Dynamic point cloud filtering method and system, terminal and storage medium

Abstract

The invention discloses a dynamic point cloud filtering method and system, a terminal and a storage medium. The method comprises the following steps: acquiring laser radar point cloud data; counting point cloud time span characteristics of the laser radar point cloud data by adopting a neighborhood time distribution analysis method, and preliminarily screening dynamic point clouds according to the point cloud time span characteristics; retrieving marked static point clouds in a neighborhood space where the preliminarily screened dynamic point clouds are located based on a static graph reverse check verification mechanism, verifying the retrieved static point clouds through a secondary time distribution method, and correcting misjudgment points; and outputting the point cloud map after dynamic point cloud filtering. The method improves the precision of dynamic point cloud filtering while maintaining the processing speed, and is significantly superior to a traditional ray tracing and geometric difference method.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and in particular to a dynamic point cloud filtering method, system, terminal, and storage medium. Background Technology

[0002] During autonomous driving and drone operation, point cloud maps need to be built using LiDAR. However, when LiDAR scans continuously, dynamic objects (such as pedestrians and vehicles) will leave trails in the global map, causing distortion of the static environment model.

[0003] Currently, mobile platforms (e.g., autonomous vehicles, drones) mainly rely on three types of technical solutions to address the problem of dynamic point clouds when constructing point cloud maps in dynamic scenes, but all of them have significant drawbacks: 1) Occupation grid method (e.g., OctoMap, a 3D probabilistic map representation method based on octree): Constructs a probabilistic voxel model through ray projection. Although it can identify changes in spatial occupancy, it has high computational complexity, is time-consuming to process a single frame, and has a high rate of missed detection in dynamic point clouds due to broken ray paths in sparse point cloud scenarios. 2) Geometric difference method (e.g., ERASOR, a dynamic obstacle removal algorithm based on pseudo occupancy ratio): It uses the coordinate alignment difference detection between real-time point cloud and global map to improve timeliness, but the accumulated coordinate transformation error will mistakenly delete static features (e.g., traffic signs, low guardrails), destroying the integrity of the map. 3) Deep learning methods: These methods utilize neural networks to learn motion patterns. While they achieve high dynamic point cloud recognition rates on specific datasets (e.g., the SemanticKITTI dataset), differences in sensor parameters (e.g., mechanical / solid-state LiDAR scanning modes) can lead to a sharp drop in accuracy. Furthermore, model deployment requires GPU memory, making it difficult to embed into automotive platforms. These shortcomings collectively prevent existing technologies from simultaneously achieving real-time performance, accuracy, and cross-platform generalization capabilities.

[0004] Therefore, existing technologies still need improvement. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a dynamic point cloud filtering method, system, terminal and storage medium to address the shortcomings of existing technologies, thereby solving the problems of low accuracy and low efficiency in existing map construction methods based on dynamic scene point clouds.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows: In a first aspect, the present invention provides a dynamic point cloud filtering method, comprising: Acquire lidar point cloud data; The neighborhood time distribution analysis method is used to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and the dynamic point cloud is initially screened based on the point cloud time span characteristics. Based on the static graph reverse lookup verification mechanism, the labeled static point cloud in the neighborhood space of the initially screened dynamic point cloud is retrieved, and the retrieved static point cloud is verified by the secondary time distribution method to correct misjudged points. Output a point cloud map after dynamic point cloud filtering.

[0007] In one implementation, acquiring the lidar point cloud data includes: Record the start time t_1 of the current frame scan and the start time t_2 of the next frame scan. Based on the physical scanning law of the laser beam, linearly map the single frame scan period [t_1, t_2] according to the point cloud sequence to obtain lidar point cloud data with timestamps.

[0008] In one implementation, the method of employing neighborhood time distribution analysis to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and to perform preliminary screening of the dynamic point cloud based on the point cloud time span characteristics, includes: The point cloud in the lidar point cloud data is divided into multiple local spatial domains with radius r1. For each local spatial domain, determine the maximum and minimum timestamps in all point clouds within the domain, and calculate the first distribution span value of the timestamps within the domain based on the maximum and minimum timestamps; Based on the calculated first distribution span value, the point cloud in the corresponding local spatial domain is classified as either a dynamic point cloud or a static point cloud, thus obtaining preliminary screening results.

