WO2024065685A1

WO2024065685A1 - Point cloud processing method and radar

Info

Publication number: WO2024065685A1
Application number: PCT/CN2022/123293
Authority: WO
Inventors: 马锋
Original assignee: 深圳市速腾聚创科技有限公司
Priority date: 2022-09-30
Filing date: 2022-09-30
Publication date: 2024-04-04

Abstract

A point cloud processing method and a radar. The point cloud processing method comprises: obtaining a point cloud, wherein points in the point cloud have reflectivities; in response to the point cloud, dividing the point cloud into first-type points and/or second-type points on the basis of the reflectivities of the points in the point cloud, wherein the reflectivity of the first-type points is higher than a first reflectivity threshold, and the reflectivity of the second-type points is lower than a second reflectivity threshold; and if the second-type points and the first-type points are connected and the second-type points meet a deletion condition, deleting the second-type points, wherein the deletion condition represents that the second-type points belong to swell noise.

Description

A method for processing point cloud and radar

Technical Field

The present application relates to the field of computer technology, and in particular to a method for processing point clouds and a radar.

Background technique

LiDAR is an active remote sensing device that uses lasers as the light source and photoelectric detection technology. It is an advanced detection method that combines laser technology with modern photoelectric detection technology. LiDAR is widely used in autonomous driving, logistics vehicles, robots, vehicle-road collaboration, public smart transportation and other fields.

There are abnormal noise points in the point cloud output by the LiDAR, which will affect the ranging accuracy of the LiDAR.

Summary of the invention

In order to solve or partially solve the problems existing in the related art, the present application provides a method and radar for processing point clouds, which can remove abnormal noise around high-reflective objects in the point clouds and improve the ranging accuracy of the laser radar.

In a first aspect, the present application provides a method for processing a point cloud, comprising: in response to obtaining a point cloud, dividing the point cloud into a first type of points and/or a second type of points based on the reflectivity of the points in the point cloud, wherein the reflectivity of the first type of points is higher than a first reflectivity threshold, and the reflectivity of the second type of points is lower than a second reflectivity threshold; obtaining at least one first connected domain corresponding to the first type of points by projection, and/or obtaining at least one second connected domain corresponding to the second type of points by projection; if the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, deleting the points in the point cloud corresponding to the second connected domain, the deletion condition characterizing that the points in the second connected domain belong to expansion noise.

The second aspect of the present application provides a device for processing point clouds, including: a point cloud classification module, a connected domain acquisition module and a noise removal module. The point cloud classification module is used to respond to the acquisition of the point cloud and divide the point cloud into a first type of points and/or a second type of points based on the reflectivity of the points in the point cloud, wherein the reflectivity of the first type of points is higher than a first reflectivity threshold, and the reflectivity of the second type of points is lower than a second reflectivity threshold; the connected domain acquisition module is used to obtain at least one first connected domain corresponding to the first type of points by projection, and/or, to obtain at least one second connected domain corresponding to the second type of points by projection; the noise removal module is used to delete the points in the point cloud corresponding to the second connected domain if the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, wherein the deletion condition indicates that the points in the second connected domain belong to expansion noise.

A third aspect of the present application provides a board comprising the device for processing point clouds as described above.

A fourth aspect of the present application provides a radar, comprising the device for processing point clouds as described above.

A fifth aspect of the present application provides an electronic device, comprising: a processor; and a memory, on which executable code is stored, and when the executable code is executed by the processor, the processor executes the method as described above.

A sixth aspect of the present application provides a computer-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor executes the above method.

A seventh aspect of the present application provides a computer program product, comprising an executable code, which implements the above method when the executable code is executed.

The technical solution provided by this application may have the following beneficial effects:

In some embodiments of the present application, the point cloud is divided into high-reflectivity points with a reflectivity higher than a first reflectivity threshold and low-reflectivity points with a reflectivity lower than a second reflectivity threshold according to the reflectivity of the points in the point cloud. When a low-reflectivity point is adjacent to a high-reflectivity point and the low-reflectivity point meets the deletion condition, the low-reflectivity point is most likely an expansion-type noise spot generated by a high-reflectivity object. By deleting the low-reflectivity point, the expansion-type noise can be effectively reduced, and the ranging accuracy of the laser radar can be improved.

In addition, in some embodiments of the present application, in order to facilitate the determination of whether low-reflectivity points are connected to high-reflectivity points, a first connected domain corresponding to the first type of points and a second connected domain corresponding to the second type of points are obtained by projection, so that whether the first type of points and the second type of points are connected can be determined by graphic processing.

In addition, in certain embodiments of the present application, when the second connected domain and the first connected domain are separated from each other in at least one of the first projection binary map, the second projection binary map, or the third projection binary map, it indicates that the corresponding second-type points and first-type points are separated from each other, and are not expansion-type noise, and need to be retained to avoid accidental deletion of low-reflectivity points corresponding to obstacles, etc.

In addition, in some embodiments of the present application, the deletion condition includes: the distance between all points corresponding to the second connected domain and the first connected domain is relatively close, and/or the ratio between the area of the second connected domain and the area of the first connected domain is less than or equal to a ratio threshold. Based on this deletion condition, it can be effectively determined whether the point cloud corresponding to the second connected domain is expansion noise.

In addition, in certain embodiments of the present application, in order to further improve driving safety, different trade-off criteria are used to select point clouds within a preset height range and point clouds within a non-preset height range, so as to reduce the number of point clouds with low risk while meeting the sensing capabilities of the lidar.

In addition, in certain embodiments of the present application, in order to further reduce the number of point clouds belonging to low risk, operations such as deleting burrs, deleting monotonically changing reflectivity, deleting highly consistent reflectivity, and deleting second connected domains with a large edge area ratio may be performed on the point clouds within the retained second connected domain, thereby effectively further reducing the number of point clouds belonging to low risk.

