguish particle group and real obstacle, significantly reduce the misjudgment rate.

Description

Technical Field

[0001] This invention relates to the field of lidar data processing technology, specifically to a method for fabricating solid-state particles in a single-line lidar system. Background Technology

[0002] Single-line lidar is a sensor that acquires two-dimensional planar distance data through rotating scanning with a single laser beam. It is widely used in AGV navigation, equipment safety protection, and personnel intrusion detection. However, in high-particle-concentration environments, such as metal processing workshops, welding areas, and oil mist environments, suspended solid particles (typically less than 5mm in diameter, such as metal shavings, dust, rain, and oil) can reflect laser signals and be easily misidentified by single-line lidar as obstacles (such as personnel intrusion), leading to erroneous equipment shutdowns or AGV obstacle avoidance errors. Traditional filtering algorithms struggle to distinguish between dynamic particles and real obstacles.

[0003] Traditional solutions have the following drawbacks:

[0004] 1. Static threshold filtering: This method ignores points with signal strength below a threshold, but it cannot distinguish between weakly reflective obstacles and particles. Random particle reflections are identified as "instantaneous obstacles," resulting in a high false alarm rate. Furthermore, a fixed threshold cannot handle dynamic changes in particle concentration, exhibiting poor adaptability.

[0005] 2. Multi-frame matching method: It relies on the continuity of the target's motion trajectory, but the random motion of particles causes the algorithm to fail.

[0006] 3. Traditional clustering algorithms: consume a lot of computing power and are difficult to meet real-time requirements. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for fabricating solid-state particles in a single-line lidar system.

[0008] The present invention employs the following technical means:

[0009] A method for fabricating solid-state particles in a single-line lidar system includes the following steps:

[0010] Step S1: The single-line lidar acquires raw point cloud data based on the flight time, extracts feature information according to the set feature model, and sends it to the particle feature analysis module for analysis, including:

[0011] Reflection intensity analysis; calculate the standard deviation of signal intensity of all signal points within the scanning window, and quantify the reflection intensity dispersion; if the reflection intensity dispersion is higher than the preset maximum dispersion, it is marked as a particle group;

[0012] Point cloud density analysis; detect the point cloud density abrupt change rate Δρ in adjacent scanned regions, and verify and label it in conjunction with the convex hull bounding box area S;

[0013] Motion trajectory analysis; calculation of the angle between displacement vectors of adjacent frames, and derivation of the motion continuity index η;

[0014] Step S2, Particle determination: Based on the feature values ​​of step S1, the particle probability P is calculated by weighted fusion, and the particle group or real obstacle is marked according to the range of P value.

[0015] Step S3, Dynamic parameter adjustment: Adaptively adjust the motion threshold in motion trajectory analysis according to the particle concentration in the environment;

[0016] Step S4, False Alarm Suppression: Remove the point cloud labeled as particle swarm and output the cleaned-up real obstacle data.

[0017] Furthermore, the reflection intensity analysis in step S1 includes:

[0018] Input the reflection intensity values ​​of all signal points within the scanning window, and calculate the average reflection intensity. Where N is the number of all signal points within the scanning window; i It is the reflection intensity value of the i-th signal point, which ranges from 0 to 255;

[0019] Calculate the standard deviation of reflection intensity ;

[0020] Calculate the normalized reflection intensity dispersion ; where σ max It is the preset maximum standard deviation;

[0021] If the normalized reflection intensity dispersion σ norm If the value is greater than 0.8, it is marked as a particle group.

[0022] Furthermore, the point cloud density analysis described in step S1 includes:

[0023] Determine the number of point clouds in the unit scan area and adjacent areas;

[0024] Calculate the point cloud density abrupt change rate Δρ = |ρ k - ρ k-1 | / ρ k-1 × 100%; ρ k The number of point clouds in the k-th unit scan region;

[0025] Verify and label the convex hull bounding box area S. If Δρ > 50% and the convex hull bounding box area S < 0.1m², then label it as a particle group.

[0026] Furthermore, the motion trajectory analysis described in step S1 includes:

[0027] Through target clustering tracking, the same target is tracked across frames, and the displacement direction angle is recorded.