[0009] In one implementation, classifying the point cloud within the corresponding local spatial domain into a dynamic point cloud or a static point cloud based on the calculated first distribution span value includes: Determine the preliminary classification thresholds for the dynamic point cloud and the static point cloud; Compare the first distribution span value corresponding to each of the local spatial domains with the preliminary classification threshold; If the value is less than the preliminary classification threshold, then the point cloud in the corresponding local spatial domain is classified as the dynamic point cloud. If the value is greater than or equal to the preliminary classification threshold, then the point cloud in the corresponding local spatial domain is classified as the static point cloud.

[0010] In one implementation, the static graph-based reverse lookup verification mechanism retrieves the labeled static point clouds within the neighborhood of the initially screened dynamic point clouds, and verifies the retrieved static point clouds using a quadratic time distribution method to correct misjudged points, including: The dynamic point cloud and the neighboring static point cloud after the initial screening are divided to obtain multiple spherical neighborhood spaces; Retrieve labeled static point clouds within each spherical neighborhood space and calculate the second distribution span value of the timestamps of the retrieved static point clouds; Based on the calculated second distribution span value, the retrieved static point cloud is verified using the quadratic time distribution method to correct the misjudged dynamic point cloud and obtain the re-screening results.

[0011] In one implementation, the preliminary screening of the dynamic point cloud and the neighboring static point cloud is divided to obtain multiple spherical neighborhood spaces, including: Centered on each dynamic point cloud, a spherical neighborhood space with radius r2 is constructed in the initially screened static point cloud to obtain multiple spherical neighborhood spaces.

[0012] In one implementation, the step of verifying the retrieved static point cloud using a quadratic time distribution method based on the calculated second distribution span value, and correcting misjudged dynamic point clouds, includes: Determine the reclassification threshold between the dynamic point cloud and the static point cloud; The second distribution span value corresponding to each of the spherical neighborhood spaces is compared with the reclassification threshold; If the value is less than the reclassification threshold, the dynamic attribute determination of the dynamic point cloud at the center position of the corresponding spherical neighborhood space is maintained. If the value is greater than or equal to the reclassification threshold, the dynamic point cloud at the center of the corresponding spherical neighborhood space will be reclassified as a static point cloud.

[0013] Secondly, the present invention provides a dynamic point cloud filtering system, comprising: The data acquisition module is used to acquire point cloud data from the LiDAR. The preliminary screening module is used to use the neighborhood time distribution analysis method to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and to perform preliminary screening of the dynamic point cloud based on the point cloud time span characteristics. The re-filtering module is used to retrieve the labeled static point cloud in the neighborhood space of the initially filtered dynamic point cloud based on the static graph reverse lookup verification mechanism, and to verify the retrieved static point cloud through the secondary time distribution method to correct misjudged points. This module is used to output a point cloud map after dynamic point cloud filtering.

[0014] Thirdly, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory stores a dynamic point cloud filtering program, and the dynamic point cloud filtering program, when executed by the processor, is used to implement the operation of the dynamic point cloud filtering method as described in the first aspect.

[0015] Fourthly, the present invention also provides a medium, which is a computer-readable storage medium, storing a dynamic point cloud filtering program, which, when executed by a processor, is used to implement the operation of the dynamic point cloud filtering method as described in the first aspect.

[0016] The present invention, by employing the above technical solution, has the following effects: This invention analyzes dynamic and static points through neighborhood time span in the coarse classification stage, thereby quickly filtering out significant dynamic points and avoiding the computational burden of traditional ray tracing. In the fine classification stage, it introduces a static graph reverse verification mechanism, which uses the spatiotemporal stability of confirmed static points to perform secondary discrimination on suspected dynamic points, correcting misclassified dynamic points and effectively reducing the probability of misclassification of dynamic points. This invention improves the accuracy of dynamic point cloud filtering while maintaining processing speed, and is significantly superior to traditional ray tracing and geometric difference methods. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the dynamic point cloud filtering method in this invention.

[0019] Figure 2 This is a flowchart of the overall solution of the present invention.

[0020] Figure 3 This is a schematic diagram of the system architecture of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the effect of the dynamic point cloud filtering method in this invention.

[0022] Figure 5 This is a functional schematic diagram of the terminal in one implementation of the present invention.