In addition, in some embodiments of the present application, the second reflectivity threshold is determined based on the average reflectivity of the first type of points, so that the second reflectivity threshold is more in line with the threshold standard of expansion-type noise, thereby improving the accuracy of expansion-type noise identification.

It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application.

BRIEF DESCRIPTION OF THE DRAWINGS

The above and other objects, features and advantages of the present application will become more apparent by describing in more detail exemplary embodiments of the present application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the present application.

FIG1 is a schematic diagram of a method for processing point clouds and an application scenario of radar according to an embodiment of the present application;

FIG2A is a schematic diagram of a laser radar system according to an embodiment of the present application;

FIG2B is a schematic diagram of a laser radar data processing system according to an embodiment of the present application;

FIG2C is a schematic diagram of a point cloud processing system according to an embodiment of the present application;

FIG3 is a flow chart of a method for processing a point cloud according to an embodiment of the present application;

FIG4 is a schematic diagram of a high-reflection expansion process shown in an embodiment of the present application;

FIG5 is a schematic diagram showing the connection between the first connected domain and the second connected domain after high-reflection expansion according to an embodiment of the present application;

FIG6 is a flow chart of a method for processing a point cloud shown in another embodiment of the present application;

FIG7 is a schematic diagram of a point cloud before high-reflection expansion noise removal processing shown in an embodiment of the present application;

FIG8 is a schematic diagram of a point cloud after high-reflection expansion noise removal processing shown in FIG7 ;

FIG9 is a schematic diagram of a point cloud before high-reflection expansion noise removal according to an embodiment of the present application;

FIG10 is a schematic diagram of the point cloud after the high reflection expansion noise removal process shown in FIG9 ;

FIG11 is a schematic diagram of a point cloud before high-reflection expansion noise removal processing shown in an embodiment of the present application;

FIG12 is a schematic diagram of a point cloud after high-reflection expansion noise removal processing shown in FIG11;

FIG13 is a schematic diagram of a point cloud before high-reflection expansion noise removal processing shown in an embodiment of the present application;

FIG14 is a schematic diagram of a point cloud after high-reflection expansion noise removal processing shown in FIG13;

FIG15 is a schematic diagram of a point cloud before high-reflection expansion noise removal processing shown in an embodiment of the present application;

FIG16 is a schematic diagram of a point cloud after high-reflection expansion noise removal processing shown in FIG15 ;

FIG17 is a schematic diagram of the structure of an apparatus for processing point clouds according to an embodiment of the present application;

FIG18 is a schematic diagram of the structure of a radar according to an embodiment of the present application;

FIG. 19 is a schematic diagram of the structure of an electronic device according to an embodiment of the present application.

Detailed ways

The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "an" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and/or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

It should be understood that although the terms "first", "second", "third", etc. may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

In order to facilitate the understanding of the present application, some concepts involved in the present application are first explained.

Vehicle-mounted radar: The detection distance is 200 to 500 meters, and the identifiable physical properties can include distance and reflectivity. It can be used for small machines such as vehicles and robots. Vehicle-mounted radars include vehicle-mounted laser radars, vehicle-mounted millimeter-wave radars, and vehicle-mounted ultrasonic radars.

LiDAR: An active remote sensing device that uses lasers as the light source and photoelectric detection technology. It is an advanced detection method that combines laser technology with modern photoelectric detection technology. It consists of a transmitting system, a receiving system, a scanning control system, a data processing system, and other parts. Its working principle is to transmit a detection signal to the target, and then process the received echo signal to obtain information such as the distance, size, speed, and reflectivity of the target. Its advantages are high resolution, high sensitivity, strong anti-interference ability, and no influence from dark conditions. It is widely used in autonomous driving, logistics vehicles, robots, vehicle-road collaboration, public smart transportation, and other fields.

Vehicle-mounted laser radar: By emitting outgoing light (such as a laser beam) with a wavelength of about 900nm, the outgoing light will be reflected by the obstacle after encountering it. The processing unit calculates the distance between the obstacle and the vehicle-mounted laser radar based on the time difference between the reflected light and the outgoing light. In addition, the processing unit can also estimate the reflectivity of the target based on the cross-section of the reflected light. Vehicle-mounted laser radar has a small size and a high degree of integration.

In the autonomous driving scenario, the system architecture applicable to the autonomous driving scenario may include mobile devices, networks, and the cloud. Mobile devices include, but are not limited to: cars, ships, robots, aircraft, etc. Mobile devices may be provided with electronic devices such as sensors to obtain information about obstacles in the surrounding environment of the mobile device, etc. Electronic devices may include: radars, image sensors, etc.

LiDAR is an active remote sensing device that uses photoelectric detection technology. It is an advanced detection method that combines laser technology with modern photoelectric detection technology. It consists of a transmitting system, a receiving system, a scanning control system, a data processing system, and other parts. Its working principle is to transmit a detection signal to the target, and then process the received echo signal to obtain information such as the distance, size, speed, and reflectivity of the target object. Its advantages are high resolution, high sensitivity, strong anti-interference ability, and no influence from dark conditions. It is widely used in autonomous driving, logistics vehicles, robots, vehicle-road collaboration, public smart transportation and other fields.

The embodiment of the present application provides a method and radar for processing point clouds, and obtains high reflectivity points and low reflectivity points from the point clouds. If the low reflectivity point is connected to the high reflectivity point, and the low reflectivity point satisfies the deletion condition representing the expansion noise, the low reflectivity point can be deleted, and the high reflectivity expansion noise can be effectively identified and deleted, thereby improving the accuracy of the distance determined based on the point cloud.

The technical solution of the embodiments of the present application is described in detail below with reference to the accompanying drawings.

FIG1 is a schematic diagram of a method for processing point clouds and an application scenario of radar according to an embodiment of the present application.

