Noise Filtering Method, Device and Autonomous Driving Vehicle for Point Cloud Data

Abstract

A method, apparatus and autonomous vehicle for filtering noise from point cloud data. The method includes: obtaining point cloud data collected by a radar and mapping the point cloud data into a three-dimensional voxel grid; determining the number of point cloud points corresponding to each three-dimensional voxel grid and filtering out the point cloud data in the three-dimensional voxel grids where the number of point cloud points is less than a first threshold; calculating the spatial volume ratio of each three-dimensional voxel grid and marking the three-dimensional voxel grids with a spatial volume less than a second threshold as potential small object grids; performing a diffusion search on the potential small object grids to determine the pseudo-small object grids connected to the entity obstacle grids in the potential small object grids, and marking the grids other than the pseudo-small object grids as real small object grids. The noise filtering method of the present invention can be executed in parallel by an in-vehicle image processing unit, greatly improving the operation speed.

Description

Technical Field

[0001] The present invention relates to the field of autonomous vehicles, and more particularly, to a method and apparatus for denoising point cloud data and an autonomous vehicle. Background Art

[0002] In recent years, autonomous driving technology has developed rapidly, and the ability to accurately avoid obstacles is particularly important in complex scenarios. With the improvement of sensor accuracy, the granularity that autonomous vehicles can perceive is getting smaller and smaller, and the data that can be perceived is getting more and more. However, among many data, there are multiple noise data that have nothing to do with the obstacle avoidance of the host vehicle. At this time, if fast denoising operations can be performed, it can not only reduce the computational workload of downstream modules, but also reduce the emergency braking of the vehicle due to noise without affecting the safety of autonomous driving, and improve the smoothness of the vehicle's autonomous driving. Summary of the Invention

[0003] A series of simplified concepts are introduced in the Summary of the Invention section, which will be further elaborated in the Detailed Description section. The Summary of the Invention section of the present invention does not mean to attempt to define the key features and essential technical features of the claimed technical solution, nor does it mean to attempt to determine the protection scope of the claimed technical solution.

[0004] In view of the deficiencies of the prior art, a first aspect of an embodiment of the present invention provides a method for denoising point cloud data, which is used for an in-vehicle image processing unit, and the method includes:

[0005] Obtain the point cloud data collected by the radar and map the point cloud data into a three-dimensional voxel grid;

[0006] Determine the number of point cloud points corresponding to each voxel grid in the three-dimensional voxel grid, and filter out the point cloud data corresponding to the voxel grids with the number of point cloud points less than a first threshold;

[0007] Calculate the spatial volume ratio of each voxel grid, and mark the voxel grids with a spatial volume less than a second threshold as potential small object grids;

[0008] Perform a diffusion search on the potential small object grids to determine the pseudo-small object grids connected to the entity obstacle grids in the potential small object grids, mark the potential small object grids other than the pseudo-small object grids as real small object grids, and mark the point cloud data corresponding to the real small object grids as noise data.

[0009] In some embodiments, the method further includes: filtering out the point cloud data corresponding to the real small object grids according to the mapping relationship between the point cloud data and the three-dimensional voxel grid.

[0010] In some embodiments, the method further includes: the diffusion search for the potential small object grids to determine the pseudo small object grids connected to the entity obstacle grids in the potential small object grids, including:

[0011] Performing one or more convolutional diffusions on the potential small object grids to connect the pseudo small object grids in the potential small object grids to the entity obstacle grids;

[0012] In some embodiments, the method further includes: the diffusion search for the potential small object grids to determine the pseudo small object grids connected to the entity obstacle grids in the potential small object grids, including:

[0013] Performing one or more convolutional diffusions on the entity obstacle grids to connect the entity obstacle grids to the pseudo small object grids in the potential small object grids.

[0014] In some embodiments, the method further includes: the calculation of the spatial volume ratio of each voxel grid, including:

[0015] Performing three-dimensional convolution on each voxel grid to obtain the spatial volume ratio of the voxel grid.