[0023] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] Exemplary methods Currently, mobile platforms (e.g., autonomous vehicles, drones) mainly rely on three types of technical solutions to address the problem of dynamic point clouds when constructing point cloud maps in dynamic scenes, but all of them have significant drawbacks: 1) Occupation grid method (e.g., OctoMap, a 3D probabilistic map representation method based on octree): Constructs a probabilistic voxel model through ray projection. Although it can identify changes in spatial occupancy, it has high computational complexity, is time-consuming to process a single frame, and has a high rate of missed detection in dynamic point clouds due to broken ray paths in sparse point cloud scenarios. 2) Geometric difference method (e.g., ERASOR, a dynamic obstacle removal algorithm based on pseudo occupancy ratio): It uses the coordinate alignment difference detection between real-time point cloud and global map to improve timeliness, but the accumulated coordinate transformation error will mistakenly delete static features (e.g., traffic signs, low guardrails), destroying the integrity of the map. 3) Deep learning methods: These methods utilize neural networks to learn motion patterns. While they achieve high dynamic point cloud recognition rates on specific datasets (e.g., the SemanticKITTI dataset), differences in sensor parameters (e.g., mechanical / solid-state LiDAR scanning modes) can lead to a sharp drop in accuracy. Furthermore, model deployment requires GPU memory, making it difficult to embed into automotive platforms. These shortcomings collectively prevent existing technologies from simultaneously achieving real-time performance, accuracy, and cross-platform generalization capabilities.

[0026] To address the above technical problems, this invention provides a dynamic point cloud filtering method. The method includes: acquiring lidar point cloud data; employing a neighborhood time distribution analysis method to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and performing preliminary screening of the dynamic point cloud based on these characteristics; using a static graph reverse lookup verification mechanism to retrieve labeled static point clouds within the neighborhood space of the preliminarily screened dynamic point clouds, and verifying the retrieved static point clouds using a secondary time distribution method to correct misjudged points; and outputting a point cloud map after dynamic point cloud filtering. This invention improves the accuracy of dynamic point cloud filtering while maintaining processing speed, significantly outperforming traditional ray tracing and geometric difference methods.

[0027] like Figure 1 As shown, this embodiment of the invention provides a dynamic point cloud filtering method, including the following steps: Step S100: Obtain lidar point cloud data.

[0028] This embodiment provides a filtering method based on a dynamic point cloud filter, which achieves dynamic interference filtering by analyzing the spatiotemporal distribution differences of point clouds. The core technology includes: (1) Dynamic point cloud filtering is performed by utilizing the characteristics of wide timestamp distribution in static point clouds and narrow distribution in dynamic point clouds; (2) Two-level classification architecture: coarse classification quickly filters out dynamic points, and fine classification corrects misjudgments by checking the static map, thereby improving the accuracy of dynamic point cloud filtering.

[0029] The method provided in this embodiment primarily addresses the core need of intelligent systems such as autonomous vehicles, mobile robots, and drones to construct high-precision environmental maps in dynamic scenes. It focuses on solving the point cloud map distortion problem caused by trajectory afterimages generated by continuous scanning of moving objects (such as vehicles and pedestrians). By innovatively mining the timestamp distribution patterns in point cloud data, it achieves efficient filtering of dynamic interference. This method can be widely applied to fields such as laser SLAM mapping optimization for autonomous driving systems, safe navigation of mobile robots in human-machine coexistence environments, high-precision static map construction, and AR / VR real-scene reconstruction. It overcomes the inherent limitations of traditional geometric difference methods and deep learning models in terms of computational efficiency and generalization, providing a new technological paradigm for dynamic environment perception.

[0030] In this embodiment, to implement the dynamic point cloud filtering method, a dynamic point cloud filtering system is also provided, such as... Figure 2 As shown, the system comprises three parts: a point cloud timestamp acquisition module, a coarse classifier for dynamic / static point clouds, and a fine classifier for dynamic / static point clouds. The point cloud timestamp acquisition module assigns precise acquisition time information to each 3D point in the point cloud map, forming spatiotemporal four-dimensional point cloud data (x, y, z, t). The coarse classifier for dynamic / static point clouds enables rapid initial screening of dynamic points, resulting in coarse-classified dynamic point cloud maps and coarse-classified static point cloud maps. The fine classifier for dynamic / static point clouds identifies and corrects misclassified points through secondary time distribution verification, resulting in fine-classified dynamic point cloud maps and fine-classified static point cloud maps.

[0031] Specifically, in one implementation of this embodiment, step S100 includes the following steps: Step S101: Record the start time t_1 of the current frame scan and the start time t_2 of the next frame scan. Based on the physical scanning law of the laser beam, linearly map the single frame scan period [t_1, t_2] according to the point cloud sequence to obtain the lidar point cloud data with timestamps.

