A rain grade determination method based on laser point cloud time series steady state analysis

Abstract

The application discloses a rain grade determination method based on laser point cloud timing steady state analysis, and belongs to the technical field of automatic driving perception. The method receives original three-dimensional point cloud data output by a vehicle-mounted laser radar, and obtains effective point cloud data through format conversion, invalid point elimination and quality checking. Suspected raindrop points are screened according to spatial distance and laser reflection intensity. Three-dimensional space neighborhood retrieval and sparsity checking are performed on the suspected raindrop points, and the number of effective raindrops in the current frame is counted. The current frame instantaneous rain grade is determined according to a preset multi-level rain grade classification standard. The current frame instantaneous rain grade is written into a timing sliding cache queue, and a continuous frame rain grade sequence is obtained. The mode of the continuous frame rain grade sequence is counted, and the steady state rain grade is determined and output. The method can reduce the influence of single frame raindrop false detection and grade jump on the rain determination result.

Description

Technical Field

[0001] This application belongs to the field of autonomous driving perception technology, specifically a method for determining rainfall level based on laser point cloud temporal steady-state analysis. Background Technology

[0002] LiDAR (Light Detection and Ranging) is a crucial sensor for autonomous vehicles in environmental perception, obstacle recognition, localization and mapping, and driving decision-making. LiDAR generates 3D point cloud data by emitting laser beams and receiving their echoes. Vehicle perception systems typically use the spatial location, reflection intensity, and object contours within the point cloud to identify environmental objects such as roads, vehicles, pedestrians, and guardrails. During rainfall, raindrops scatter or reflect the laser beam, resulting in numerous discrete or weakly reflective points in the point cloud data. These points can affect obstacle contour recognition and cause fluctuations in ranging results. Therefore, how to utilize existing point cloud data from vehicle-mounted LiDAR to perform real-time and stable determination of rainfall levels is a critical issue in the adverse weather perception and processing of autonomous vehicles.

[0003] Current vehicle-mounted rain detection methods typically rely on mechanical rain sensors, optical rain sensors, or wiper status for judgment. These methods mostly perform single-point detection at localized locations on the vehicle, and the detection results do not match the global point cloud noise distribution formed by lidar around the vehicle. For example, if a rain sensor is installed in the windshield area, when the vehicle is traveling at high speed, the water film on the windshield may be affected by vehicle speed, wiper operation, and wind direction, while the raindrop distribution in the lidar's field of view may show different densities in front of, to the side of, or above the vehicle. The single-point rain measurement result cannot directly reflect the true degree of interference from raindrops in the point cloud.

[0004] Existing raindrop recognition methods based on laser point clouds also suffer from insufficient recognition stability. Some methods determine the presence of rainfall solely based on the reflection intensity, distance, or number of points in a single frame of the point cloud, failing to fully consider the discrete and sparse distribution of raindrops in three-dimensional space. This easily leads to misidentification of clutter at vehicle edges, reflections from road surface water, distant weak reflective objects, or sensor anomalies as raindrops. To illustrate, when a vehicle passes through a streetlamp reflection or a splash from another vehicle, a batch of low-intensity discrete points may suddenly appear in a single frame of the point cloud. If the system only considers the number of points in a single frame, it may misidentify a brief splash as moderate or heavy rain.

[0005] Furthermore, rainfall exhibits significant temporal fluctuations, with single-frame point clouds being greatly affected by gusts, changes in vehicle attitude, sudden changes in local raindrops, and data acquisition anomalies. If the rainfall level is directly output based on the number of raindrops in a single frame, the rainfall level may frequently jump between "no rain, light rain, moderate rain, and heavy rain," causing repeated adjustments to subsequent perception parameters, planning strategies, or safety redundancy strategies, thus affecting the operational stability of the autonomous driving system.

[0006] Therefore, a method is needed to quantify rainfall levels based on native laser point cloud data, combined with multi-feature raindrop identification, spatial sparsity verification, and multi-frame temporal steady-state analysis. Summary of the Invention

[0007] To address the shortcomings of existing vehicle-mounted rainfall detection methods in reflecting the degree of raindrop interference in point clouds within the field of view of lidar, and the problems that existing rainfall determination methods based on lidar point clouds are easily affected by single-frame false detections, local splashing, gust disturbances, or abnormal acquisition, resulting in abrupt changes in rainfall levels, this application provides a rainfall level determination method based on the temporal steady-state analysis of lidar point clouds. This method uses the original three-dimensional point cloud data output by vehicle-mounted lidar, combined with spatial distance, lidar reflection intensity, three-dimensional spatial sparsity, and continuous frame temporal statistics, to stably determine the rainfall level.

[0008] To achieve the above objectives, this application provides a method for determining rainfall level based on laser point cloud time-series steady-state analysis, comprising: The system receives raw 3D point cloud data output from an onboard LiDAR and preprocesses it to obtain effective point cloud data. Based on the spatial distance and laser reflection intensity of each point in the effective point cloud data, it filters out potential raindrop points. It performs a 3D spatial neighborhood search on the potential raindrop points and performs sparsity verification on them based on the neighborhood search results, counting the number of effective raindrops in the current frame. Based on the number of effective raindrops in the current frame and a preset multi-level rainfall classification standard, it determines the instantaneous rainfall level of the current frame. It writes the instantaneous rainfall level of the current frame into a time-series sliding buffer queue to obtain a continuous frame rainfall level sequence. It performs mode statistics on the continuous frame rainfall level sequence to determine the steady-state rainfall level and outputs the steady-state rainfall level as the current rainfall level determination result.