[0028] Calculate the motion continuity index ; where, θ1 is the displacement direction angle of the previous frame, and θ2 is the displacement direction angle of the current frame.

[0029] If the motion continuity index η < the motion threshold, it is determined as random motion and marked as a particle swarm.

[0030] Furthermore, in step S2, the reflectance intensity dispersion σ norm , the point cloud density change rate Δρ, and the motion discontinuity (1 - η) are normalized for eigenvalues:

[0031] Particle probability calculation: , where w1, w2, and w3 are preset weights; the determination rules include:

[0032] If P ≤ 0.4, it is directly marked as a real obstacle;

[0033] If 0.4 < P ≤ 0.7, cross-frame verification is triggered. If the P value is < 0.7 for three consecutive frames, it is marked as a real obstacle;

[0034] If P > 0.7, it is directly marked as a particle swarm.

[0035] Furthermore, the dynamic parameter adjustment in step S3 includes:

[0036] Calculate the environmental particle concentration C 颗粒 = (number of particle marked points / total number of point cloud points) × 100%;

[0037] If C 颗粒 > 30%, switch to the loose mode, reduce the motion threshold of the motion continuity index η; and return to step S1 to re - analyze the motion trajectory.

[0038] Furthermore, the false alarm suppression in step S4 includes:

[0039] Remove the point cloud marked as a particle swarm: Delete the points marked as a particle swarm from the point cloud set;

[0040] Output the real obstacle data, including the real obstacle coordinates and motion vectors.

[0041] A method for simulating solid particles by a single - line lidar, which dynamically distinguishes particle swarms from real obstacles by analyzing target reflectance intensity, motion patterns, and spatial distribution characteristics, significantly reducing the misjudgment rate; it has the following effects:

[0042] One, the multi - feature fusion mechanism solves traditional defects.

[0043] Traditional limitations: Traditional static threshold filtering cannot distinguish between weakly reflective obstacles and particles, and multi-frame matching methods fail due to the random motion of particles.

[0044] This scheme, 1. Reflection intensity analysis: Based on the reflection intensity dispersion σ... norm 1. Quantify particle group reflection instability; 2. Point cloud density abrupt change + convex hull verification: identify contourless targets; 3. Motion continuity index: distinguish Brownian motion (η<0.3) from continuous motion (η=0.577→obstacles).

[0045] Results: The false alarm rate decreased from 18.2% to 1.3%.

[0046] Second, dynamic thresholds enhance environmental adaptability.

[0047] Traditional drawback: Fixed thresholds cannot cope with dynamic changes in particle concentration (such as oil mist injection causing an instantaneous spike in concentration to 33.7%).

[0048] This plan, 1, is based on C. 颗粒 1. Adjust the motion threshold of the motion continuity index η in real time; 2. Trigger cross-frame verification by weighted particle probability P (P=0.427→3 consecutive frames verification).

[0049] Results: False alarm rate fluctuations were controlled within ±0.2% (for the fixed threshold scheme, the fluctuations reached ±15%).

[0050] Third, lightweight design ensures real-time performance.

[0051] Traditional drawbacks: Traditional clustering algorithms consume a lot of computing power and are difficult to meet the real-time requirements of 25Hz scanning.

[0052] This solution achieves the following: 1. Algorithm complexity optimized to O(N); 2. Embedded deployment (NVIDIA Jetson TX2) with single-frame processing latency <10ms.

[0053] Results: CPU usage <15%, supports high-precision scenes with a 270° scanning angle and 0.25° resolution. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the principle of the present invention;

[0055] Figure 2 This is a flowchart of the reflection intensity analysis. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Terminology explanation.

[0058] Suppression algorithm: An algorithm that suppresses specific interference signals through data processing.

[0059] Point cloud density: The number of valid data points returned by the lidar per unit angle.

[0060] Motion continuity index: A quantification of the smoothness of the motion trajectory between adjacent frames of the target.