[0032] Accurate acquisition of point cloud timestamps is fundamental for dynamic filtering during continuous LiDAR scanning. In this embodiment, point cloud timestamps are acquired using the aforementioned point cloud timestamp acquisition module to obtain timestamped LiDAR point cloud data. The specific acquisition method is as follows: A linear interpolation method is used to assign a time stamp to each frame of the point cloud: First, the start time t_1 of the current frame scan and the start time t_2 of the next frame are recorded. Based on the physical scanning characteristics of the laser beam (such as the constant angular velocity of a rotating radar), the single-frame scan period [t_1, t_2] is linearly mapped according to the point cloud sequence. Specifically, for the k-th point in a single frame (out of a total of N points), its timestamp t_k is determined by the linear interpolation method, as shown in the following formula: t_k = t_1 + (t_2 - t_1) * (k / (N - 1)); This method fully considers the scanning continuity of lidar, avoids pulse-type time allocation errors, and provides a sub-second precision time reference for subsequent spatiotemporal distribution analysis, fundamentally ensuring the reliability of dynamic point identification.

[0033] like Figure 1 As shown, this embodiment of the invention provides a dynamic point cloud filtering method, including the following steps: Step S200: Using the neighborhood time distribution analysis method, the point cloud time span characteristics of the lidar point cloud data are statistically analyzed, and the dynamic point cloud is initially screened based on the point cloud time span characteristics.

[0034] In this embodiment, after obtaining the lidar point cloud data with timestamps, the lidar point cloud data is input into a coarse classifier for preliminary screening. This process mainly adopts the neighborhood time distribution analysis method to statistically analyze the time span characteristics of the point cloud in a local spherical space. Based on the essential difference that static targets continuously exist and form a wide time distribution, while dynamic targets appear instantaneously and form a narrow time distribution, the dynamic points are quickly screened.

[0035] In this embodiment, the principle of using the different time distribution characteristics of static and dynamic targets to achieve rapid initial screening of dynamic points is as follows: Because the timestamp distributions of static elements and dynamic objects differ fundamentally during the continuous scanning and point cloud map construction process of LiDAR, this characteristic forms the core basis for dynamic filtering. Static objects (such as roads and buildings) have fixed spatial locations, and the same spatial coordinate point is repeatedly captured by multiple consecutive scans, resulting in a point cloud timestamp covering a large time span (for example, a point on the ground can accumulate samples from four different times within a 4-second scan cycle, forming a wide temporal distribution band). Dynamic objects (such as vehicles and pedestrians) occupy a specific spatial location for only a very short time; their point clouds are composed of a small number of consecutive frames (usually 1-2 frames) of instantaneous scanning, with timestamps highly concentrated within a short time window, forming peak-shaped time clusters. For example... Figure 3As shown, the timestamps of static point clouds exhibit a diffuse distribution along the time axis (similar to low-frequency background noise), while dynamic targets appear as isolated narrowband pulses (similar to time-domain abrupt signals). This difference in distribution stems from the kinematic nature of the problem: the spatiotemporal coordinates of static elements are invariant, while dynamic targets exhibit transient coupling characteristics of their spatiotemporal coordinates. This embodiment is based on this physical law; by quantifying the timestamp dispersion of local point cloud clusters, the separation of static and dynamic point clouds can be achieved without relying on easily invalidated geometric assumptions or computationally time-consuming ray tracing.

[0036] Specifically, in one implementation of this embodiment, step S200 includes the following steps: Step S201: Divide the point cloud in the lidar point cloud data with r1 as the radius to obtain multiple local spatial domains; Step S202: For each local spatial domain, determine the maximum and minimum timestamps in all point clouds within the domain, and calculate the first distribution span value of the timestamps within the domain based on the maximum and minimum timestamps. Step S203: Based on the calculated first distribution span value, classify the point cloud in the corresponding local spatial domain into dynamic point cloud or static point cloud to obtain preliminary screening results.

[0037] In this embodiment, after obtaining the lidar point cloud data with timestamps, this lidar point cloud data (i.e., point cloud map) is input into a coarse classifier. This module will perform spherical neighborhood analysis on the point cloud: First, it traverses the lidar points in the point cloud map, and constructs a local spatial domain with each point as the center and a radius r1. The radius r1 can be a randomly selected radius value within the range of the point cloud map. For example, the random radius generation method is to determine the center coordinates and the maximum radius range, and then generate a random angle of 0-2π through a random number generator. Combined with trigonometric functions, the coordinates of the random points are calculated, and the radius r1 can be obtained based on the coordinates of the random points.

[0038] It is understood that in this embodiment, the point cloud map is divided with a radius r1 to obtain multiple spherical regions with a radius of r1. Then, each spherical region is used as the calculation area for the time distribution span value C1 of the laser point at its center. In the process of calculating the time distribution span value C1, the difference between the maximum and minimum timestamps in the region is defined as the time distribution span value C1 (i.e., the first distribution span value) in the current spherical region. Based on this time distribution span value C1, the point cloud in the corresponding local spatial domain is classified as a dynamic point cloud or a static point cloud.