[0016] In some embodiments, the method further includes: the mapping of the point cloud data into the three-dimensional voxel grid, including:

[0017] Establishing the mapping relationship between the point cloud data and the three-dimensional voxel grid in parallel through multiple parallel operation units of the vehicle-mounted image processing unit.

[0018] In some embodiments, the method further includes: determining the number of point cloud points corresponding to each voxel grid in the three-dimensional voxel grid and filtering the point cloud data corresponding to the voxel grids with the number of point cloud points less than the first threshold, including:

[0019] Traversing each voxel grid in parallel through multiple parallel operation units of the vehicle-mounted image processing unit to filter the point cloud data corresponding to the voxel grids with the number of point cloud points less than the first threshold.

[0020] In some embodiments, the method further includes: calculating the spatial volume ratio of each voxel grid and marking the voxel grids with the spatial volume less than the second threshold as potential small object grids, including:

[0021] Traversing each voxel grid in parallel through multiple parallel operation units of the vehicle-mounted image processing unit to mark the potential small object grids.

[0022] In some embodiments, the method further includes: the performing diffusion search on the potential small object grids, including:

[0023] Performing the diffusion search on each of the potential small object grids in parallel through a plurality of parallel operation units of the vehicle-mounted image processing unit.

[0024] A second aspect of the embodiments of the present invention provides a noise filtering device for point cloud data. The device includes a memory and a graphics processing unit. A computer program is stored on the memory and is run by the graphics processing unit. When the computer program is run by the graphics processing unit, it executes the noise filtering method for point cloud data as described above.

[0025] A third aspect of the embodiments of the present invention provides an autonomous driving vehicle. The autonomous driving vehicle includes: a vehicle body; a radar mounted on the vehicle body, where the radar is used to collect point cloud data; the noise filtering device for point cloud data as described above, connected to the radar and used to execute the noise filtering method for point cloud data as described above to filter the point cloud data.

[0026] A fourth aspect of the embodiments of the present invention provides a storage medium. A computer program is stored on the storage medium. When the computer program runs, it executes the noise filtering method for point cloud data as described above.

[0027] The noise filtering method, device, autonomous driving vehicle, and storage medium of the embodiments of the present invention perform noise filtering in units of three-dimensional voxel grids. Different voxel grids are independent of each other, and the vehicle-mounted image processing unit can execute operations on different voxel grids in parallel, greatly improving the operation speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] By describing the embodiments of the present invention in more detail in conjunction with the drawings, the above and other objects, features, and advantages of the present invention will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings, the same reference numerals generally represent the same components or steps.

[0029] Figure 1 FIG. is a schematic flowchart of a noise filtering method for point cloud data according to an embodiment of the present invention;

[0030] Figure 2 FIG. is a schematic block diagram of a noise filtering device for point cloud data according to an embodiment of the present invention;

[0031] Figure 3 FIG. is a schematic block diagram of an autonomous driving vehicle according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] In order to make the objectives, technical solutions and advantages of the present application more apparent, exemplary embodiments according to the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein. Based on the embodiments of the present application described herein, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0033] In the following description, numerous specific details are given to provide a more thorough understanding of the present application. However, it is obvious to those skilled in the art that the present application can be implemented without one or more of these details. In other instances, some technical features well known to those skilled in the art are not described to avoid confusion with the present application.

[0034] It should be understood that the present application can be implemented in different forms and should not be construed as limited to the embodiments presented herein. On the contrary, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0035] The purpose of the terms used herein is only to describe specific embodiments and is not a limitation of the present application. When used herein, the singular forms "a", "an" and "the" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms "comprising" and / or "including", when used in this specification, determine the presence of the stated features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups. When used herein, the term "and / or" includes any and all combinations of the related listed items.

[0036] To thoroughly understand the present application, detailed structures will be presented in the following description to illustrate the technical solutions proposed by the present application. The optional embodiments of the present application are described in detail as follows. However, in addition to these detailed descriptions, the present application can also have other implementation manners.