[0009] In one optional implementation, the original 3D point cloud data is preprocessed, including: reading the 3D coordinates and laser reflection intensity of each point in the original 3D point cloud data, converting the original 3D point cloud data into a unified point cloud structure; removing blank frames, invalid coordinate points, points with missing reflection intensity, and near-field vehicle noise points located within a preset vehicle interference range; when the number of remaining points and field completeness meet preset point cloud quality conditions, the remaining points are determined as the valid point cloud data. Thus, the impact of format abnormalities, missing fields, and near-field vehicle interference on subsequent rainfall statistics can be eliminated before raindrop point filtering.

[0010] In one optional implementation, screening the suspected raindrop points includes: determining the corresponding spatial distance based on the three-dimensional coordinates of each point in the effective point cloud data; comparing the spatial distance with a preset raindrop detection spatial range, and comparing the laser reflection intensity with a preset reflection intensity screening interval; when the same point is simultaneously located within the preset raindrop detection spatial range and the preset reflection intensity screening interval, the point is identified as the suspected raindrop point. This avoids judging raindrop points solely based on a single reflection intensity or a single distance feature, improving the matching accuracy between the initial screening results and the raindrop point cloud features.

[0011] In one optional implementation, sparsity verification of the suspected raindrop points includes: establishing a KD-Tree index structure based on the valid point cloud data; using each suspected raindrop point as a retrieval center, searching for neighboring points within a preset neighborhood search range and obtaining the number of neighboring points; comparing the number of neighboring points with a preset sparsity judgment condition, and including suspected raindrop points that meet the preset sparsity judgment condition into the valid raindrop points. This allows for the exclusion of continuous obstacle outlines, road surface reflection fragments, or densely distributed non-raindrop points from the raindrop statistical chain.

[0012] In one optional implementation, counting the number of valid raindrops in the current frame includes: accumulating the valid raindrop points in the current frame that meet the preset sparsity determination condition to obtain the number of valid raindrops in the current frame; not including suspected raindrop points that do not meet the preset sparsity determination condition in the number of valid raindrops in the current frame; and clearing the temporary counting parameter of the current frame after obtaining the number of valid raindrops in the current frame. This allows the number of valid raindrops in each frame to be generated independently, reducing the interference of the temporary counting result of the previous frame on the determination of the current frame.

[0013] In one optional implementation, determining the instantaneous rainfall level of the current frame includes: reading a preset multi-level rainfall classification standard, wherein the preset multi-level rainfall classification standard includes effective raindrop count ranges corresponding to no rain, light rain, moderate rain, and heavy rain levels, respectively; comparing the effective raindrop count of the current frame with the effective raindrop count range level by level, and determining the rainfall level that matches the comparison as the instantaneous rainfall level of the current frame. Thus, the effective raindrop count of the current frame can be converted into a quantified rainfall level that can be used for autonomous driving perception, planning, or safety strategy invocation.

[0014] In one optional implementation, obtaining a continuous frame rainfall level sequence includes: reading a preset temporal buffer window capacity; when the temporal sliding buffer queue has not reached the preset temporal buffer window capacity, appending the instantaneous rainfall level of the current frame to the temporal sliding buffer queue; when the temporal sliding buffer queue reaches the preset temporal buffer window capacity, deleting the earliest written rainfall level and writing the instantaneous rainfall level of the current frame to obtain an updated continuous frame rainfall level sequence. This ensures that the rainfall level determination result is always based on the most recent continuous frames, balancing real-time performance and temporal stability.

[0015] In one optional implementation, determining the steady-state rainfall level includes: counting the occurrence frequency of each rainfall level in the continuous frame rainfall level sequence; determining the rainfall level with the highest occurrence frequency as the steady-state rainfall level; when at least two rainfall levels have the same occurrence frequency, maintaining the steady-state rainfall level output from the previous frame, or determining the steady-state rainfall level from the rainfall levels with the same occurrence frequency according to a preset level priority rule. This can suppress frequent jumps in rainfall levels caused by single-frame abnormal point clouds, localized splashing, or short-term acquisition anomalies.

[0016] In one optional implementation, when the current frame fails the quality check, spatial distance and laser reflection intensity filtering are not performed, the number of valid raindrops for the current frame is not generated, and the instantaneous rainfall level for the current frame to be written into the temporal sliding buffer queue is not generated; instead, the steady-state rainfall level output from the previous frame is used. When the number of consecutive frames that fail the quality check reaches a preset unusable frame threshold, updating the steady-state rainfall level based on consecutively failed quality checks is stopped. This avoids low-quality point cloud frames from participating in rainfall level statistics, reducing the impact of abnormal frames on the steady-state output results.

[0017] In one optional implementation, the preset multi-level rainfall classification standard is generated through sample calibration, including: statistically analyzing the distribution of valid raindrop counts in the current frame corresponding to labeled samples of no rain, light rain, moderate rain, and heavy rain; determining a first, second, and third quantity thresholds based on the boundary intervals of quantity distributions between adjacent labeled rainfall levels; and writing the first, second, and third quantity thresholds into the preset multi-level rainfall classification standard. This ensures that the rainfall classification standard matches the actual raindrop count distribution in the lidar point cloud.