FIG1 shows the hardware structure of a vehicle 10 that supports assisted driving or automatic driving functions. For example, at least one Light Detection and Ranging (LIDAR) 11 is mounted on the roof and/or the side of the vehicle body of the vehicle 10. The detection area of the LIDAR 11 may be fixed, such as a certain LIDAR 11 may be used only for detection of a preset area. The detection area of the LIDAR 11 may be adjustable, such as the LIDAR on the vehicle body may scan multiple detection areas by adjusting the posture, etc., or may scan multiple detection areas by adjusting the field of view of the LIDAR itself. Specifically, the vehicle 10 may be equipped with five LIDARs 11: the top of the vehicle, the front side of the vehicle, the rear side of the vehicle, the left side of the vehicle, and the right side of the vehicle. Through multiple LIDARs 11, the outline of objects existing in the area around the vehicle and the distance to the object can be detected.

In addition, a camera may be mounted on the vehicle 10. The camera can capture the environment in front of the vehicle at a predetermined viewing angle. For example, the camera may be a monocular camera, a multi-camera camera, or the like.

In addition, the vehicle 10 may be equipped with a plurality of millimeter wave radars in a manner surrounding the vehicle 10. For example, the vehicle 10 may be equipped with four millimeter wave radars, with the left side in front of the vehicle, the right side in front of the vehicle, the left side behind the vehicle, and the right side behind the vehicle as the detection range. The millimeter wave radars can detect the distance of an object existing in each detection area, and can detect the relative speed between the object and the vehicle 10.

Furthermore, the vehicle 10 may also be equipped with a positioning device 12, such as a Beidou positioning device, a global positioning system (GPS), etc. The positioning device 12 can determine the current position of the vehicle 10.

In addition, the vehicle 10 may also be equipped with an electronic control unit (ECU). The detection signal of at least one of the above-mentioned LIDAR 11, millimeter wave radar and positioning device 12 is sent to the ECU. The ECU can detect and identify obstacles (such as roadblocks, moving objects, trees, adjacent vehicles, etc. around the vehicle 10) based on these signals. In addition, the ECU can be physically divided into multiple units according to function, and this application will collectively refer to them as ECU.

It should be noted that although the movable device is described as a car, such description is not restrictive and is applicable to a variety of movable devices, such as land robots, water robots, etc.

The method for processing point clouds and radar of the embodiments of the present application can be applied to any one or more electronic devices that require the use of a clock, such as LIDAR 11, millimeter wave radar, positioning device 12, ECU or communication system as shown in Figure 1.

In the above-mentioned application scenarios such as assisted driving, autonomous driving, and smart transportation, quickly and accurately perceiving the surrounding environment of movable equipment is a key point.

Here, the vehicle's autonomous driving scenario is used as an example for illustrative explanation. According to the vehicle position information, obstacle information, and road information perceived by the sensor system, road signal control is coordinated to improve the quality and efficiency of road management. Specifically, the corresponding autonomous driving vehicle decision can be determined based on the information perceived by the sensor system, and the safe distance between autonomous driving vehicles can be adjusted, so that the vehicle can drive safely and reliably on the road.

FIG. 2A is a schematic diagram of a laser radar system according to an embodiment of the present application.

Referring to FIG. 2A , taking a vehicle-mounted laser radar as an example, the laser radar may include a transmitting system, a receiving system, a data processing system, and a point cloud output system.

The transmitting system includes a signal generator, a laser transmitter, and a lens, etc., which are used to transmit laser signals. The signal generator generates a signal to trigger the laser transmitter to emit a laser with a wavelength of 905nm or 1550nm. After the laser is emitted, it passes through the lens and irradiates the target object (target object). The target object can be regarded as a Lambertian body, that is, a diffusely reflecting object.

The laser receiving system may include a filter, an attenuation plate, a photosensitive unit, and a temperature control module, a boost circuit, a quenching circuit, and a signal extraction module used in conjunction with the photosensitive unit, etc., for receiving the laser reflection signal and converting the laser reflection signal into an electrical signal. For example, the echo signal reflected by the obstacle is narrow-band filtered and then passes through the attenuation plate (whose transmittance affects the detection performance of the laser radar). After passing through the attenuation plate, the echo signal is irradiated onto the photosensitive element (such as a single-photon avalanche photodiode SPAD). At this time, when the photosensitive element works in the Geiger mode, a single photon can trigger an avalanche, and the avalanche current increases dramatically. Since continuous large currents will cause the avalanche photodiode (APD) device to heat up, the device may be damaged if the time is too long, and when the APD works in the Geiger mode, the self-sustaining avalanche process cannot end automatically, resulting in the system being unable to perform the next round of receiving process. Therefore, quenching is required to terminate the avalanche and increase the APD receiving rate. The quenching circuit mainly includes passive quenching, active quenching, and gated quenching. Passive quenching mainly uses resistor voltage division to reduce the voltage and exit the avalanche state. Active quenching is mainly to actively reduce the reverse bias voltage through the feedback circuit, thereby exiting the avalanche. Gated quenching keeps the frequencies of the emitter and the detector the same, while ensuring that the detector is in Geiger mode when photons arrive, and in the cutoff state when photons should not arrive. This quenching method reduces the dark count. After quenching, the SPAD exits the avalanche state, and after being boosted by the boost circuit, it works in Geiger mode again and can receive again. The period of boosting is called dead time, and the length of the dead time has a great impact on the performance of the lidar.

The data processing system processes and analyzes the collected electrical signals to generate point clouds. The point cloud output system is used to output point clouds for obstacle recognition, lane line recognition, etc.

FIG. 2B is a schematic diagram of a laser radar data processing system according to an embodiment of the present application.

2B , the data processing system may specifically include an echo signal filtering and processing system, an echo signal detection system, a distance, reflectivity, and angle information calculation system, a point cloud data generation system, and a point cloud processing system.