[0039] Specifically, in one implementation of this embodiment, step S203 includes the following steps: Step S203a: Determine the preliminary classification thresholds for the dynamic point cloud and the static point cloud; Step S203b, compare the first distribution span value corresponding to each local spatial domain with the preliminary classification threshold; Step S203c, if it is less than the preliminary classification threshold, classify the point cloud within the corresponding local spatial domain as the dynamic point cloud; Step S203d, if it is greater than or equal to the preliminary classification threshold, classify the point cloud within the corresponding local spatial domain as the static point cloud.

[0040] Specifically, in this embodiment, after calculating the first distribution span value C1 of the timestamps within the local spatial domain, set the dynamic / static classification threshold as t1 (i.e., the preliminary classification threshold). Herein, the dynamic / static classification threshold t1 can be selected according to random values. For example, it is dynamically adjusted based on the first distribution span value C1. The random range of the dynamic / static classification threshold t1 is set as [C1×α, C1×β], where α and β are empirical coefficients in the range of 0.5 - 1.5.

[0041] After setting the dynamic / static classification threshold t1, compare the first distribution span value C1 with the dynamic / static classification threshold t1. If C1 < t1, determine that the neighborhood point cloud (i.e., the point cloud at the center of the local spatial domain) is a dynamic point cloud; otherwise, classify it as a static point cloud.

[0042] This embodiment is based on a coarse classifier and uses the neighborhood time span analysis method to quickly filter out significantly dynamic points, avoiding the computational burden of traditional ray tracing, and can achieve a quick preliminary screening of dynamic points, thereby obtaining a coarsely classified dynamic point cloud map and a coarsely classified static point cloud map.

[0043] As Figure 1 shown, an embodiment of the present invention provides a dynamic point cloud filtering method, including the following steps: Step S300, based on the static map reverse lookup verification mechanism, retrieve the marked static point clouds within the neighborhood space where the preliminarily screened dynamic point cloud is located, and verify the retrieved static point clouds through the secondary time distribution method to correct misjudged points.

[0044] In this embodiment, input the coarsely classified dynamic point cloud map and the coarsely classified static point cloud map obtained through preliminary screening into a fine classifier for re-screening. This process mainly introduces a static map reverse lookup verification mechanism, retrieves the marked static point clouds within the neighborhood space of the initially screened dynamic points, and identifies and corrects misjudged points through secondary time distribution verification, thereby improving the dynamic point cloud filtering accuracy.

[0045] Specifically, in an implementation manner of this embodiment, Step S300 includes the following steps: Step S301, divide the preliminarily screened dynamic point cloud and the neighborhood static point cloud to obtain multiple spherical neighborhood spaces; In one implementation of this embodiment, step S301 includes the following steps: Step S301a: Using each dynamic point cloud as the center, construct a spherical neighborhood space with a radius of r2 in the initially screened static point cloud to obtain multiple spherical neighborhood spaces.

[0046] In this embodiment, the core function of the fine classifier is to correct and optimize the coarse classification results, and to reclassify the misclassified dynamic point cloud as a static point cloud through a spatiotemporal consistency verification mechanism.

[0047] The specific process is as follows: Using the coarse-classified dynamic point cloud map and the coarse-classified static point cloud map output by the coarse classifier as input, a refined analysis is performed on each dynamic point cloud D_i. Centered on the dynamic point cloud D_i, a spherical neighborhood space with radius r2 is constructed in the static point cloud. Here, radius r2 can also be a randomly selected radius value within the range of the point cloud map (for example, similar to the random selection method of radius r1, the random selection range of radius r2 is smaller than the random selection range of radius r1).

[0048] Understandably, the processing method of the fine classifier in this embodiment is as follows: traverse the laser points in the point cloud map, and construct a spherical region with a radius of r2 centered on each dynamic point cloud. This r2 spherical region is used as the region for calculating the time distribution span of the laser point D_i. Through this partitioning method, multiple spherical neighborhood spaces can be obtained.

[0049] Step S302: Retrieve the labeled static point cloud within each spherical neighborhood space, and calculate the second distribution span value of the timestamp of the retrieved static point cloud; Step S303: Based on the calculated second distribution span value, the retrieved static point cloud is verified using the quadratic time distribution method to correct the misjudged dynamic point cloud and obtain the re-screening results.