[0037] For the filtering algorithm in the existing technical solutions, there are basically two solutions. The first is the filtering of sparse random noise; the second is the filtering of small objects with dense point clouds. By reasonably filtering the perceived "noise" through the filtering algorithm, non - safety obstacle avoidance is reduced and the driving smoothness of the vehicle is increased.

[0038] The point cloud data volume of autonomous driving is extremely large, and the requirements for system hardware are also very high. At the same time, various noises are generated by sensors due to environmental factors, which is not conducive to the smooth autonomous driving of vehicles. Therefore, noise is filtered before and after perception.

[0039] There are mainly two current filtering algorithms. The first one is for the sparse random noise generated inside the sensor due to the environment. This method is usually completed at the front end of the perception module. After receiving the point cloud data, filtering is performed with the point cloud data points as units. Although this method supports parallelization, it cannot filter the "noise" in specific scenarios. For example, in snowy days, rainy days, and scenes where leaves are blown down by strong winds.

[0040] The second one is for the "noise" generated by external environmental factors in specific scenarios. This method is usually filtered after the point cloud is clustered into objects during perception. However, this method can only filter the "noise" objects based on the volume and spatial position of the objects after the point cloud is clustered. During clustering, the "noise" participates in the calculation process, resulting in redundant calculations, and this method does not support parallelization.

[0041] An autonomous driving system without using filtering is prone to sudden braking due to noise. But if both of the above two filtering methods are used simultaneously, not only will redundant calculations be generated, but it will also be very time-consuming.

[0042] In view of the above problems, the embodiment of the present invention proposes a method for filtering noise from point cloud data. This algorithm can filter out the internal radar noise and small object noise irrelevant to the safety of autonomous driving from the point cloud data in an extremely short time. And this algorithm supports GPU parallel acceleration and can be deployed on the computing units with limited computing power of driverless vehicles. It can utilize the on-vehicle GPU chip for parallel acceleration without adding additional computing devices, filter non-safety noise in an extremely short time, reduce the additional downstream calculations, and increase the smoothness of autonomous driving. Next, the method for filtering noise from point cloud data and the autonomous driving vehicle proposed by the embodiment of the present invention will be described with reference to the accompanying drawings.

[0043] First, refer to Figure 1 , Figure 1 shows a schematic flowchart of a method 100 for filtering noise from point cloud data according to an embodiment of the present invention. The method 100 for filtering noise from point cloud data according to the embodiment of the present invention can be used in an autonomous driving vehicle, which can also be called a driverless vehicle, and is an intelligent vehicle that does not require a driver to perform driving operations and can automatically complete the vehicle driving task on behalf of the driver. An autonomous driving vehicle is equipped with a radar, which identifies obstacles around the vehicle through the point cloud data collected by the radar, so as to achieve obstacle avoidance. As Figure 1 shown, the method 100 for filtering noise from point cloud data according to the embodiment of the present invention includes the following steps:

[0044] In step S110, obtain the point cloud data collected by the radar and map the point cloud data into a three-dimensional voxel grid;

[0045] In step S120, determine the number of point cloud points corresponding to each voxel grid in the three-dimensional voxel grid, and filter out the point cloud data corresponding to the voxel grids with the number of point cloud points less than the first threshold;

[0046] In step S130, calculate the spatial volume ratio of each voxel grid, and mark the voxel grids with a spatial volume less than the second threshold as potential small object grids;

[0047] In step S140, perform a diffusion search on the potential small object grids to determine the pseudo small object grids connected to the entity obstacle grids in the potential small object grids, and mark the potential small object grids other than the pseudo small object grids as real small object grids.

[0048] The noise filtering method 100 for point cloud data according to the embodiment of the present invention is implemented in a vehicle-mounted image processing unit (Graphics Processing Unit, GPU), and the vehicle-mounted image processing unit is wired or wirelessly connected to the radar of the vehicle to obtain the point cloud data collected by the radar. The unique hardware structure of the GPU enables it to support large-scale parallel computing. The noise filtering method 100 according to the embodiment of the present invention is constructed based on the idea of parallelization and is particularly suitable for the parallel computing of the GPU.