The echo signal filtering processing system can filter the point cloud. For example, the point cloud can be filtered by using related technical means such as echo intensity threshold filtering and outlier correction algorithm.

Echo data processing is an important part of lidar detection. The detection of echo signals is the source of information such as distance and speed.

In the process of simulating radar target echoes to calculate distance, reflectivity, angle information, etc., a high-speed analog-to-digital converter (ADC) can be used to collect radar pulse signals, and the collected signals are transmitted to a field-programmable gate array (FPGA) for detection processing to determine the rising edge time of the detection signal. Based on the rising edge time, the signal is delayed and then modulated to simulate the target echo signal, i.e., the target simulated echo signal. After the target simulated echo signal is received by the radar, the radar performance (such as radar ranging accuracy, etc.) can be tested.

The point information can then be processed by point cloud interpolation (such as upsampling, upsampling) to generate a point cloud. For example, ArcGIS can be used to create a las dataset, then convert the las dataset to a raster, and then the global mapper raster to a DEM to export the generated point cloud.

The method for processing point cloud provided in the present application can be applied to a point cloud processing system. FIG2C is a schematic diagram of a point cloud processing system shown in an embodiment of the present application.

Referring to FIG. 2C , the point cloud processing system includes but is not limited to: a point cloud denoising processing system, a point cloud rain and snow detection system, a point cloud strong light detection system, a high reflection expansion processing system, a point cloud filtering processing system, a point cloud interpolation processing system, a point cloud enhancement processing system, and a point cloud correction processing system.

For example, the point cloud denoising system can be used to remove some noise points that will not affect driving safety, such as removing outliers whose height is higher than the height of the car.

For example, in rainy and snowy weather, the LiDAR will detect raindrops and snowflakes in the air during scanning, thus forming a large number of discrete point clouds that affect the normal judgment of the perception system and cause the vehicle to be unable to drive normally. These point clouds floating in the air due to rainy and snowy weather can be called "noise points". Using the point cloud rain and snow detection system to reduce the noise of the point cloud and filter out these noise points is very necessary for the safe driving of the vehicle.

For example, the point cloud enhancement processing system can enhance the point cloud based on the basic data processing and enhancement methods of point cloud in deep learning. Among them, the point cloud enhancement processing includes point cloud normalization, random shuffling, random translation, random rotation, random scaling and random discarding, etc., continuous summary and update, etc.

For example, the point cloud correction processing system can correct the point cloud distortion caused by the movement of the laser radar. Assuming that the frequency of the vehicle-mounted laser radar is 10HZ, the time for a complete point cloud frame is 0.1s. Within this 0.1s, if the vehicle is moving at high speed, the coordinate system of the vehicle-mounted laser radar is always moving, which will cause distortion: the output point cloud is not in the initial vehicle-mounted laser radar coordinates. The coordinate system changes with the movement of the car, resulting in point cloud distortion. The distorted point cloud can be corrected through point cloud correction, which helps to improve the accuracy of the determined obstacle location information and improve driving safety.

Flash LiDAR is a typical radar in the field of LiDAR. Since there is no scanning process, the mechanical structure of Flash LiDAR is extremely simple, the difficulty of installation is very low, the cost advantage is obvious, and the difficulty of implementation in the entire LiDAR ecosystem is relatively low. However, due to the floodlight characteristics of Flash LiDAR, it is easy to produce expansion noise spots that scanning radars do not produce. For example, the noise formed by hitting high-reflectivity objects is particularly obvious. The echo of high-reflectivity objects will affect the floodlight echo signals of surrounding points, forming expansion noise around high reflectivity, which will affect point cloud clustering and object recognition. It is a major failure mode of Flash LiDAR.

The related technology is not convenient for determining whether the target point cloud corresponds to a real object or the expansion noise caused by a high-reflectivity object. If the high-reflectivity expansion is not processed, the point cloud formed by the expansion noise will lead to the misjudgment of the existence of a false target object. In addition, if low-reflectivity objects near high-reflectivity objects are over-removed (deleted by mistake), the real objects near the high-reflectivity objects will also be misjudged, which will affect the accuracy of the judgment results in application scenarios such as autonomous driving.

In certain embodiments of the present application, based on the distance relationship between high-reflectivity points and low-reflectivity points in the point cloud (such as whether they are connected) and the deletion conditions determined in combination with the characteristics of the expansion noise, at least some of the low-reflectivity points are deleted. This effectively reduces the amount of expansion noise while ensuring the accuracy of detection of real objects near high reflections, thereby improving the accuracy of detection results in application scenarios such as autonomous driving.

FIG. 3 is a flow chart of a method for processing a point cloud according to an embodiment of the present application.

3 , the method for processing a point cloud includes operations S310 to S330 .

In operation S310 , a point cloud is obtained, and points in the point cloud have reflectivity.

In this embodiment, please refer to Figures 2A and 2B. After the transmitting system transmits the laser signal, the receiving system receives the echo signal. After the data processing system processes the echo signal filtering, echo signal detection, distance, reflectivity, angle information calculation, etc., a point cloud can be obtained, and the point cloud has information such as reflectivity. Among them, the laser radar can receive the reflected signal, and the received reflected light needs to reach a certain light intensity. However, the reflectivity of the laser for targets at the working distance is uncertain. The reflectivity of cars, people, and roads to the laser is different. Therefore, this results in the reflectivity of multiple points in a frame of point cloud may be different, and the reflectivity of each point in the point cloud may be different.

In operation S320, in response to the point cloud, the point cloud is divided into first-category points and/or second-category points based on reflectivity of points in the point cloud, wherein the reflectivity of the first-category points is higher than a first reflectivity threshold and the reflectivity of the second-category points is lower than a second reflectivity threshold.