[0050] In this embodiment, after constructing a spherical neighborhood space with radius r2 in the static point cloud, the labeled static point cloud is retrieved within this space domain, and its timestamp distribution span C2 (i.e., the second distribution span value C2) is calculated; where C2 is defined as the difference between the maximum and minimum timestamps of the static point cloud in the neighborhood; then, based on the calculated second distribution span value C2, the retrieved static point cloud is verified using the quadratic time distribution method to correct misjudged dynamic point clouds, thereby obtaining a refined dynamic point cloud map and a refined static point cloud map.

[0051] In one implementation of this embodiment, step S303 includes the following steps: Step S303a: Determine the reclassification threshold between the dynamic point cloud and the static point cloud; Step S303b: Compare the second distribution span value corresponding to each of the spherical neighborhood spaces with the reclassification threshold. Step S303c: If it is less than the reclassification threshold, maintain the dynamic attribute determination of the dynamic point cloud at the center position of the corresponding spherical neighborhood space. Step S303d: If it is greater than or equal to the reclassification threshold, reclassify the dynamic point cloud at the center position of the corresponding spherical neighborhood space as a static point cloud.

[0052] Specifically, in this embodiment, after calculating the second distribution span value C2, a dynamic / static reclassification threshold t2 is set. Here, the reclassification threshold t2 can be selected according to a random value (for example, similar to the selection method of the preliminary classification threshold t1). Then, compare the second distribution span value C2 with the reclassification threshold t2. If C2 < t2, it indicates that there are no static reference points with a wide time distribution in this neighborhood, and maintain the dynamic attribute determination of D_i. On the contrary, when C2 ≥ t2, it proves that there are continuously existing static feature points at this position, and at this time, reclassify D_i as a static point cloud.

[0053] In this embodiment, by introducing a static graph back-check verification mechanism, the spatio-temporal stability of the already confirmed static points is used to perform a secondary discrimination on the suspected dynamic points, effectively reducing the misclassification probability of dynamic points.

[0054] As Figure 1 shown, an embodiment of the present invention provides a method for filtering dynamic point clouds, including the following steps: Step S400: Output a point cloud map after filtering the dynamic point cloud.

[0055] In this embodiment, after the above two consecutive levels of filtering (coarse classification and fine classification), a point cloud map after filtering the dynamic point cloud is output. Of course, in order to further improve the filtering accuracy of the dynamic point cloud, multiple levels of filters (for example, the first-level classification, the second-level classification, the third-level classification) can be set for classification, or a combination with other filtering techniques (for example, the deep learning dynamic segmentation method) can be used to implement the dynamic point cloud filtering solution.

[0056] The effects of this embodiment based on the above dynamic point cloud filtering method are as follows: As Figure 4As shown, after processing the original point cloud map using the dynamic point cloud filtering method provided in this embodiment, interference point clouds generated by moving vehicles, pedestrians, and other dynamic objects were successfully eliminated. In the visualization results, the red point cloud clearly marks dynamic objects (such as moving vehicles and pedestrians), while the gray point cloud completely preserves the static environment structure (including roads, buildings, and infrastructure). The processing effect shows that this method can accurately identify and separate the trajectory afterimages generated by high-speed moving objects, while effectively filtering out transient dynamic targets, ultimately outputting a clean static environment model. Its technical value lies in the efficient removal of dynamic interference and the complete preservation of static structure, significantly improving the geometric realism and topological consistency of the point cloud map.

[0057] The solution in this embodiment can provide a positioning reference map without dynamic pollution for autonomous driving systems, ensuring the navigation reliability of vehicles in complex urban scenarios; in the field of robot path planning, it can construct an environment model without motion artifacts to optimize obstacle avoidance decisions; at the same time, it can be applied to high-precision static map construction, AR / VR real-scene reconstruction, etc., providing basic data support for digital twin cities.

[0058] In summary, this embodiment proposes for the first time that timestamp dispersion is an essential discriminative feature for dynamic and static point clouds, completely breaking through the limitations of traditional geometric constraints and data-driven paradigms. By revealing the physical laws that static objects exhibit a wide temporal distribution due to continuous accumulation over multiple frames, while dynamic targets form a narrow temporal cluster due to instantaneous spatial occupation, it innovatively transforms spatiotemporal distribution characteristics into a quantifiable classification criterion.