[0018] Compared with related technologies, this application has the following advantages: First, this application utilizes the raw 3D point cloud data output by an onboard LiDAR to determine the rainfall level. Within the point cloud processing chain, spatial distance and laser reflection intensity filtering, 3D spatial neighborhood retrieval, and sparsity verification are performed sequentially, ensuring that the rainfall level determination result corresponds to the point cloud noise distribution within the actual perceived field of view of the LiDAR. Therefore, on the one hand, it eliminates the need for additional reliance on mechanical rain sensors, optical rain sensors, or wiper status; on the other hand, it reduces the probability of vehicle body noise, distant weak reflective objects, localized reflective points, continuous obstacle outlines, and road surface reflection fragments entering the effective raindrop count statistical chain, thereby reducing interference from non-raindrop points on the current frame's rainfall determination result.

[0019] Second, after converting the number of effective raindrops in the current frame into the instantaneous rainfall level, this application does not directly use the instantaneous result of a single frame as the final output. Instead, it writes the instantaneous rainfall level of the current frame into a time-series sliding buffer queue and performs mode statistics based on the continuous frame rainfall level sequence to determine the steady-state rainfall level. This suppresses rainfall level jumps caused by anomalous point clouds in a single frame, gust disturbances, local splashing, changes in vehicle attitude, or sensor acquisition anomalies, making the output more suitable as input for adjusting autonomous driving perception parameters, selecting planning strategies, or triggering safety redundancy strategies.

[0020] Third, this application outputs quantitative rainfall levels such as no rain, light rain, moderate rain, and heavy rain based on a preset multi-level rainfall classification standard. Compared with the scheme that only judges whether there is rainfall, it can provide input basis for severe weather risk assessment and adaptive optimization of sensing parameters by level. Furthermore, the preset multi-level rainfall classification standard can be sample calibrated by the effective raindrop number distribution corresponding to the labeled samples of no rain, light rain, moderate rain, and heavy rain, so that the rainfall level classification results match the actual raindrop number distribution of lidar point clouds. Attached Figure Description

[0021] Figure 1 A flowchart of a rainfall level determination method based on laser point cloud time-series steady-state analysis provided in this application embodiment; Figure 2 A flowchart for multi-dimensional raindrop feature initial screening and spatial neighborhood sparsity verification provided in this application embodiment; Figure 3 A flowchart of time-series sliding cache update and mode steady-state correction provided in an embodiment of this application; Figure 4 This is a schematic diagram of a rainfall level determination system based on laser point cloud time-series steady-state analysis, provided as an embodiment of this application. Detailed Implementation

[0022] The technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings. The described embodiments are only used to explain this application and should not be construed as limiting the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0023] The original three-dimensional point cloud data in this embodiment can be data output frame by frame by an onboard LiDAR. Each point can include three-dimensional coordinates and laser reflection intensity. The three-dimensional coordinates can be represented as the spatial position in the vehicle coordinate system or the LiDAR coordinate system. The laser reflection intensity can represent the echo intensity after the laser beam is reflected by the target. The rainfall levels in this embodiment can include no rain, light rain, moderate rain, and heavy rain, and can also be expanded to more levels according to the needs of the vehicle perception system. For ease of description, the following uses four levels—no rain, light rain, moderate rain, and heavy rain—as an example.

[0024] like Figure 1 As shown in the embodiments of this application, the method provided may include: original laser point cloud input, point cloud preprocessing and quality verification, multi-dimensional raindrop feature initial screening, spatial neighborhood sparsity verification, single-frame instantaneous rainfall level generation, temporal sliding buffer update, steady-state mode correction and level output, and outputting quantified rainfall level results. After outputting the quantified rainfall level results, the system cyclically receives the next frame of original three-dimensional point cloud data and repeats the above processing flow.

[0025] In one embodiment, a parameter configuration table can be configured during system initialization. The parameter configuration table may be as shown in Table 1.

[0026]

[0027] Table 1 Parameter Configuration Table

[0028] The parameters in Table 1 can be configured during the vehicle's factory calibration phase, or updated during road testing based on the LiDAR model, installation location, point cloud frame rate, vehicle speed range, and historical rainy day samples. The parameter configuration table can be stored in the vehicle computing platform's configuration file, database, or firmware parameter area.

[0029] The parameter configuration table can be generated through sample calibration. During sample calibration, continuous point cloud frames of the vehicle-mounted LiDAR are collected under rainless, light rain, moderate rain, and heavy rain conditions, and the manually labeled rainfall level or external rainfall reference level is recorded. Each sample frame can simultaneously record the frame timestamp, vehicle speed, LiDAR installation attitude, LiDAR model, point cloud frame rate, and manually labeled rainfall level. The system or calibration personnel will statistically analyze the spatial distance distribution, laser reflection intensity distribution, neighborhood point number distribution, and single-frame effective raindrop number distribution of the point cloud samples under each labeled level. Based on the distribution difference between raindrop samples and non-raindrop samples, the preset raindrop detection spatial range, preset reflection intensity screening interval, and preset sparsity judgment conditions are determined. Based on the effective raindrop number distribution under different labeled rainfall levels, the preset multi-level rainfall classification standard is determined.