[0061] In this scheme, the single-line lidar acquires raw point cloud data based on flight time, extracts feature information according to the set feature model, and sends it to the particle feature analysis module for analysis, including reflection intensity analysis, point cloud density analysis, and motion trajectory analysis. After comprehensive judgment, it is marked as a particle group for simulation. If it is a real obstacle, it performs corresponding protection processing (stopping / obstacle avoidance / optimizing navigation route, etc.).

[0062] Figure 1 This is a schematic diagram illustrating the principle of the present invention, such as... Figure 1 As shown, a method for fabricating solid-state particles in a single-line lidar system includes the following steps:

[0063] Step S1, dynamic feature extraction.

[0064] This step extracts key features from the raw point cloud data, providing a foundation for particle identification. The input is the raw point cloud dataset acquired by LiDAR, and the output is quantified feature values: reflection intensity dispersion, point cloud density change rate, and motion continuity index. Feature extraction is based on a sliding window mechanism, with a configurable window size (e.g., a 30° sector area corresponding to 120 data points).

[0065] Step S101: Reflection intensity analysis: Calculate the standard deviation of the signal intensity of all signal points within the scanning window and quantify the reflection intensity dispersion; if the reflection intensity dispersion is higher than the preset maximum dispersion, it is marked as a particle group. Figure 2 This is a flowchart of the reflection intensity analysis, such as... Figure 2 As shown.

[0066] Solid particles, due to their varying size and material, typically exhibit discrete and unstable reflected signal intensity, while real obstacles (such as metal equipment) show more uniform reflected intensity. This solution effectively distinguishes weakly reflective targets (particles) from real obstacles through reflection intensity analysis, avoiding the limitations of static threshold filtering. It also boasts low computational complexity (only O(N) time per window), making it suitable for real-time processing.

[0067] Example: AGV protection system in a turning workshop.

[0068] Hardware configuration: LiDAR: S300-C-3 (270° scanning angle, 0.25° resolution); Processor: NVIDIA Jetson TX2 (ARM Cortex-A57);

[0069] Environmental parameters: Metal scrap density: 150-200 pieces / m³ (diameter 0.5-2mm); Scanning frequency: 25Hz (40ms per frame).

[0070] In this example, the window size is a 30° sector area (corresponding to 120 data points, with a resolution of 0.25°).

[0071] Table 1 shows examples of point clouds within the window in this embodiment (reflection intensity values ​​0-255).

[0072]

[0073] Setting a 30° sector region balances feature sensitivity and computational efficiency: a window that is too small (<10°) is susceptible to noise interference; a window that is too large (>60°) will smooth out local features.

[0074] The steps for extracting the reflection intensity dispersion, as described in the example, are as follows:

[0075] Step S101a: Input the reflection intensity values ​​of all signal points within the scanning window, and calculate the average reflection intensity μ.

[0076] The formula for calculating the mean reflection intensity μ is: Where N is the number of all signal points within the scanning window, which is 120 in this embodiment; i It is the reflection intensity value of the i-th signal point, which ranges from 0 to 255.

[0077] The mean reflection intensity μ reflects the central tendency of signal intensity within the window. A low mean value may indicate a weakly reflective target (such as a dark obstacle or distant particles).

[0078] Step S101b: Calculate the standard deviation σ of the reflection intensity.

[0079] The formula for calculating the standard deviation σ of reflection intensity is: .

[0080] The standard deviation σ of the reflection intensity quantifies the dispersion of the signal intensity. Particle groups, due to their irregular shapes, exhibit large fluctuations in reflection intensity (high σ value), while flat obstacles (such as walls) show stable reflection intensity (low σ value).

[0081] Step S101c: Calculate the normalized reflection intensity dispersion σ norm .

[0082] Normalized reflection intensity dispersion σ norm The calculation formula is: ; where σ max It is a preset maximum standard deviation, based on experimental calibration presets. In this embodiment, σ max It is 40.

[0083] Normalized reflection intensity dispersion σ norm Normalization eliminates the influence of dimensions, compressing the dispersion to the [0,1] interval, which facilitates unified threshold determination. The fluctuation range of reflection intensity varies significantly under different environmental particle concentrations (e.g., the average intensity in an oil mist environment is only 20-50, while in a dusty environment it can reach 100-200). Normalization enables the algorithm to adapt to diverse scenarios.