[0049] Exemplarily, the radar of the vehicle can be a lidar, and the lidar can be either a lidar with regular repeated scanning or a lidar with a complex scanning trajectory with non-repeated scanning characteristics. The radar is used to sense the environmental information outside the vehicle, for example, the distance information, azimuth information, reflection intensity information, speed information, etc. of the environmental target. The point cloud data includes at least distance information and azimuth information.

[0050] As an example, the radar may include a transmitting module, a receiving module, a sampling module, and an arithmetic module. Among them, the transmitting module can transmit an optical pulse sequence (such as a laser pulse sequence). The receiving module can receive the optical pulse sequence reflected by the detected object, perform optoelectronic conversion on the optical pulse sequence to obtain an electrical signal, and then output it to the sampling module after processing the electrical signal. The sampling module can sample the electrical signal to obtain a sampling result. The arithmetic module can determine the distance between the radar and the detected object based on the sampling result of the sampling module.

[0051] In addition to the above modules, the radar may further include a scanning module configured to change the propagation direction of at least one laser pulse sequence emitted by the transmitting module and then emit it. The scanning module may include a plurality of optical elements for changing the propagation path of the light beam. The optical element may change the propagation path of the light beam by means of reflection, refraction, diffraction, etc. of the light beam. Exemplarily, each optical element in the scanning module may project light to different directions by rotation, thereby scanning the space around the radar.

[0052] In one implementation, the radar may detect the distance from the radar to the detected object by measuring the time of light propagation between the radar and the detected object, i.e., the time-of-flight (TOF). Alternatively, the radar may also detect the distance from the radar to the detected object by other techniques, such as a ranging method based on phase shift measurement, or a ranging method based on frequency shift measurement, etc. The embodiments of the present invention are not limited thereto.

[0053] Noise usually exists in the point cloud data obtained by the radar. The embodiments of the present invention still divide the noise into two parts: the first part is the internal noise of the radar, and the second part is the external environmental noise. The internal noise of the radar is usually noise sporadically generated in the point cloud data due to intrinsic sensor defects or external environmental temperature / humidity, etc. Usually, such noise is very sparse in space. The external noise of the radar is usually non-security noise generated by specific environments or objects, such as fallen leaves, flying insects, waste paper scraps. This part of the noise is usually small and dense point clouds and is in a suspended state in space.

[0054] The noise filtering method of the embodiments of the present invention is constructed based on the idea of parallelization, and mainly performs diffusion filtering on potential noise through three stages, namely: 1) internal noise filtering of the radar, 2) scanning of small potential noise objects, and 3) noise diffusion regression.

[0055] Specifically, first project the point cloud data into a three-dimensional voxel grid. A three-dimensional voxel grid is a plurality of volume spaces divided in three-dimensional space, and each volume space is a voxel, which is short for Volume Pixel. Exemplarily, the three-dimensional voxel grid can be a three-dimensional voxel grid within the region of interest. The region of interest can be selected according to actual needs. For example, for a moving vehicle, the region of interest can be the area within a certain range around the vehicle. Objects outside this range do not need to be concerned temporarily, so the point cloud points within the region of interest can be retained, and the point cloud points outside the region of interest can be discarded. Then, the above-mentioned region of interest can be divided into a voxel matrix of size Nx * Ny * Nz, where Nx is the number of voxels divided along the x-axis, Ny is the number of voxels divided along the y-axis, and Nz is the number of voxels divided along the z-axis. Since the method of the embodiment of the present invention is implemented on an in-vehicle GPU, therefore, through multiple parallel computing units of the in-vehicle GPU, the mapping relationship between the point cloud data and the three-dimensional voxel grid can be established in parallel, thereby greatly improving the computing speed.