[0030] Point cloud samples collected under rainless conditions can be used to statistically analyze the distance, reflection intensity, and distribution of neighboring points of vehicle body clutter, road reflection points, guardrail reflection points, and sensor anomalies. Point cloud samples collected under rainy conditions can be used to statistically analyze the distance, reflection intensity, and distribution of neighboring points of raindrops under different rainfall levels. External rainfall reference levels can be provided by manual observation records, road test rainfall records, or the output results of rainfall simulation equipment at the vehicle test site. Manual intervention can include sample level labeling, anomaly sample removal, parameter boundary verification, and parameter calibration at different vehicle model installation locations.

[0031] The system receives raw 3D point cloud data output from the vehicle-mounted LiDAR. This raw 3D point cloud data can use data formats from different LiDAR manufacturers, and may include fields such as point number, 3D coordinates, reflection intensity, scan line number, or timestamp. The system first converts the different input formats into a unified point cloud structure, which includes at least the 3D coordinates of the points and the LiDAR reflection intensity.

[0032] Format conversion can be achieved using existing point cloud data parsing methods. For example, the software development kit provided by the LiDAR manufacturer can be used to read the coordinate and reflection intensity fields from the original data packet; the ROS point cloud message parsing method can be used to read the point coordinates and intensity fields in the PointCloud2 format; or predefined binary field offset rules can be used to read the 3D coordinates and intensity fields from the point cloud file. The input of the above implementation methods is the original 3D point cloud data, the processing method is to read the fields according to the corresponding format and convert them into a unified point cloud structure, and the output is point cloud data with consistent field names and field types. The improvement focus of this application is not on the point cloud format conversion method itself, but on the fact that the format-converted point cloud data enters the subsequent raindrop initial screening, neighborhood sparsity verification, and time-series steady-state output chain to generate rainfall level determination results.

[0033] After format conversion, the system performs invalid point removal. Invalid points can include anomalous coordinates (NaN), points lacking reflection intensity fields, points whose distance values ​​do not meet the device's effective detection range, and near-field vehicle noise within a preset vehicle interference range. NaN stands for Not a Number, indicating non-numerical anomalous data. Near-field vehicle noise can be determined by the LiDAR installation location, vehicle body contour, and vehicle reflection area. For example, a LiDAR installed on the roof may create fixed noise near the vehicle edge, hood, or windshield wipers; this type of noise can be removed using the preset vehicle interference range. The output of invalid point removal is the remaining point cloud, which is used for quality verification.

[0034] After invalid point removal, the system performs quality verification. Specifically, the preset point cloud quality conditions can be pre-written into a parameter configuration table. These preset point cloud quality conditions include blank frame determination conditions, a preset minimum number of remaining points, a preset reflection intensity field integrity rate, and a preset point cloud timeliness threshold. The blank frame determination condition is used to determine whether there are no readable points in the current frame; the preset minimum number of remaining points is used to determine whether the number of remaining points after invalid point removal meets the data volume required for subsequent raindrop feature initial screening; the preset reflection intensity field integrity rate is used to determine the proportion of points with laser reflection intensity fields in the current frame to the number of remaining points; and the preset point cloud timeliness threshold is used to determine whether the time difference between the current frame timestamp and the current system processing time meets the timeliness requirements.

[0035] During quality verification, the system reads the remaining point cloud after invalid point removal, the current frame timestamp, and the preset point cloud quality conditions from the parameter configuration table. First, it determines whether the remaining point cloud is a blank frame. If it is, the current frame does not meet the preset point cloud quality conditions. If it is not a blank frame, the system counts the number of remaining points in the remaining point cloud and compares this number with the preset minimum number of remaining points. If the number of remaining points is less than the preset minimum number of remaining points, the current frame does not meet the preset point cloud quality conditions. If the number of remaining points is not less than the preset minimum number of remaining points, the system counts the number of points with a laser reflection intensity field and calculates the difference between the number of points with a laser reflection intensity field and the number of remaining points. The ratio is used to obtain the completeness rate of the reflection intensity field. When the completeness rate of the reflection intensity field is less than the preset completeness rate of the reflection intensity field, it is determined that the current frame does not meet the preset point cloud quality condition. When the completeness rate of the reflection intensity field is not less than the preset completeness rate of the reflection intensity field, the time difference between the timestamp of the current frame and the current processing time of the system is calculated, and the time difference is compared with the preset point cloud timeliness threshold. When the time difference is greater than the preset point cloud timeliness threshold, it is determined that the current frame does not meet the preset point cloud quality condition. When the time difference is not greater than the preset point cloud timeliness threshold, it is determined that the current frame meets the preset point cloud quality condition, and the remaining point cloud is determined as valid point cloud data for subsequent multi-dimensional raindrop feature initial screening.

[0036] The preset point cloud quality conditions can be determined based on the vehicle-mounted LiDAR model, point cloud frame rate, theoretical number of points per frame, point cloud field format, allowable processing latency of the vehicle perception system, and actual vehicle-collected samples. For example, when the number of points output per frame by the LiDAR under normal operating conditions is concentrated within a preset range, the number of points below the lower limit of this preset range can be set as the preset minimum remaining number of points; when the point cloud field format requires each point to have a laser reflection intensity field, the preset reflection intensity field integrity rate can be set as a proportional value corresponding to the integrity requirement of this field; after the LiDAR frame rate and the vehicle perception system processing cycle are determined, the preset point cloud timeliness threshold can be set according to the time interval between adjacent frames and the allowable processing latency. After the above parameters are written into the parameter configuration table, they are read and compared during each frame quality check.