[0084] Step S101d: Determination of particle groups.

[0085] Judgment rule: If the normalized reflection intensity dispersion σ norm If the value is greater than 0.8, it is marked as a particle group.

[0086] In this embodiment, the point cloud intensity values ​​within the window are: 121, 120, 125, 78, 21, ..., 32, 200, 210, 80. Calculation: , , .

[0087] If σ norm Values ​​greater than 0.8 are labeled as particle groups. In this embodiment, the normalized reflection intensity dispersion σ norm =0.23≤0.8, no flag is triggered.

[0088] For particle group identification, a threshold of 0.8 was set, corresponding to an actual standard deviation of 32 (i.e., 40 × 0.8). In the turning workshop experiment, 32 was the critical value for distinguishing particle groups from obstacles. The reason why particle marking was not triggered in this embodiment is that there are continuous high-reflectivity points (such as 121, 120, 125) within the window, which may be the edges of metal obstacles; the discrete values ​​mainly come from a few outliers (such as 21, 32), which may be individual distant particles, but because the proportion of high values ​​is large, the overall dispersion does not exceed the standard.

[0089] The process of calculating the discreteness is shown using LaTeX format:

[0090] \begin{align*}

[0091] \mu&= \frac{1}{120} \sum_{i=1}^{120} I_i = \frac{121+120+125+\cdots+117}{120} = 51.3 \\

[0092] \sigma&= \sqrt{\frac{1}{120} \sum_{i=1}^{120} (I_i - 51.3)^2} = 9.2\\

[0093] \sigma_{\text{norm}}&= \frac{9.2}{40} = 0.23 \quad (\text{Default}\sigma_{\max}=40)

[0094] \end{align*}.

[0095] S102, Point Cloud Density Analysis: Detect the point cloud density abrupt change rate Δρ in adjacent scanned areas, and verify and label it in conjunction with the convex hull bounding box area S: If Δρ>50% and the convex hull bounding box area S<0.1m², it is labeled as a particle group.

[0096] Due to the random distribution of particles, solid particles cause abrupt changes in point cloud density; while the density of real obstacles changes gradually. The contour is verified by comparing the rate of change of point cloud density in adjacent regions and combining it with the area S of the convex hull bounding box.

[0097] Table 2 is a comparison table of the number of point clouds in region K and its adjacent regions in this embodiment.

[0098]

[0099] Step S102a: Determine the number of point clouds in the unit scan area and adjacent areas.

[0100] The unit scanning area is 30°. When scanning the environment with lidar, the 360° field of view is divided into 12 30° sector regions. Abrupt changes in the number of point clouds in adjacent regions reflect drastic changes in local spatial distribution.

[0101] Step S102b: Calculate the point cloud density abrupt change rate Δρ for a unit scan area.

[0102] The formula for calculating the point cloud density abrupt change rate Δρ is: Δρ = |ρ k - ρ k-1 | / ρ k-1× 100%; ρ k The number of point clouds in the k-th unit scan region;

[0103] In this embodiment, Δρ = 64.4%. The mutation rate quantifies the severity of density changes. 64.4% is much higher than the typical threshold (50%) for industrial scenarios, indicating the presence of discontinuously distributed targets in this area.

[0104] Step S102c: Verify and label based on the area S of the convex hull bounding box. If Δρ > 50% and the area S of the convex hull bounding box < 0.1 m², label it as a particle group.

[0105] A convex hull bounding box area S < 0.1 m² indicates no clustered contour. In this embodiment, the convex hull detection area S = 0.08 m². 2 It handles dynamic changes in particle concentration, avoids the failure of fixed thresholds, and uses convex hull detection to ensure that only targets without contours are filtered.

[0106] The area of ​​the convex hull bounding box is the area of ​​the smallest bounding rectangle of the convex hull. It is calculated using the Andrew or Graham scan method, and the specific steps are as follows:

[0107] 1. Calculate the vertex coordinates of the convex hull.