[0056] After projecting the point cloud data into the three-dimensional voxel grid, first filter the sparse noise generated inside the radar. Utilize the characteristic of sparse noise and distinguish the sparse noise from the objects with dense point clouds by calculating the number of data points in the grid. Specifically, in the three-dimensional voxel grid where the noise is located, the number of point cloud points is scarce; while in the three-dimensional voxel grid where the real object is located, since multiple laser beams scan the object, the point cloud points are dense and continuous, and the number of point clouds is large. Therefore, the number of point cloud points in each three-dimensional voxel grid can be determined, and the point cloud data in the three-dimensional voxel grid with the number of point cloud points less than the first threshold can be filtered out. By setting a reasonable threshold, the real object can be distinguished from the noise inside the radar, thereby achieving the purpose of filtering the noise inside the radar.

[0057] In this step, similarly, through multiple parallel computing units of the in-vehicle image processing unit, each voxel grid can be traversed in parallel to determine the number of point cloud points in each voxel grid, so as to filter out in parallel the point cloud data corresponding to the voxel grid with the number of point cloud points less than the first threshold, thereby improving the computing speed of this step.

[0058] After filtering out the internal noise of the radar, potential small objects in the remaining point cloud data are identified. The small objects include fallen leaves, flying insects, waste paper scraps, etc. Since the small objects are real objects, multiple dense point cloud beams of the radar will hit the object, so they cannot be filtered out by the internal noise filtering algorithm in the first stage. However, the volume of the small objects is small, and it is very likely that they will not pose a safety hazard to the vehicle driving. The purpose of filtering out the point cloud data generated by the small objects is mainly to reduce the subsequent computational workload. In the embodiment of the present invention, without using a clustering algorithm, the spatial volume ratio of each voxel grid is calculated, and the voxel grid corresponding to the potential small object is identified according to its spatial volume. For the convenience of description, it is called the potential small object grid.

[0059] Specifically, taking the voxel grid as the target, by performing three-dimensional convolution on each voxel grid, the occupancy rate of each voxel grid in a specific subspace is calculated to reflect the spatial volume of the grid cluster. By performing three-dimensional convolution on the voxel grid, in the adjacent voxel grids of each voxel grid, the valid voxel grids with point cloud data can be determined, and the proportion of the valid voxel grids in the adjacent voxel grids can be obtained, so as to determine the spatial volume ratio of the voxel grid. Since the small object corresponds to a voxel grid with a small volume in the subspace, the voxel grid with a spatial volume lower than the second threshold can be determined as the voxel grid corresponding to the potential small object.

[0060] In the embodiment of the present invention, the step of performing three-dimensional convolution on each voxel grid is executed in parallel by the in-vehicle image processing unit. Since the three-dimensional convolution of each voxel grid is completely independent, executing this step in parallel by the GPU can greatly reduce the operation time.

[0061] After identifying the potential small object grid corresponding to the potential small object, the steps of the third-stage noise diffusion regression are executed. Specifically, in the point cloud data of the potential small object, in fact, there is some valuable data, such as the edge of an obstacle, which is very helpful for accurately describing the size / direction of the obstacle. Therefore, the main purpose of the third stage is to distinguish real small object noise (such as fallen leaves, flying insects, waste paper scraps, etc.) from pseudo-small object noise (such as the edge of the road, branches, etc.). Pseudo-small object noise is usually connected to an object with a high spatial occupancy rate, while real small object noise is often completely suspended and in a floating state. Therefore, it is possible to distinguish the real small object grid from the pseudo-small object grid in the potential small object grid according to whether the potential small object grid is adjacent to the entity obstacle grid.

[0062] In some embodiments, convolution diffusion search can be performed on the voxel grid where all potential small objects are located, so as to gradually spread the information of the high spatial occupancy of the object to the adjacent pseudo-small object noise. Among them, the potential small object grid can be subjected to one or more convolution diffusions, so that the pseudo-small object grid in the potential small object grid is connected to the entity obstacle grid, or the entity obstacle grid can be subjected to one or more convolution diffusions, so that the entity obstacle grid is connected to the pseudo-small object grid in the potential small object grid. That is, it can spread from the pseudo-small object grid to the entity obstacle grid, or from the entity obstacle grid to the pseudo-small object grid. Exemplarily, the entity obstacle grid is a voxel grid that is not marked as a potential small object grid.