[0037] Quality verification determines whether the current frame enters the suspected raindrop screening process. Frames that fail quality verification are not screened for spatial distance and laser reflection intensity, and no valid raindrop count or instantaneous rainfall level for writing to the temporal sliding buffer queue is generated. The preset unusable frame threshold can be determined based on the LiDAR frame rate and the allowed duration of the rainfall level to be maintained by the vehicle perception system. For example, when the LiDAR frame rate is 10 frames per second and the vehicle perception system allows the previous rainfall level to be maintained for 1 second, the preset unusable frame threshold can be set to 10 frames, or, after test calibration, to another number of frames corresponding to that maintenance duration.

[0038] In one implementation, to reduce computational load, the system can pre-calculate the squared distance value for each point in the valid point cloud data. For the i-th valid point, its three-dimensional coordinates can be represented as... The squared distance from this point to the origin of the lidar coordinate system can be expressed as:

[0039] in, This represents the squared distance to the i-th valid point; This represents the coordinate value of the i-th valid point along the first coordinate axis. This represents the coordinate value of the i-th valid point along the second coordinate axis. This represents the coordinate value of the i-th valid point along the third coordinate axis. The system can directly use this coordinate value when performing spatial distance filtering later. The distance squared is compared with the preset raindrop detection spatial range to avoid repeatedly performing the square root operation for each point.

[0040] Figure 2 A flowchart illustrating the initial screening of multi-dimensional raindrop features and the verification of spatial neighborhood sparsity. (Reference) Figure 2 The system iterates through each point in the valid point cloud data, reading the squared distance value and laser reflection intensity of that point. The preset raindrop detection spatial range can include a lower and upper distance limit; the preset reflection intensity filtering interval can include a lower and upper reflection intensity limit. When the squared distance value of a point is between the squared lower and upper distance values, and the laser reflection intensity of that point is between the lower and upper reflection intensity limits, the system identifies that point as a suspected raindrop point. The suspected raindrop point consists of the corresponding point's three-dimensional coordinates, laser reflection intensity, and point number, used for subsequent spatial neighborhood sparsity verification.

[0041] In one implementation, the initial screening of multi-dimensional raindrop features can be performed using the following judgment relationship:

[0042] in, This indicates the lower limit of the preset raindrop detection range. This indicates the upper limit of the preset raindrop detection range; This represents the laser reflection intensity at the i-th effective point; This indicates the lower limit of the preset reflection intensity filtering range; This indicates the upper limit of the preset reflection intensity filtering range. When both the distance and reflection intensity conditions are met, the system adds the corresponding valid point to the suspected raindrop point set.

[0043] The lower distance limit can be used to exclude near-field fixed noise points on the vehicle body, while the upper distance limit can be used to exclude distant weakly reflective background points. The reflection intensity screening range can be calibrated through real-vehicle sampling in rainy and non-rainy weather. For example, laser point cloud samples can be collected under rainless, drizzling, light, moderate, and heavy rain conditions, and the differences in reflection intensity distribution between raindrop points and road points, vehicle points, and guardrail points can be statistically analyzed to determine the reflection intensity screening range. The above values ​​can be adjusted according to different LiDAR models and installation locations.

[0044] After screening suspected raindrop points, the system establishes a KD-Tree index structure based on the valid point cloud data. The KD-Tree organizes points in the 3D point cloud space, enabling the system to retrieve neighboring points within a preset search range. The KD-Tree index structure can be implemented using existing 3D point cloud indexing methods, such as using KD-Tree construction functions from open-source point cloud libraries, using the spatial index interface pre-built into the vehicle computing platform, or using a KD-Tree construction program that recursively partitions the point cloud space according to coordinate axes. The input to this stage is valid point cloud data; the processing method involves building a spatial index according to 3D coordinates; and the output is a tree index structure that returns neighboring points based on the search center and search radius. The improvement focus of this application is not on the KD-Tree index structure itself, but on using the number of neighboring points it outputs to participate in the sparsity verification of suspected raindrop points, and using the verified number of valid raindrops for subsequent rainfall level determination.

[0045] For each suspected raindrop point, the system uses that suspected raindrop point as the search center to search for neighboring points within a preset neighborhood search range and obtains the number of neighboring points. In one implementation, the suspected raindrop point... The corresponding set of neighborhood points can be represented as:

[0046] in, Indicates suspected raindrops The set of neighborhood points; Represents any point in the valid point cloud data; Indicates suspected raindrops With point Spatial distance between them; This indicates the search radius corresponding to the preset neighborhood search range; j≠i indicates that the neighborhood point set does not include the suspected raindrop point itself, which serves as the search center. The system can... The number of midpoints indicates suspected raindrops. Does it meet the preset sparsity determination criteria?

[0047] When counting the number of neighboring points, the suspected raindrop point that serves as the retrieval center can be excluded from the count. If the retrieval center point is included in the count, the neighboring point count threshold in the preset sparsity determination criteria should be calibrated according to the counting rules that include the retrieval center point. This avoids inconsistencies between the statistical calibrator for the number of neighboring points and the calibrator for the sparsity threshold.