[0059] The two-tiered collaborative architecture established based on this enables efficient and accurate processing: the coarse classification stage quickly filters out significant dynamic points through neighborhood time span analysis, avoiding the computational burden of traditional ray tracing; the fine classification stage introduces a static graph reverse verification mechanism, utilizing the spatiotemporal stability of confirmed static points to perform secondary discrimination on suspected dynamic points, effectively reducing the probability of misclassification of dynamic points. This architecture forms a closed-loop processing flow from rapid initial screening to accurate error correction, enabling rapid response to moving targets and providing a technical path for dynamic environment perception that combines theoretical innovation and engineering practicality, laying the foundation for highly reliable environmental modeling in scenarios such as autonomous driving and robot navigation.

[0060] This embodiment achieves the following technical effects through the above technical solution: This embodiment analyzes dynamic and static points through neighborhood time span analysis in the coarse classification stage, thereby quickly filtering out significant dynamic points and avoiding the computational burden of traditional ray tracing. In the fine classification stage, a static graph reverse verification mechanism is introduced, which uses the spatiotemporal stability of confirmed static points to perform secondary discrimination on suspected dynamic points, correcting misclassified dynamic points and effectively reducing the probability of misclassification of dynamic points. This embodiment improves the accuracy of dynamic point cloud filtering while maintaining processing speed, which is significantly better than traditional ray tracing and geometric difference methods.

[0061] Exemplary device Based on the above embodiments, the present invention also provides a dynamic point cloud filtering system, comprising: The data acquisition module is used to acquire point cloud data from the LiDAR. The preliminary screening module is used to use the neighborhood time distribution analysis method to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and to perform preliminary screening of the dynamic point cloud based on the point cloud time span characteristics. The re-filtering module is used to retrieve the labeled static point cloud in the neighborhood space of the initially filtered dynamic point cloud based on the static graph reverse lookup verification mechanism, and to verify the retrieved static point cloud through the secondary time distribution method to correct misjudged points. This module is used to output a point cloud map after dynamic point cloud filtering.

[0062] This embodiment achieves the following technical effects through the above technical solution: This embodiment analyzes dynamic and static points through neighborhood time span analysis in the coarse classification stage, thereby quickly filtering out significant dynamic points and avoiding the computational burden of traditional ray tracing. In the fine classification stage, a static graph reverse verification mechanism is introduced, which uses the spatiotemporal stability of confirmed static points to perform secondary discrimination on suspected dynamic points, correcting misclassified dynamic points and effectively reducing the probability of misclassification of dynamic points. This embodiment improves the accuracy of dynamic point cloud filtering while maintaining processing speed, which is significantly better than traditional ray tracing and geometric difference methods.

[0063] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 5 As shown.

[0064] The terminal includes: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein, the processor of the terminal provides computing and control capabilities; the memory of the terminal includes a storage medium and internal memory; the storage medium stores the operating system and computer programs; the internal memory provides an environment for the operation of the operating system and computer programs in the storage medium; the interface is used to connect to external devices; the display screen is used to display relevant information; and the communication module is used to communicate with a cloud server or other devices.

[0065] This computer program is executed by the processor to implement the dynamic point cloud filtering method.

[0066] It will be understood by those skilled in the art that Figure 5 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0067] In one embodiment, a terminal is provided, comprising: a processor and a memory, the memory storing a dynamic point cloud filtering program, which, when executed by the processor, is used to implement the operation of the dynamic point cloud filtering method described above.

[0068] In one embodiment, a storage medium is provided, wherein the storage medium stores a dynamic point cloud filtering program, which, when executed by a processor, is used to implement the operation of the dynamic point cloud filtering method described above.

[0069] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include both non-volatile and volatile memory.

[0070] In summary, this invention provides a dynamic point cloud filtering method, system, terminal, and storage medium, comprising: acquiring lidar point cloud data; employing a neighborhood time distribution analysis method to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and performing preliminary screening of the dynamic point cloud based on the point cloud time span characteristics; based on a static graph reverse lookup verification mechanism, retrieving the labeled static point clouds in the neighborhood space of the preliminarily screened dynamic point clouds, and verifying the retrieved static point clouds using a secondary time distribution method to correct misjudged points; and outputting a point cloud map after dynamic point cloud filtering. This invention improves the accuracy of dynamic point cloud filtering while maintaining processing speed, significantly outperforming traditional ray tracing and geometric difference methods.

[0071] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A dynamic point cloud filtering method, characterized in that, include: Acquire lidar point cloud data; The neighborhood time distribution analysis method is used to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and the dynamic point cloud is initially screened based on the point cloud time span characteristics. Based on the static graph reverse lookup verification mechanism, the labeled static point cloud in the neighborhood space of the initially screened dynamic point cloud is retrieved, and the retrieved static point cloud is verified by the secondary time distribution method to correct misjudged points. Output a point cloud map after dynamic point cloud filtering.