In this embodiment, high reflectivity points and low reflectivity points are classified so that it is possible to determine whether the current low reflectivity point is an expansion noise point based on the spatial position relationship between the high reflectivity points and the low reflectivity points and the characteristics of the low reflectivity points themselves.

In order to facilitate subsequent point cloud processing, such as determining whether the second type of points are expansion noise, the first type of points and the second type of points can be labeled separately. For example, the labels can be high reflectivity, low reflectivity, etc.

For example, the first reflectivity threshold and the second reflectivity threshold may be fixed values, such as those determined based on expert experience or historical statistical results. The first reflectivity threshold and the second reflectivity threshold may be dynamic values, such as those determined based on reflectivity statistical results of a previous cycle. The first reflectivity threshold may be higher than the second reflectivity threshold. For example, the first type of point is a high reflectivity point, and the second type of point is a low reflectivity point.

In some embodiments, the second reflectivity threshold is determined based on the reflectivity average value of the first type of points. For example, the low reflectivity threshold can be determined based on the reflectivity statistics (such as the average reflectivity) of at least some of the high reflectivity points in the current frame or at least one historical frame. Specifically, a typical value of the threshold for low reflectivity calibration is: subtracting at least one of 100, 110, 120, 130 or 140 from the high reflectivity average value.

In operation S330, if the second type point is connected to the first type point and the second type point satisfies a deletion condition, the second type point is deleted, and the deletion condition indicates that the second type point belongs to expansion noise.

In this embodiment, if the second type of point is isolated from the first type of point, it can be determined that the second type of point is not the expansion noise of the high reflectivity point corresponding to the high reflectivity object. If the second type of point is connected to the first type of point, it is necessary to further determine whether the current second type of point is the expansion type noise based on the deletion condition. For example, there are usually only low reflectivity points in the expansion type noise, and the deletion condition may include requirements for low reflectivity. For example, the reflectivity of the points in the expansion noise is usually relatively uniform, and the deletion condition may include requirements for reflectivity uniformity. For example, the distance between the expansion noise point and the high reflectivity point is usually close, and the deletion condition may include requirements for the distance between the point and the high reflectivity point. For example, the number of points in the expansion noise is usually small, and the deletion condition may include requirements for the number of points. In this way, the recognition accuracy of the expansion type noise in the second type of point can be further improved based on the deletion condition.

Among them, the second type of points are connected to the first type of points. For example, for a point in the second type of points, it can be determined whether the preset distance radius of the point includes the first type of points. If the preset radius of the second type of points includes the first type of points, it is determined that the second type of points are connected to the first type of points.

Among them, it can be understood that before determining whether the second type of points and the first type of points are connected, the method also includes: determining the number of first type of points and second type of points that meet the conditions, and when the number of first type of points and the number of second type of points meet the preset conditions, determining whether the second type of points and the first type of points are connected.

The point cloud processing method provided in this embodiment can effectively eliminate high anti-expansion noise in the second type of points based on whether the first type of points and the second type of points are connected and based on deletion conditions, thereby improving the accuracy of laser radar recognition.

In some embodiments, due to the huge number of points in a frame of point cloud, calculating the distance between two points separately requires a large amount of computing resources, takes a long time to calculate, and has a slow response speed. In order to solve the above problem, the first type of points and/or the second type of points can be processed as connected domains in the image by image processing to reduce the consumption of computing resources and improve the response speed.

Specifically, if the second type of point is connected to the first type of point, and the second type of point meets the deletion condition, deleting the second type of point may include the following operations.

First, at least one first connected domain corresponding to the first type of points is obtained by projection, and/or at least one second connected domain corresponding to the second type of points is obtained by projection. The connected domain of an image refers to the area composed of pixels with the same pixel value and adjacent positions in the image. Connected domain analysis refers to finding independent connected domains in the image and marking them.

Then, if the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, the points in the point cloud corresponding to the second connected domain are deleted, and the deletion condition indicates that the points corresponding to the second connected domain belong to expansion noise.

It can be understood that the second connected domain is connected to the first connected domain, which means whether any point in the first connected domain is included in an area with any point in the second connected domain as the center and a preset distance as the radius. If included, it means that the first connected domain and the second connected domain are connected.

Among them, optionally, when the second connected domain is a regular shape, the center point in the second connected domain can be used as the center of the circle to determine whether the area with a preset distance as the radius includes any point in the first connected domain. If included, it means that the first connected domain and the second connected domain are connected.

For example, the first connected domain is a connected domain in the projected binary graph corresponding to the first type of points. For example, the second connected domain is a connected domain in the projected binary graph corresponding to the second type of points.

Take the XYZ coordinate system with the spatial position of the vehicle-mounted laser radar as the origin of the coordinate system as an example. In the XYZ three-dimensional coordinate system, the projected binary image includes: the first projected binary image in the XY plane, the second projected binary image in the XZ plane, or the third projected binary image in the YZ plane. In this way, a frame of point cloud can be projected onto the XY plane, XZ plane, and YZ plane respectively. For a frame of point cloud, only the three projected binary images need to be analyzed, and there is no need to calculate each point in a frame of point cloud separately.

As mentioned above, as long as the first-class points and the second-class points are separated from each other, it can be determined that the second-class points are not the expansion noise points of the first-class points. Correspondingly, if the projections of the first-class points and the second-class points of a frame of point cloud on any of the three planes (XY plane, XZ plane, YZ plane) are separated from each other, it can be shown that the first-class points and the second-class points are separated from each other in space, and the second-class points are not the expansion noise of the first-class points. That is, the second-class points corresponding to the projection are not expansion noise and need to be retained.

Specifically, the method may further include the following operation: if the second connected domain is separated from the first connected domain in at least one of the first projected binary image, the second projected binary image or the third projected binary image, then retaining the points in the second connected domain.