[0108] Input data: The coordinate set A of the point cloud of the abrupt change region = {(x1,y1),(x2,y2),...,(x...} n ,y n ...,(x N ,y N )}; where n is the index of the point cloud, (x n ,y n ) are the coordinates of the nth point cloud, and N is the total number of points in the input point set.

[0109] Convex Hull Generation: Extract the convex hull vertices H={h1,h2,...,h} of the point cloud using the Andrew method or Graham scan method. m ...,h M} (sorted clockwise or counterclockwise); where m is the index of the convex hull vertex, and M is the total number of vertices of the output convex hull, usually M < N.

[0110] 2. Calculate the minimum bounding rectangle of the convex hull.

[0111] The minimum bounding rectangle must satisfy the following conditions: its edges are parallel to at least one edge of the convex hull; and all vertices are located inside the rectangle.

[0112] Taking a two-dimensional plane as an example, traversing each edge of the convex hull: opposite edge h i h i+1 Calculate its direction vector =(dx,dy); hi Let be the vertex of the i-th convex hull.

[0113] Calculate the perpendicular distances from all vertices of the convex hull to the edge, and take the maximum perpendicular distance d. max (i.e., height).

[0114] The side length is ;

[0115] The area S of the rectangle corresponding to the side of index i i =L×d max .

[0116] Choose the minimum area: the area S of the rectangle with all sides i Take the minimum value; take the minimum area S as the area of ​​the convex hull bounding box.

[0117] Example calculation: Assume the convex hull of the abrupt change region is a quadrilateral, and the vertex coordinates are as follows (unit: meters):

[0118] h1(0,0),h2(0.4,0),h3(0.4,0.1),h4(0,0.1);

[0119] Traversing edges h1h2: direction vector =(0.4,0);

[0120] Large vertical distance d max =0.1m (y coordinates of vertices h3 and h4);

[0121] Area S1 = 0.4 × 0.1 = 0.04 m² 2 .

[0122] Traversing edges h3h4: direction vector =(0,0.1);

[0123] Maximum vertical distance d max =0.4m;

[0124] Area S2 = 0.1 × 0.4 = 0.04 m 2 .

[0125] The same logic applies to the other sides, resulting in a final minimum area S = 0.04m. 2 .

[0126] The mutation rate calculation process is shown in LaTeX format:

[0127] \Delta\rho = \frac{|42 - 118|}{118} \times 100\% = 64.4\%.

[0128] Step S103, motion trajectory analysis: calculate the angle between the displacement vectors of adjacent frames and derive the motion continuity index η.

[0129] Solid particles move randomly (Brownian motion), while real obstacles move continuously (with gradual changes in direction). By tracking target clusters (such as a group of points with ID:100), the angle between displacement vectors of adjacent frames is calculated, and the motion continuity index η is derived.

[0130] Step S103a: Record the displacement direction angle through target clustering tracking.

[0131] Input data: Cluster of points with target ID:100 (stable target generated by point cloud clustering algorithm).

[0132] Table 3 is a data table of cluster points with target ID 100.

[0133]

[0134] In this embodiment, three consecutive frames of position data (T-2, T-1, T) are used to track the same target across frames, avoiding target ID jumps caused by point cloud jitter. The displacement direction angle in the previous frame is θ1 = 45°; the displacement direction angle in the current frame is θ2 = 45°. The displacement direction angle is recorded, and the continuity of target motion is quantified by the change in the motion direction angle between adjacent frames.

[0135] Step S103b: Calculate the motion continuity index η. .

[0136] In this embodiment, θ1 = 45°, θ2 = 63.4°. .

[0137] Step S103c, Judgment rule: If the motion continuity index η < motion threshold, it is judged as random motion and marked as particle swarm; otherwise, it is not triggered. The default motion threshold is 0.3.

[0138] In this embodiment, η=0.577>0.3, so it is not triggered.

[0139] This step addresses the failure of multi-frame matching methods under random particle motion. The algorithm is efficient (single-target computation latency <0.1ms) and supports real-time tracking.

[0140] The motion trajectory analysis and calculation process is demonstrated using LaTeX format:

[0141] \eta = \frac{\cos45^\circ + \cos63.4^\circ}{2} = \frac{0.707 + 0.447}{2} = 0.577.