[0063] After multiple diffusion searches, the voxel grid where the pseudo-small object noise is located will be marked as an entity obstacle grid with high spatial occupancy. On the contrary, for the real small object noise, since it is not adjacent to the voxel grid of any entity obstacle with high spatial occupancy, the classification information of the voxel grid will not be updated after the diffusion search. Therefore, the pseudo-small object grid can be filtered out from the list of potential small object grids through the diffusion search, and only the real small object grid is retained.

[0064] Since the mapping from the point cloud data to the three-dimensional voxel grid is established in the first stage, after determining the real small object grid, the point cloud data corresponding to the real small object grid can be marked as noise data according to the mapping relationship between the point cloud data and the three-dimensional voxel grid. After that, the point cloud data corresponding to the real small object grid can also be filtered out. Thus, the real small object noise in the point cloud data can be filtered out without clustering.

[0065] In the embodiments of the present invention, the diffusion search can be performed on each potential small object grid in parallel through multiple parallel computing units of the vehicle-mounted image processing unit. Since the diffusion search of each voxel grid is completely independent, the operation time can be greatly reduced by parallel execution of this step by the GPU.

[0066] In the above algorithm, it is assumed that there are N point cloud points in the point cloud data, which can be projected into K voxel grids in the space of M three-dimensional voxel grids, where M >> N > K. Therefore, the traversal projection of each point cloud in the first stage will take O(N) operations, and the second and third stages are traversal operations based on the three-dimensional voxel grid, so each stage will take O(K) operations. Finally, the point cloud data in the noise grid will be deleted one by one using the operation relationship between the point cloud data and the three-dimensional voxel grid. In the parallelized algorithm, it will take O(N) operations. The whole process will require O(2N + 2KL) operations.

[0067] Since the above algorithm is completely independent when operating on different voxel grids, it can be highly parallelized by the GPU. In a GPU with J parallel computing units, the above operation process will take O(2N / J + 2KL / J) time to complete the filtering algorithm per unit time. Since the number of parallel units of the GPU is extremely large, that is, the value of J is large, the time consumption of the entire algorithm is very short, and noise filtering can be completed in a very short time.

[0068] In summary, the noise filtering method 100 for point cloud data according to the embodiments of the present invention integrates the advantages of high parallelism of the GPU. By decoupling serial dependencies, a highly parallelized noise filtering algorithm is designed, which can quickly filter the noise in the point cloud data at the input data level, reduce the redundant calculations of subsequent downstream modules, reduce sudden braking caused by non-safe obstacles, and increase the smoothness of vehicle driving of the autonomous driving system.

[0069] The embodiments of the present invention also provide a noise filtering device for point cloud data. Refer to Figure 2 , the noise filtering device 200 for point cloud data includes a memory 210 and a graphics processing unit 220. A computer program run by the graphics processing unit 220 is stored on the memory 210. When the computer program is run by the graphics processing unit 220, it executes the noise filtering method 100 for point cloud data.

[0070] Exemplarily, the memory 210 may include one or more computer program products, and the computer program products may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. The graphics processing unit 220 can run the program instructions stored in the memory 210 to implement the functions (implemented by the graphics processing unit) in the embodiments of the present invention described herein and / or other desired functions. Since the image processing unit 220 includes multiple parallel computing units, the operations of each step in the above method can be executed in parallel, filtering non-safe noise in a very short time, reducing the additional operations downstream, and increasing the smoothness of autonomous driving.

[0071] The embodiments of the present invention also provide an autonomous vehicle, which can be used to implement the noise filtering method 100 for point cloud data described above. The autonomous vehicle is an intelligent vehicle that does not require a driver to perform driving operations and can automatically complete vehicle driving tasks on behalf of the driver; the autonomous vehicle can also have an artificial driving function. Refer to Figure 3 , Figure 3 shows a schematic block diagram of an autonomous vehicle according to an embodiment of the present invention.