In some embodiments, due to the irregular shape of the target object, the quality of the point cloud data collected by the lidar, etc., after the point cloud is projected, there are missing parts inside and on the edges of the connected domain, which may lead to the misjudgment that the second connected domain is separated from the first connected domain.

In order to improve the above problem, morphological processing can be performed on the connected domain. Specifically, the above method can also include the following operation: performing morphological processing on the first connected domain, and the morphological processing includes: at least one of morphological closing operation or morphological expansion processing. Among them, the morphological closing operation or morphological expansion processing can adopt relevant technologies, which are not limited here.

Correspondingly, if the second connected domain is connected to the first connected domain and the second connected domain satisfies the deletion condition, deleting the points in the point cloud corresponding to the second connected domain may include the following operations: if the second connected domain is connected to the first connected domain after morphological processing and the second connected domain satisfies the deletion condition, deleting the points in the point cloud corresponding to the second connected domain.

FIG. 4 is a schematic diagram of a high-reflection expansion process shown in an embodiment of the present application.

Referring to FIG4 , the left figure is a schematic diagram of a connected domain that has not been processed by high-reflection expansion. The three points at the top of the left figure are separated from each other, and image processing needs to be performed according to three connected domains. The right figure is a schematic diagram of a connected domain that has been processed by high-reflection expansion. The three points at the top of the right figure form a connected domain as a whole after high-reflection expansion processing, and image processing is performed according to a single connected domain.

FIG. 5 is a schematic diagram showing the connection between the first connected domain and the second connected domain after high-reflection expansion according to an embodiment of the present application.

Referring to FIG5 , the left figure shows a black high-reflection connected domain and a gray low-reflection connected domain. It can be seen that the high-reflection connected domain and the low-reflection connected domain are adjacent and can be considered to be connected or unconnected. The right figure shows the black high-reflection connected domain after expansion. There is a connecting edge between the black high-reflection connected domain and the gray low-reflection connected domain after expansion, that is, they are connected. If the low-reflection connected domain in the right figure meets the deletion condition, it can be determined as expansion noise.

In some embodiments, the deletion condition includes at least one of the following. For example, the distance between the point in the point cloud corresponding to the second connected domain and the first connected domain is less than or equal to the distance threshold. Among them, the distance between the two is less than or equal to the distance threshold, which indicates that the two are adjacent. On the other hand, it indicates that there are no points that are too far apart. For example, if there is a long strip-shaped point cloud far away from the first connected domain in the second connected domain, it is not expansion noise. The distance threshold can be determined based on expert experience, usage effect, etc.

For example, the ratio of the area of the second connected domain to the area of the first connected domain is less than or equal to the ratio threshold. Since the area of the connected domain corresponding to the expansion noise is usually small, if the area of the second connected domain is large, it is not the expansion noise. The ratio threshold can be determined based on expert experience, usage effect, etc. For example, the ratio threshold can be 1/6, 1/5, 1/4, 1/3, etc.

In some embodiments, since the point cloud is projected, a connected domain may correspond to one or more objects, and the connected domain may be split to improve the richness of the information expressed by the connected domain.

Specifically, the above method may also include the following operations: splitting the first connected domain to obtain multiple sub-connected domains, each of which corresponds to a different object. For example, if there are pedestrians and vehicles in the axial direction of the laser radar, the shape of the first connected domain is the superposition of the outer contours of the pedestrians and the vehicles. If the first connected domain is split into at least two connected domains, it is helpful to improve the accuracy of the determined expansion noise and the accuracy of the distance of the target object.

In a specific embodiment, the distance, reflectivity and elevation angle information of the point cloud uploaded by the laser radar device is processed. First, the point cloud distance, point cloud reflectivity and elevation angle of each point on the entire image are obtained, and each point is traversed.

First, the reflectivity value of each point is determined. Points above the high reflectivity threshold HighRefThd (a typical high reflectivity threshold for 8-bit processing is 200) are labeled as high reflectivity points. Points below the low reflectivity threshold LowRefThd (a typical low reflectivity threshold for 8-bit processing is 50) are labeled as low reflectivity points. For example, the overall high reflectivity mark and low reflectivity mark are similar to a binary process, where values above a certain threshold are set to 1, values below a certain threshold are set to 0, and the remaining values are discarded.

It should be noted that the above operation can also be improved in the following way: determine the reflectivity value of each point, and label the points higher than the high reflectivity threshold HighRefThd (a typical high reflectivity threshold of 8-bit processing is 200) as high reflectivity points; calculate the average reflectivity based on the reflectivity of all high reflectivity points, and then determine the threshold for low reflectivity calibration based on the average reflectivity (a typical value is to subtract 120 from the high reflectivity average value), and then label the points lower than the low reflectivity threshold LowRefThd.

Then, multiple highly reflective labels that are at the same distance and very close to each other are merged into one and considered as one object. It should be noted that there may be multiple highly reflective images connected on the point cloud, which need to be split into multiple labels.

Next, a morphological closing operation is performed on all high-reflection connected domains, which can remove the hollow part of the high-reflection connected domain, make the edge of the high-reflectivity object smoother, and facilitate the judgment of the contour of the high-reflectivity object. Then, a morphological expansion process is performed on the distance neighborhood of each high-reflection connected domain. A typical expansion method is: obtain the distance average, reflectivity average and area size of the current high-reflectivity area, perform convolution calculation with the current high-reflection connected domain, and then perform expansion and morphological processing, so that the high-reflection connected domain can be expanded and connected with the low-reflection connected domain. The low-reflection connected domain without intersection is regarded as non-expansion noise and is retained.