[0142] In step S1, if any of the following conditions is met, it is marked as a particle group:

[0143] Condition 1: The dispersion degree of reflection intensity > the preset maximum dispersion degree (e.g., 0.8);

[0144] Condition 2: The mutation rate of point cloud density Δρ > 50% and the area S of the convex hull bounding box < 0.1 m² (no stable clustering is formed);

[0145] Condition 3: The motion continuity index < 0.3; where, the motion continuity index of 0 means completely random, and the motion continuity index of 1 means uniform linear motion.

[0146] Step S2: The determination rule of the particle group.

[0147] In this step, based on the feature extraction results, it comprehensively judges whether the point cloud is a solid particle and triggers the verification mechanism. The input is the output feature value of step 1, and the output is the particle probability P and the preliminary determination. In this step, a weighted fusion mechanism is adopted to avoid misjudgment by a single feature. The particle probability P is the weighted sum of the dispersion degree of reflection intensity σ norm , the change rate of point cloud density Δρ, and the motion discontinuity (1 - η).

[0148] Normalize the feature values of the dispersion degree of reflection intensity σ norm , the change rate of point cloud density Δρ, and the motion discontinuity (1 - η):

[0149] Normalized dispersion degree σ norm = 0.23, as the input value;

[0150] The mutation rate of point cloud density Δρ = 64.4%, and it is obtained that: normalized Δρ = 0.644;

[0151] The motion continuity index η = 0.577, and it is obtained that: the motion discontinuity (1 - η) = 0.423.

[0152] Calculation of particle probability: , where w1, w2, and w3 are preset weights.

[0153] In this embodiment, w1 = 0.5, w2 = 0.3, w3 = 0.2; .

[0154] Determination logic:

[0155] If P ≤ 0.4, it is directly marked as a real obstacle;

[0156] If 0.4 < P ≤ 0.7, cross-frame verification is triggered. If the P value is < 0.7 for 3 consecutive frames, it is marked as a real obstacle;

[0157] If P > 0.7, it is directly labeled as a particle group.

[0158] In this embodiment, P=0.427 ∈ (0.4,0.7], triggering cross-frame verification; for three consecutive frames, the P value [0.427, 0.402,0.415] is all <0.7, and it is finally marked as a real obstacle.

[0159] This step uses multi-feature fusion to reduce the false positive rate, and cross-frame verification enhances robustness. The algorithm is lightweight and supports embedded deployment.

[0160] Step S3: Dynamic parameter adjustment: Input the number of particle marker points and the total number of point cloud points per unit time, and calculate the particle concentration C. 颗粒 If C 颗粒 If the preset concentration threshold is reached, the system switches to a relaxed mode, lowers the motion threshold of the motion continuity index η, and returns to step S1 to re-analyze the motion trajectory.

[0161] This step adaptively adjusts the judgment threshold based on the concentration of environmental particles, improving the adaptability of the solution.

[0162] Technical principle: When the particle concentration is high, the judgment threshold is relaxed (the motion threshold is lowered) to avoid oversensitivity; when the concentration is low, the motion threshold is tightened to improve accuracy.

[0163] Calculation formula: Particle concentration: C 颗粒 = (Number of particle markers / Total number of point cloud points) × 100%.

[0164] Motion threshold adjustment: If C 颗粒 If the threshold is >30%, switch to relaxed mode, and the exercise threshold will be reduced from 0.3 to 0.2.

[0165] In this embodiment, a 10-second cycle is used for a total point cloud of 9500 points and a particle marker of 3200 points.

[0166] calculate: .

[0167] Adjustment: C 颗粒 =33.7%>30%, switch to relaxed mode, the exercise threshold drops to 0.2.

[0168] This step enables environmental adaptation, reducing the need for manual parameter tuning. In actual measurements, it can handle variations in metal debris density of 150-200 pieces / m³.

[0169] Step S4: False Alarm Suppression: Point clouds marked as particle swarms do not participate in subsequent clustering and trajectory tracking; only real obstacle data is retained. The input is the judgment result, and the output is the purified real obstacle data, including real obstacle coordinates and motion vectors.