[0072] As shown Figure 3 in the figure, the autonomous vehicle includes a vehicle body 300, a radar 310, and a noise filtering device 320. The radar 310 is used to collect point cloud data and may include one or more radars. The noise filtering device 320 is connected to the radar 310 in a wired or wireless manner to receive the point cloud data collected by the radar 310. The noise filtering device 320 may be the noise filtering device for point cloud data as described above with reference to Figure 2 and is used to perform the noise filtering method 100 for point cloud data as described above to filter out the noise in the point cloud data, specifically including the discrete noise generated inside the radar and the noise of small objects generated by the external environment. It should be noted that the autonomous vehicle also includes other component structures, which are not limited in the embodiments of the present invention. The noise filtering method 100 for point cloud data performed by the noise filtering device 320 may refer to the above, and will not be elaborated here.

[0073] In addition, according to the embodiments of the present invention, a computer storage medium is also provided. Program instructions are stored on the computer storage medium and are used to execute the corresponding steps of the noise filtering method 100 for point cloud data in the embodiments of the present invention when the program instructions are run by a computer or a processor. The specific details can be seen above. The computer storage medium may include, for example, a memory card of a smart phone, a storage component of a tablet computer, a hard disk of a personal computer, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disc read-only memory (CD-ROM), a USB memory, or any combination of the above storage media. The computer-readable storage medium may be any combination of one or more computer-readable storage media.

[0074] In summary, the noise filtering method, device, autonomous vehicle, and storage medium for point cloud data in the embodiments of the present invention perform noise filtering in units of three-dimensional voxel grids. Different voxel grids are independent of each other, and the vehicle-mounted image processing unit can perform operations on different voxel grids in parallel, greatly improving the operation speed.

[0075] Although example embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above example embodiments are merely exemplary and are not intended to limit the scope of the present application thereto. Those of ordinary skill in the art can make various changes and modifications therein without departing from the scope and spirit of the present application. All such changes and modifications are intended to be included within the scope of the present application as claimed in the appended claims.

[0076] Those of ordinary skill in the art will appreciate that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0077] In several embodiments provided by this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed.

[0078] In the specification provided herein, a large number of specific details are set forth. However, it can be understood that the embodiments of this application can be practiced without these specific details. In some instances, well-known methods, structures, and technologies are not shown in detail so as not to obscure the understanding of this specification.

[0079] Similarly, it should be understood that, in order to streamline this application and assist in understanding one or more of the various inventive aspects, in the description of the exemplary embodiments of this application, the various features of this application are sometimes grouped together into a single embodiment, figure, or description thereof. However, the methods of this application should not be construed as reflecting the intention that the claimed application requires more features than are expressly recited in each claim. Rather, as reflected by the corresponding claims, the inventive point lies in being able to solve the corresponding technical problems with features fewer than all the features of a single disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into the detailed description, where each claim itself serves as a separate embodiment of this application.

[0080] Those skilled in the art can understand that, except for features being mutually exclusive, any combination can be adopted for all the features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all the processes or units of any method or device so disclosed. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) can be replaced by an alternative feature that provides the same, equivalent, or similar purpose.

[0081] In addition, those skilled in the art can understand that although some embodiments described herein include certain features included in other embodiments rather than other features, the combination of features of different embodiments means that it is within the scope of this application and forms different embodiments. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0082] Each component embodiment of this application can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) can be used in practice to implement some or all of the functions of some of the modules according to the embodiments of this application. This application can also be implemented as a device program (for example, a computer program and a computer program product) for executing part or all of the methods described herein. Such a program for implementing this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0083] It should be noted that the above embodiments illustrate rather than limit this application, and those skilled in the art can design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claims. This application can be implemented by means of hardware including several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices can be embodied by the same hardware item. The use of the words first, second, and third, etc. does not denote any order. These words can be interpreted as names.

[0084] The above is only the specific implementation manner of this application or the description of the specific implementation manner. The protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by this application should be covered by the protection scope of this application. The protection scope of this application shall be subject to the protection scope of the claims.