Then, for a certain high-reflection connected domain obtained in the previous operation, all its distance connected domains are traversed and judged based on the deletion conditions. For example, the deletion conditions may include at least one of the following conditions: Condition 1, whether this area is a low-reflection connected domain, and whether the reflectivity satisfies the low-reflection threshold. Condition 2, whether the average distance between each point in its connected domain and the high-reflection connected domain is close. Condition 3, whether the ratio of the area of the low-reflection connected domain to the area of the high-reflection connected domain is less than a certain ratio, and a typical value is that the area of the low-reflection connected domain accounts for less than 1/5 of the high-reflection connected domain.

This embodiment can accurately determine whether a low reflection point connected to a high reflection point is expansion noise on the basis of consuming less computing resources and having a higher response speed.

In some embodiments, in order to further improve driving safety and improve the accuracy of obstacle identification based on point cloud, the above method may also include the following operations: selecting the second type of points according to the selection criteria based on the height information of the second type of points and the preset height range.

Specifically, the preset height range includes at least two height ranges, and the rejection criteria corresponding to each of the at least two height ranges are different. For example, for a vehicle-mounted laser radar, the height of obstacles that may affect the passage of vehicles is within a specific height range, such as obstacles within a height range from the ground to 2 meters above the ground may interfere with the vehicle. However, within the height range of 2 meters above the ground, obstacles do not have much impact on the passage of vehicles. Therefore, point clouds can be rejected according to the height range using corresponding different rejection criteria. For example, point clouds within the height range from the ground to 2 meters above the ground are retained as much as possible, and point clouds within the height range above 2 meters above the ground can be retained less. It should be noted that the above two height ranges are only exemplary descriptions and cannot be understood as limitations on the present application. In addition, the height range can also be three, four or more, and accordingly, there can also be three, four or more rejection criteria.

In some embodiments, after selecting the second type of points according to the selection criteria based on the height information of the second type of points and the preset height range, the above method may further include the following operations: first, for the retained second connected domain, extract the burr part of the contour of the second connected domain. Then, remove the burr part in the second connected domain. After research and analysis, the burr part of the low reflectivity connected domain is mainly noise, and the burr part can be removed.

In some embodiments, since the high-reflection expansion noise is mainly located around the high-reflection connected domain, and its shape changes with the edge shape of the high-reflection connected domain, it can appear as a long strip, that is, the side length area is relatively large. Therefore, the above method can also include the following operations.

After the second type of points are selected according to the selection criteria based on the height information of the second type of points and the preset height range, for the retained second connected domain, the edge of the second connected domain is determined based on the reflectivity gradient change information of the second connected domain.

Then, if the ratio of the area occupied by the edge of the second connected domain to the area of the second connected domain is smaller than a preset area ratio, the second connected domain is deleted.

In some embodiments, the reflections of different parts of the real object are different and variable, and accordingly, the reflectivity of the point cloud of different parts of the real object is different and variable. The reflectivity of the expansion noise may be uniform or monotonically changing. Therefore, it can also be judged whether the second type of point is expansion noise based on this.

Specifically, the above method may also include the following operations: after selecting the second type of points according to the selection criteria based on the height information of the second type of points and the preset height range, for the retained second connected domain, if the reflectivity consistency of the second connected domain is higher than the preset consistency threshold, or the reflectivity of the second connected domain changes monotonically along a specific direction, the second connected domain is deleted.

The reflectivity consistency can be determined based on the reflectivity variance. The variance (sample variance) in statistics is the average of the square values of the difference between each sample value and the average of all sample values. In addition, other indicators such as standard deviation or mean square error can also be used to judge the reflectivity consistency, which is not limited here.

In a specific embodiment, a height judgment of a high-reflectivity connected domain is added. For example, if the laser radar is installed on a vehicle, if the height value of the high-reflectivity area is less than 1 meter, the point cloud may be an obstacle. At this time, all low-reflectivity connected domains are not deleted to ensure driving safety. However, if the height information of the point cloud is greater than 1 meter but less than 2.5 meters, a more stringent threshold judgment can be used to prevent accidental deletion of obstacles. Objects with a height greater than 2.5 meters can normally judge low-reflectivity connected domains.

The low-reflection connected domain that meets the conditions obtained by the previous operation can also be further screened to make a judgment on mistaken deletion. For example, the outline of the low-reflection connected domain can be extracted and the burrs can be removed. This step can eliminate the interference of thin rods. For example, the reflectivity gradient change information can be extracted as the edge, and the proportion of the area occupied by the edge information of the low-reflection connected domain can be counted. If the edge information is less (the edge is relatively smooth). For example, to judge the reflectivity consistency of the low-reflection connected domain, it can be judged by judging the reflectivity variance of the low-reflection connected domain. If the reflectivity consistency of the low-reflection connected domain is good, or it changes monotonically in the row direction or column direction, it can be judged that it is low-reflection noise caused by expansion.

For each high-reflection connected domain, the above operation is repeated to filter the low-reflection connected domain of the high-reflection connected domain. Finally, the low-reflection connected domain filtered out in the original point cloud image is deleted to remove the noise point cloud generated by the high-reflection expansion.

FIG. 6 is a flow chart of a method for processing a point cloud according to another embodiment of the present application.

Referring to FIG6 , first, the distance and reflectivity information of the LiDAR point cloud is obtained.

Then, the highly reflective connected components and the distance data connected components can be extracted based on the reflectivity data.

Next, for a certain high-reflection connected domain, all distance connected domains are traversed to obtain the pending area that meets the artifact condition.

Then, the pending areas are identified for false deletion and the pending areas that do not meet the conditions are eliminated.

Next, all high-reflection area judgments are completed and data points in the artifact area are eliminated.

The following is an exemplary description of the point cloud processing effect.

Fig. 7 is a schematic diagram of a point cloud before high-reflection expansion noise removal according to an embodiment of the present application. Fig. 8 is a schematic diagram of a point cloud after high-reflection expansion noise removal according to Fig. 7 .