[0170] This step performs final output processing to suppress false point clouds.

[0171] Processing flow: Remove particle point cloud: Delete points marked as particle groups from the point cloud collection (e.g., 1200 points are removed in this embodiment). Output real obstacle data: including structured data such as ID, position, velocity, and contour area.

[0172] AGV navigation or equipment protection: Upon receiving real obstacle data, appropriate protection measures are taken (such as stopping the machine, obstacle avoidance, or optimizing the navigation route).

[0173] This step ensures that AGV navigation or equipment protection only responds to real threats. The measured false alarm rate decreased from 18.2% to 1.3% (72-hour data from the turning workshop).

[0174] In-depth technical analysis by scenario:

[0175] 1. Scene of flying metal shavings.

[0176] The challenge lies in the fact that the turning process generates 0.5-2mm metal chips, forming a high-density suspension layer of 150-200 particles / m³, which causes drastic fluctuations in reflection intensity (60-220). Traditional static threshold filtering cannot distinguish weakly reflective obstacles.

[0177] Innovative approach: Employing reflection intensity analysis (reflection intensity dispersion σ) norm The system employs a dual mechanism: triggering a marker with a value of >0.8 and analyzing motion trajectory (determining random motion based on a motion continuity index η < 0.3).

[0178] Results: The number of false alarms was reduced from 127 to 9, which is equivalent to avoiding 1.2 unplanned downtimes per hour on the production line, and the overall equipment efficiency (OEE) was improved by 13%.

[0179] 2. Oil mist spraying scenario.

[0180] The challenge lies in the fact that the cooling oil mist generates 1-3μm ultrafine particles with weak reflection signals (20-50) but a density mutation rate exceeding 70%, which causes the multi-frame matching method to fail due to Brownian motion.

[0181] Innovative approach: Introduce point cloud density analysis (started when point cloud density mutation rate Δρ>50%) and convex hull profile verification (convex hull bounding box area S<0.1m², confirming no solid profile).

[0182] Results: The number of false alarms was reduced from 89 to 6, ensuring the continuous operation stability of the high-pressure cooling system and extending the maintenance cycle by 40%.

[0183] 3. AGV dust pollution scenario.

[0184] The challenge lies in the fact that AGVs raise a 0.3-0.8m low-altitude dust layer as they move, which traditional clustering algorithms misjudge as "low obstacles," triggering frequent obstacle avoidance interruptions.

[0185] Innovative approach: Accurate identification is achieved through dynamic threshold adjustment (the motion threshold of the motion continuity index η decreases from 0.3 to 0.2 when the dust concentration is >25%) and multi-feature weighted fusion (particles with a probability P >0.7 are directly labeled as particle groups).

[0186] Results: The number of false alarms was reduced from 54 to 4, AGV throughput efficiency increased by 22%, and the annual logistics throughput increased by 150,000 pieces.

[0187] The advantages of this solution are as follows:

[0188] 1. Multi-feature fusion mechanism:

[0189] For the first time, the reflection intensity discreteness σ norm The triple feature fusion of point cloud density mutation rate Δρ and motion continuity index η solves the problem of high false positive rate of single feature. It remains stable within the dynamic range of particle concentration (150-350 particles / m³) with a false negative rate of <0.5%. In contrast, the false positive rate of single feature scheme is still as high as 8.2-12.7%.

[0190] 2. Dynamic threshold adjustment:

[0191] The threshold for judgment is adaptively optimized based on real-time particle concentration to adapt to dynamic environmental changes (such as instantaneous concentration spikes caused by oil mist injection).

[0192] 3. Lightweight real-time processing:

[0193] The algorithm complexity is optimized to O(n), the single-frame processing latency on the embedded platform is <10ms, and the CPU usage on the Jetson TX2 platform is <15%, meeting the real-time requirements of industrial scenarios.

[0194] This solution uses a single-line LiDAR (such as the S300-C-3 model) to collect point cloud data and combines it with multi-feature analysis (reflection intensity, point cloud density, motion trajectory) to dynamically distinguish between solid particles and real obstacles (such as people or AGVs). The algorithm runs on an embedded platform (such as the NVIDIA Jetson TX2), with a processing latency of <10ms, supporting real-time applications (scanning frequency of 25Hz). The advantages of this solution include a false alarm rate reduction of over 80% (from 18.2% to 1.3%) and environmental adaptability.