Claims

1. A method for filtering noise from point cloud data, characterized in that, the method is used for an in-vehicle image processing unit, and the method includes: acquiring point cloud data collected by a radar and mapping the point cloud data into a three-dimensional voxel grid; determining the number of point cloud points corresponding to each voxel grid in the three-dimensional voxel grid, and filtering the point cloud data corresponding to the voxel grids with the number of point cloud points less than a first threshold; calculating the spatial volume ratio of each voxel grid, and marking the voxel grids with a spatial volume less than a second threshold as potential small object grids; performing a diffusion search on the potential small object grids to determine pseudo small object grids in the potential small object grids that are connected to entity obstacle grids, marking potential small object grids other than the pseudo small object grids as real small object grids, and marking the point cloud data corresponding to the real small object grids as noise data according to the mapping relationship between the point cloud data and the three-dimensional voxel grid; filtering the point cloud data corresponding to the real small object grids.

2. The method for filtering noise from point cloud data according to claim 1, characterized in that, the performing a diffusion search on the potential small object grids to determine pseudo small object grids in the potential small object grids that are connected to entity obstacle grids includes: performing one or more convolutional diffusions on the potential small object grids to connect the pseudo small object grids in the potential small object grids to the entity obstacle grids.

3. The method for filtering noise from point cloud data according to claim 1, characterized in that, the performing a diffusion search on the potential small object grids to determine pseudo small object grids in the potential small object grids that are connected to entity obstacle grids includes: performing one or more convolutional diffusions on the entity obstacle grids to connect the entity obstacle grids to the pseudo small object grids in the potential small object grids.

4. The method for filtering noise from point cloud data according to claim 1, characterized in that, the calculating the spatial volume ratio of each voxel grid includes: performing three-dimensional convolution on each voxel grid to obtain the spatial volume ratio of the voxel grid.

5. The method for filtering noise from point cloud data according to claim 1, characterized in that, the mapping the point cloud data into a three-dimensional voxel grid includes: parallelly establishing the mapping relationship between the point cloud data and the three-dimensional voxel grid through multiple parallel operation units of the in-vehicle image processing unit.

6. The method for filtering noise from point cloud data according to claim 1, characterized in that, the determining the number of point cloud points corresponding to each voxel grid in the three-dimensional voxel grid, and filtering the point cloud data corresponding to the voxel grids with the number of point cloud points less than a first threshold includes: parallelly traversing each voxel grid through multiple parallel operation units of the in-vehicle image processing unit to filter the point cloud data corresponding to the voxel grids with the number of point cloud points less than a first threshold.

7. The method for filtering noise from point cloud data according to claim 1, characterized in that, Calculating the spatial volume rate of each of the voxel grids, and marking the voxel grids with a spatial volume smaller than a second threshold as potential small object grids, includes: Traversing each of the voxel grids in parallel through a plurality of parallel operation units of the vehicle-mounted image processing unit to mark the potential small object grids.

8. The noise filtering method for point cloud data according to claim 1, wherein, Performing diffusion search on the potential small object grids includes: Performing the diffusion search on each of the potential small object grids in parallel through a plurality of parallel operation units of the vehicle-mounted image processing unit.

9. A noise filtering device for point cloud data, wherein, The device includes a memory and a graphics processing unit. A computer program run by the graphics processing unit is stored on the memory. When the computer program is run by the graphics processing unit, it executes the noise filtering method for point cloud data according to any one of claims 1-8.

10. An autonomous vehicle, wherein, The autonomous vehicle includes: A vehicle body; A radar mounted on the vehicle body, the radar being used to collect point cloud data; The noise filtering device for point cloud data according to claim 9, connected to the radar, and used to execute the noise method for point cloud data according to any one of claims 1-8 to filter the point cloud data.

11. A storage medium, on which a computer program is stored. When the computer program runs, it executes the noise filtering method for point cloud data according to any one of claims 1-8.

Images