2. The dynamic point cloud filtering method according to claim 1, characterized in that, The acquisition of lidar point cloud data includes: Record the start time t_1 of the current frame scan and the start time t_2 of the next frame scan. Based on the physical scanning law of the laser beam, linearly map the single frame scan period [t_1, t_2] according to the point cloud sequence to obtain lidar point cloud data with timestamps.

3. The dynamic point cloud filtering method according to claim 1, characterized in that, The method employs neighborhood time distribution analysis to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and performs preliminary screening of the dynamic point cloud based on these characteristics, including: The point cloud in the lidar point cloud data is divided into multiple local spatial domains with radius r1. For each local spatial domain, determine the maximum and minimum timestamps in all point clouds within the domain, and calculate the first distribution span value of the timestamps within the domain based on the maximum and minimum timestamps; Based on the calculated first distribution span value, the point cloud in the corresponding local spatial domain is classified as either a dynamic point cloud or a static point cloud, thus obtaining preliminary screening results.

4. The dynamic point cloud filtering method according to claim 3, characterized in that, The step of classifying the point cloud within the corresponding local spatial domain into dynamic point cloud or static point cloud based on the calculated first distribution span value includes: Determine the preliminary classification thresholds for the dynamic point cloud and the static point cloud; Compare the first distribution span value corresponding to each of the local spatial domains with the preliminary classification threshold; If the value is less than the preliminary classification threshold, then the point cloud in the corresponding local spatial domain is classified as the dynamic point cloud. If the value is greater than or equal to the preliminary classification threshold, then the point cloud in the corresponding local spatial domain is classified as the static point cloud.

5. The dynamic point cloud filtering method according to claim 1, characterized in that, The static graph-based reverse lookup verification mechanism retrieves the labeled static point clouds within the neighborhood of the initially screened dynamic point clouds, and verifies the retrieved static point clouds using a quadratic time distribution method to correct misjudged points, including: The dynamic point cloud and the neighboring static point cloud after the initial screening are divided to obtain multiple spherical neighborhood spaces; Retrieve labeled static point clouds within each spherical neighborhood space and calculate the second distribution span value of the timestamps of the retrieved static point clouds; Based on the calculated second distribution span value, the retrieved static point cloud is verified using the quadratic time distribution method to correct the misjudged dynamic point cloud and obtain the re-screening results.

6. The dynamic point cloud filtering method according to claim 5, characterized in that, The process involves dividing the initially screened dynamic point cloud and its neighboring static point cloud to obtain multiple spherical neighborhood spaces, including: Centered on each dynamic point cloud, a spherical neighborhood space with radius r2 is constructed in the initially screened static point cloud to obtain multiple spherical neighborhood spaces.

7. The dynamic point cloud filtering method according to claim 5, characterized in that, The step of verifying the retrieved static point cloud using the quadratic time distribution method based on the calculated second distribution span value and correcting misjudged dynamic point clouds includes: Determine the reclassification threshold between the dynamic point cloud and the static point cloud; The second distribution span value corresponding to each of the spherical neighborhood spaces is compared with the reclassification threshold; If the value is less than the reclassification threshold, the dynamic attribute determination of the dynamic point cloud at the center position of the corresponding spherical neighborhood space is maintained. If the value is greater than or equal to the reclassification threshold, the dynamic point cloud at the center of the corresponding spherical neighborhood space will be reclassified as a static point cloud.

8. A dynamic point cloud filtering system, characterized in that, include: The data acquisition module is used to acquire point cloud data from the LiDAR. The preliminary screening module is used to use the neighborhood time distribution analysis method to statistically analyze the point cloud time span characteristics of the lidar point cloud data, and to perform preliminary screening of the dynamic point cloud based on the point cloud time span characteristics. The re-filtering module is used to retrieve the labeled static point cloud in the neighborhood space of the initially filtered dynamic point cloud based on the static graph reverse lookup verification mechanism, and to verify the retrieved static point cloud through the secondary time distribution method to correct misjudged points. This module is used to output a point cloud map after dynamic point cloud filtering.

9. A terminal, characterized in that, include: The processor and memory, wherein the memory stores a dynamic point cloud filtering program, which, when executed by the processor, is used to implement the operation of the dynamic point cloud filtering method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a dynamic point cloud filtering program, which, when executed by a processor, is used to implement the operation of the dynamic point cloud filtering method as described in any one of claims 1-7.

Images