[0048] The preset sparsity determination condition can include that the number of neighboring points is not greater than the preset sparse neighboring point count. This means that raindrops are typically discretely distributed in 3D space, while the outlines of obstacles such as vehicles, guardrails, pedestrians, and road signs are typically densely distributed in their local neighborhoods. When the number of neighboring points of a suspected raindrop is not greater than the preset sparse neighboring point count, the system can determine the suspected raindrop as a valid raindrop; when the number of neighboring points of a suspected raindrop is greater than the preset sparse neighboring point count, the system can treat the suspected raindrop as a continuous obstacle point cloud or environmental reflection point, and not include it in the valid raindrop count. A valid raindrop consists of the corresponding suspected raindrop's 3D coordinates, laser reflection intensity, and point number, used for counting the valid raindrops in the current frame.

[0049] In one implementation, the number of valid raindrops in the current frame can be counted as follows:

[0050] in, This represents the number of valid raindrops in the current frame (frame t). This represents the number of suspected raindrops in frame t; The Σ represents the valid raindrop determination value of the i-th suspected raindrop point, which is 1 when the suspected raindrop point meets the preset sparsity determination condition, and 0 otherwise; Σ represents the sum of the determination values ​​of all suspected raindrop points in the current frame; and the number of valid raindrops in the current frame. Used for subsequent instantaneous rainfall level determination.

[0051] After obtaining the number of valid raindrops in the current frame, the system reads the preset multi-level rainfall classification standard. The preset multi-level rainfall classification standard is shown in Table 2.

[0052]

[0053] Table 2 Preset Multi-level Rainfall Classification Standards

[0054] The first, second, and third quantity thresholds in Table 2 can be determined based on the lidar frame rate, detection range, raindrop sample statistics, manually labeled rainfall levels, and vehicle operating scenarios. To avoid overlapping of level intervals, the first quantity threshold is lower than the second quantity threshold, and the second quantity threshold is lower than the third quantity threshold. The system will then calculate the effective raindrop count for the current frame. Compare with the effective raindrop count range in Table 2 to determine the instantaneous rainfall level for the current frame. The instantaneous rainfall level Lt for the current frame consists of the frame number, rainfall level, and the number of valid raindrops in the current frame, and is used for time-series sliding buffer updates.

[0055] In one implementation, the distribution of effective raindrop counts in the current frame corresponding to no rain, light rain, moderate rain, and heavy rain labeled samples can be statistically analyzed separately. A first, second, and third quantity thresholds are determined based on the boundary intervals of the quantity distributions between adjacent labeled rainfall levels. Specifically, the system reads the maximum concentration interval of effective raindrop counts in the current frame for no rain labeled samples and the minimum concentration interval for effective raindrop counts in the current frame for light rain labeled samples, and determines the boundary value between these two intervals as the first quantity threshold. The system then reads the boundary value of effective raindrop counts between light rain labeled samples and moderate rain labeled samples to determine the second quantity threshold. Finally, the system reads the boundary value of effective raindrop counts between moderate rain labeled samples and heavy rain labeled samples to determine the third quantity threshold. Once the thresholds are determined, they are written into a preset multi-level rainfall classification standard for use in the single-frame instantaneous rainfall level generation process.

[0056] Figure 3 Flowchart for time-series sliding cache update and mode steady-state correction. (Reference) Figure 3 The system writes the instantaneous rainfall level of the current frame to a time-series sliding buffer queue. This queue can store the instantaneous rainfall levels of the most recent W frames, where W is the preset buffer window capacity. When the queue is not full, the system directly appends the instantaneous rainfall level of the current frame to the tail of the queue; when the queue is full, the system deletes the earliest written rainfall level from the head of the queue and writes the instantaneous rainfall level of the current frame to the tail. Thus, the system always maintains a sequence of the most recent consecutive frame rainfall levels. This sequence is formed by arranging the instantaneous rainfall levels of each frame within the buffer window in chronological order of their write times, and is used for mode statistics.

[0057] In one implementation, a sequence of consecutive frame rainfall levels can be represented as:

[0058] in, This represents the sequence of consecutive frame rainfall levels corresponding to frame t. represents the instantaneous rainfall level in frame t; W represents the preset time-series buffer window capacity. If the number of frames processed so far is less than W, then... This can include all instantaneous rainfall levels that have been processed so far. If the current frame fails the quality check, no frame will be generated for writing. of , Maintain the queue contents corresponding to the previous valid frame.

[0059] The system performs mode statistics on the continuous frame rainfall level sequence. The candidate rainfall level set includes no rain level, light rain level, moderate rain level, and heavy rain level. The system counts the frequency of each rainfall level in the continuous frame rainfall level sequence and determines the rainfall level with the highest frequency as the steady-state rainfall level. The steady-state rainfall level can be represented as:

[0060] in, Indicates the steady-state rainfall level; This represents the k-th candidate rainfall level in the candidate rainfall level set; Indicates candidate rainfall level In a continuous frame rainfall level sequence The number of times it appears in; This indicates the candidate rainfall level number that appears most frequently. `argmax` represents the candidate rainfall level with the highest frequency; `argmax` represents the index corresponding to the candidate rainfall level with the highest frequency. When multiple candidate rainfall levels have the highest frequency, the system determines the current steady-state rainfall level according to the preset level priority rule or the steady-state rainfall level of the previous frame.