[0195] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method of monostatic lidar solid particle simulation, characterized by, Comprising the following steps: Step S1, single-line laser radar acquires raw point cloud data according to time of flight, extracts feature information according to the set feature model, and sends it to the particle feature analysis module for analysis, including: Reflection intensity analysis; calculate the standard deviation of the signal intensity of all signal points in the scanning window, quantify the reflection intensity dispersion; if the reflection intensity dispersion is higher than the preset maximum dispersion, mark it as a particle group; Point cloud density analysis; detect the point cloud density mutation rate Δρ of adjacent scanning areas, and verify and mark in combination with the convex hull bounding box area S; Motion trajectory analysis; calculate the angle of adjacent frame displacement vectors, and deduce the motion continuity index η; Through target clustering tracking, the same target is tracked across frames, and the displacement direction angle is recorded; Computing motion continuity index ; wherein θ1 is the previous frame displacement direction included angle, θ2 is the current frame displacement direction included angle; If the motion continuity index η is less than the motion threshold, it is determined as random motion, and marked as a particle group; Step S2, particle determination: based on the characteristic value of step S1, calculate the particle probability P through weighted fusion, and mark the particle group or real obstacle according to the P value range; Step S3, dynamic parameter adjustment: according to the particle concentration of the environment, adaptively adjust the motion threshold in the motion trajectory analysis; Step S4, false alarm suppression: remove the point cloud marked as a particle group, and output the purified real obstacle data.

2. The method of claim 1, wherein the single-beam lidar solid particle fabrication method is characterized by, The reflection intensity analysis in step S1 includes: Input the reflection intensity value of all signal points in the scanning window, calculate the average value of the reflection intensity ; wherein N is the number of all signal points in the scanning window; l i is the reflection intensity value of the i-th signal point, ranging from 0 to 255. Computing the standard deviation of the reflected intensity ; Computing a normalized reflectance intensity dispersion ; wherein σ max is a predetermined maximum standard deviation; If the normalized intensity dispersion σ norm > 0.8, the particle group is marked.

3. The method of claim 2, wherein the single-beam lidar solid particle fabrication method is characterized by, The point cloud density analysis in step S1 includes: Determine the number of point clouds in the unit scanning area and adjacent areas; The point cloud density mutation rate Δρ of the unit scanning area is calculated as follows: Δρ = |ρ k - ρ k-1 | / ρ k-1 × 100%; ρ k is the point cloud number of the kth unit scanning area; In combination with the convex hull bounding box area S, verify and mark, if Δρ>50% and the convex hull bounding box area S<0.1m², mark it as a particle group.

4. The method of claim 3, wherein the single-beam lidar solid particle fabrication method is characterized by, In step S2, the reflection intensity dispersion σ norm characteristic value normalization of the point cloud density variation rate Δρ and the motion discontinuity (1-η): Particle probability calculation: where w1, w2, and w3 are preset weights. The determination rules include: If P ≤ 0.4, directly mark it as a real obstacle; If 0.4 < P ≤ 0.7, trigger cross-frame verification, if the P value of the last 3 frames is less than 0.7, mark it as a real obstacle; If P > 0.7, directly mark it as a particle group.

5. The method of claim 4, wherein the single-beam lidar solid particle fabrication method is characterized by, The dynamic parameter adjustment of step S3 includes: Computing the particle concentration C 颗粒 = (number of particle marker points / total number of point cloud points) x 100%; If C 颗粒 > 30%, switch to a relaxed mode, reduce the motion threshold of the motion continuity index η; and return to step S1 to re-analyze the motion trajectory.

6. The method of claim 5, wherein the single-beam lidar solid particle fabrication method is characterized by, The false alarm suppression of step S4 includes: Remove the point cloud marked as a particle group: delete the point cloud marked as a particle group from the point cloud set; Output real obstacle data, including real obstacle coordinates and motion vector.

Images