[0061] For example, when the preset time-series buffer window capacity is 5, and the consecutive frame rainfall level sequence is light rain, light rain, moderate rain, light rain, and no rain, if the light rain level appears 3 times, the moderate rain level appears once, the no rain level appears once, and the heavy rain level appears 0 times, the system will determine the light rain level as the steady-state rainfall level. This steady-state rainfall level will be output as the current rainfall level determination result.

[0062] When at least two candidate rainfall levels appear the same number of times in a consecutive frame rainfall level sequence, the system can maintain the steady-state rainfall level output from the previous frame, thereby reducing level oscillations in cases of tied modes. If no steady-state rainfall level exists in the previous frame, the steady-state rainfall level can be determined according to a preset level priority rule. The preset level priority rule can be set according to the vehicle safety strategy. For example, in vehicles with stricter requirements for triggering safety redundancy strategies, the system can prioritize levels ranked lower in the order of no rain, light rain, moderate rain, and heavy rain. In vehicles that avoid erroneous strategy switching, the system can prioritize levels ranked higher in the order of no rain, light rain, moderate rain, and heavy rain, or default to the no rain level.

[0063] After the steady-state rainfall level is determined, the system outputs the steady-state rainfall level as the current rainfall level judgment result. The output can be submitted to the vehicle-mounted perception fusion module, the autonomous driving decision-making module, the planning and control module, or the vehicle status monitoring module. These modules can adjust perception parameters, obstacle confidence, speed planning strategies, following distance, or safety redundancy strategies based on the steady-state rainfall level. For example, when the steady-state rainfall level is moderate or heavy rain, the vehicle can increase the intensity of point cloud noise reduction processing, reduce the confidence weight for weakly reflective targets in a single frame, or increase the distance setting value in the following distance strategy.

[0064] After completing the processing of the current frame, the system clears the temporary count parameters for the current frame, including the number of suspected raindrop points, the number of valid raindrop points, and the intermediate results of the current frame's neighborhood search, and waits for the input of the next frame's original 3D point cloud data. Once the next frame's data arrives, the system repeats the above processing procedure to achieve continuous iterative operation.

[0065] Figure 4 This is a schematic diagram of a rainfall level determination system based on laser point cloud time-series steady-state analysis. (Reference) Figure 4 The system may include a point cloud preprocessing and quality verification module, a multi-dimensional raindrop feature initial screening module, a spatial neighborhood sparsity verification module, a single-frame instantaneous rainfall level generation module, a time-series sliding cache update module, and a steady-state mode correction and level output module.

[0066] The point cloud preprocessing and quality verification module is used to receive the raw 3D point cloud data output by the vehicle-mounted LiDAR, and to perform format conversion, invalid point removal and quality verification on the raw 3D point cloud data to obtain valid point cloud data.

[0067] The multi-dimensional raindrop feature screening module is used to read the spatial distance and laser reflection intensity of each point in the effective point cloud data, and to screen suspected raindrop points according to the preset raindrop detection spatial range and preset reflection intensity screening interval.

[0068] The spatial neighborhood sparsity verification module is used to establish a KD-Tree tree index structure based on effective point cloud data, perform three-dimensional spatial neighborhood retrieval on suspected raindrop points, determine effective raindrop points based on the number of neighborhood points and preset sparsity judgment conditions, and count the number of effective raindrops in the current frame.

[0069] The single-frame instantaneous rainfall level generation module is used to read the preset multi-level rainfall classification standard and determine the instantaneous rainfall level of the current frame based on the number of valid raindrops in the current frame.

[0070] The time-series sliding cache update module is used to write the instantaneous rainfall level of the current frame into the time-series sliding cache queue. When the queue reaches the preset time-series cache window capacity, it performs a new and old data overwrite iteration to obtain a continuous frame rainfall level sequence.

[0071] The steady-state mode correction and grade output module is used to count the number of occurrences of each rainfall grade in a continuous frame rainfall grade sequence, determine the steady-state rainfall grade, and output the steady-state rainfall grade as the current rainfall grade determination result.

[0072] The system in this embodiment can be deployed in an in-vehicle computing platform, an autonomous driving domain controller, a LiDAR processing unit, or a perception fusion processor. The functional modules can be integrated into the same processor or distributed across different processing units. The above module division is only a logical functional division and should not be construed as a limitation on the hardware structure.

[0073] This application may also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described rainfall level determination method based on laser point cloud time-series steady-state analysis. The computer-readable storage medium may include a read-only memory, random access memory, flash memory, hard disk, solid-state drive, or other media capable of storing a computer program.

[0074] The above description is merely a preferred embodiment of this application. It should be noted that, due to the limitations of written expression, there are objectively infinite specific structures. For those skilled in the art, without departing from the principles of this application, several improvements, modifications, or variations can be made, and the above technical features can be combined in an appropriate manner. These improvements, modifications, variations, or combinations, or the direct application of the concept and technical solution of this application to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for determining rainfall level based on laser point cloud time-series steady-state analysis, characterized in that, include: Receive the raw 3D point cloud data output by the vehicle-mounted LiDAR, and preprocess the raw 3D point cloud data to obtain effective point cloud data; Based on the spatial distance and laser reflection intensity of each point in the effective point cloud data, suspected raindrop points are screened from the effective point cloud data; A three-dimensional spatial neighborhood search is performed on the suspected raindrops, and the sparsity of the suspected raindrops is checked based on the neighborhood search results. The number of valid raindrops in the current frame is counted. The instantaneous rainfall level of the current frame is determined based on the number of valid raindrops in the current frame and the preset multi-level rainfall classification standard. Write the instantaneous rainfall level of the current frame into a time-series sliding buffer queue to obtain a continuous frame rainfall level sequence; The mode of the continuous frame rainfall level sequence is statistically analyzed to determine the steady-state rainfall level, and the steady-state rainfall level is output as the current rainfall level determination result.

2. The method for determining rainfall level based on laser point cloud time-series steady-state analysis according to claim 1, characterized in that, The original 3D point cloud data undergoes format conversion, invalid point removal, and quality verification, including: Read the three-dimensional coordinates and laser reflection intensity of each point in the original three-dimensional point cloud data, and convert the original three-dimensional point cloud data into a unified point cloud structure; Remove blank frames, invalid coordinates, points with missing reflection intensity, and near-field vehicle noise within the preset vehicle interference range; When the number of remaining points and the integrity of the fields meet the preset point cloud quality conditions, the remaining points are determined as the valid point cloud data.

3. The rainfall level determination method based on laser point cloud time-series steady-state analysis according to claim 1, characterized in that, Based on the spatial distance and laser reflection intensity of each point in the effective point cloud data, suspected raindrop points are screened from the effective point cloud data, including: The corresponding spatial distance is determined based on the three-dimensional coordinates of each point in the effective point cloud data; The spatial distance is compared with a preset raindrop detection spatial range, and the laser reflection intensity is compared with a preset reflection intensity screening range; When the same point is simultaneously located within the preset raindrop detection space range and the preset reflection intensity screening range, the point is identified as the suspected raindrop point.

4. The rainfall level determination method based on laser point cloud time-series steady-state analysis according to claim 1, characterized in that, Perform a three-dimensional spatial neighborhood search on the suspected raindrops, and perform sparsity verification on the suspected raindrops based on the neighborhood search results, including: A KD-Tree tree index structure is established based on the effective point cloud data; with each suspected raindrop point as the retrieval center, neighboring points are retrieved within a preset neighborhood search range and the number of neighboring points is obtained; The number of neighboring points is compared with a preset sparsity determination condition, and suspected raindrop points that meet the preset sparsity determination condition are counted as valid raindrop points.

5. The rainfall level determination method based on laser point cloud time-series steady-state analysis according to claim 4, characterized in that, The counting of valid raindrops in the current frame includes: The effective raindrop points that satisfy the preset sparsity determination condition in the current frame are accumulated to obtain the effective raindrop count in the current frame; Suspected raindrops that do not meet the preset sparsity determination criteria are not included in the current frame's valid raindrop count; After obtaining the number of valid raindrops in the current frame, clear the temporary count parameter for the current frame.

6. The method for determining rainfall level based on laser point cloud time-series steady-state analysis according to claim 1, characterized in that, Determine the instantaneous rainfall level for the current frame, including: Read the preset multi-level rainfall classification standard, which includes the effective raindrop quantity range corresponding to no rain level, light rain level, moderate rain level and heavy rain level respectively; The effective number of raindrops in the current frame is compared with the effective raindrop number range level by level, and the rainfall level that is matched by the comparison is determined as the instantaneous rainfall level of the current frame.

7. The rainfall level determination method based on laser point cloud time-series steady-state analysis according to claim 1, characterized in that, The continuous frame rainfall level sequence is obtained, including: Read the preset timing buffer window capacity; When the time-series sliding buffer queue does not reach the preset time-series buffer window capacity, the instantaneous rainfall level of the current frame is appended to the time-series sliding buffer queue; When the time-series sliding buffer queue reaches the preset time-series buffer window capacity, the earliest written rainfall level is deleted and the instantaneous rainfall level of the current frame is written to obtain an updated continuous frame rainfall level sequence.

8. The method for determining rainfall level based on laser point cloud time-series steady-state analysis according to claim 1, characterized in that, Determining steady-state rainfall levels includes: The frequency of occurrence of each rainfall level in the continuous frame rainfall level sequence is counted. The rainfall level that occurs most frequently is determined as the steady-state rainfall level; When at least two rainfall levels occur the same number of times, the steady-state rainfall level output from the previous frame is maintained, or the steady-state rainfall level is determined from the rainfall levels that occur the same number of times according to a preset level priority rule.

9. The method for determining rainfall level based on laser point cloud time-series steady-state analysis according to claim 2, characterized in that, When the current frame fails the quality check, spatial distance and laser reflection intensity filtering are not performed, the number of valid raindrops in the current frame is not generated, and the instantaneous rainfall level of the current frame for writing to the time-series sliding buffer queue is not generated. Instead, the steady-state rainfall level output from the previous frame is used. When the number of consecutive frames that fail quality verification reaches a preset unusable frame threshold, the steady-state rainfall level will stop being updated based on consecutive frames that fail quality verification.

10. The method for determining rainfall level based on laser point cloud time-series steady-state analysis according to claim 6, characterized in that, The preset multi-level rainfall classification standard is generated through sample calibration, including: The distribution of the number of valid raindrops in the current frame corresponding to the labeled samples of no rain, light rain, moderate rain, and heavy rain is counted respectively. A first number threshold, a second number threshold, and a third number threshold are determined based on the boundary interval of the number distribution between adjacent labeled rainfall levels. The first number threshold, the second number threshold, and the third number threshold are written into the preset multi-level rainfall classification standard.

Images