See Figure 7, which shows a point cloud of a road sign with high reflectivity. Due to the floodlight characteristics of the laser radar, expansion noise appears on the side of the road sign in Figure 7, which may lead to the misjudgment that there are other obstacles on the side of the road sign. In addition, recognizing the expansion noise not only results in incorrect recognition results, but also wastes a lot of computing resources and reduces the response speed.

See Figure 8, which shows a point cloud of a road sign with high reflectivity after high reflectivity expansion noise removal. In Figure 8, the expansion noise on the side of the road sign with high reflectivity has been removed. In addition, the point clouds of the road sign and other objects (including low reflectivity points) are all intact, and no point clouds are deleted by mistake.

Fig. 9 is a schematic diagram of a point cloud before high-reflection expansion noise removal according to an embodiment of the present application. Fig. 10 is a schematic diagram of a point cloud after high-reflection expansion noise removal according to Fig. 9.

See FIG9 , which shows a point cloud of a road sign. Expansion noise appears on the side of the road sign in FIG9 , which may lead to a misjudgment that there are other obstacles on the side of the road sign. Unlike FIG7 , the road sign in FIG9 includes not only a high reflectivity portion, but also a low reflectivity portion. The low reflectivity portion in FIG9 is easily judged as high reflectivity expansion noise and mistakenly deleted.

See Figure 10, which shows a point cloud of a road sign after high-reflection expansion noise removal. In Figure 10, the expansion noise on the side of the road sign has been removed. At the same time, the low-reflectivity part of the road sign and the point clouds of other objects (including low-reflection points) are all preserved intact, and no point clouds are deleted by mistake.

Fig. 11 is a schematic diagram of a point cloud before high-reflection expansion noise removal according to an embodiment of the present application. Fig. 12 is a schematic diagram of a point cloud after high-reflection expansion noise removal according to Fig. 11 .

Figures 11 and 12 are tests for misjudgment of low-reflective objects next to high-reflective objects in the case of non-high-reflective expansion. As shown in Figures 11 and 12, the road sign has both high-reflective parts and small areas of low-reflective parts (such as the area indicated by the arrow). After point cloud processing, the small-sized low-reflective parts around the high-reflective parts of the road sheet were not mistakenly deleted, and the test passed. For example, a low-reflective object was placed above a high-reflective sign, and the low-reflective object was not treated as expansion, indicating that the test results were good.

Fig. 13 is a schematic diagram of a point cloud before high-reflection expansion noise removal according to an embodiment of the present application. Fig. 14 is a schematic diagram of a point cloud after high-reflection expansion noise removal according to Fig. 13 .

Figures 13 and 14 are tests for misjudgment of a low-reflective object between two high-reflective objects. As shown in Figures 13 and 14, a low-reflective object is placed between two high-reflective objects, and the point cloud collected by the LiDAR is processed. After point cloud processing, the point cloud of the low-reflective object between the two high-reflective objects is not mistakenly deleted, and the test passes. For example, there is a low-reflective object with a small area between two high-reflective objects. After the algorithm is used, the point cloud of the low-reflective object is well preserved.

Fig. 15 is a schematic diagram of a point cloud before high-reflection expansion noise removal according to an embodiment of the present application. Fig. 16 is a schematic diagram of a point cloud after high-reflection expansion noise removal according to Fig. 15 .

Figures 15 and 16 are high-reflection expansion tests for road test data. Comparing Figures 15 and 16, it can be seen that the expansion points around the high-reflection road sign are basically removed, but the point cloud of the pole connected to the high-reflection road sign is not removed and is well preserved, and the test results are good.

The point cloud processing method provided in this application can effectively eliminate the artifact point cloud caused by high-reflection expansion, and well retain the low-reflection objects near the high-reflection, thereby improving the ranging accuracy.

Another aspect of the present application also provides a device for processing point clouds.

FIG. 17 is a schematic diagram of the structure of an apparatus for processing point clouds according to an embodiment of the present application.

17 , the device 1700 for processing point clouds includes: a point cloud classification module 1710 , a connected domain acquisition module 1720 , and a noise removal module 1730 .

The point cloud classification module 1710 is used to classify the point cloud into first-category points and/or second-category points based on the reflectivity of the points in the point cloud in response to obtaining the point cloud, wherein the reflectivity of the first-category points is higher than a first reflectivity threshold, and the reflectivity of the second-category points is lower than a second reflectivity threshold.

The connected domain acquisition module 1720 is used to obtain at least one first connected domain corresponding to the first type of points by means of projection, and/or to obtain at least one second connected domain corresponding to the second type of points by means of projection.

The noise removal module 1730 is used to delete points in the point cloud corresponding to the second connected domain if the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, wherein the deletion condition indicates that the points in the second connected domain belong to expansion noise.

Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated again here.

Another aspect of the present application provides a radar.

FIG. 18 is a schematic diagram of the structure of a radar according to an embodiment of the present application.

Referring to FIG. 18 , the radar 1800 may include a circuit. For example, the circuit may implement the method for processing a point cloud as shown above. The circuit may be disposed on a circuit board 1810 , and a plurality of chips, such as a central control chip, may be disposed on the circuit board 1810 . The circuit board 1810 may be disposed in a housing 1820 .

The radar may be a scanning radar or a non-scanning radar. Among them, scanning laser radars include MEMS laser radars, mechanical laser radars, laser radars including multiple scanning devices, etc. Non-scanning laser radars include flash laser radars, phased array laser radars, etc. This application does not limit the type of laser radar.

Another aspect of the present application provides an electronic device.

19 , the electronic device 1900 may include a memory 1910 and a processor 1920. In addition, the electronic device 1900 may also be provided with various sensors such as a laser radar.

The processor 1920 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

The memory 1910 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage. Among them, ROM may store static data or instructions required by the processor 1920 or other modules of the computer. The permanent storage may be a readable and writable storage device. The permanent storage may be a non-volatile storage device that does not lose the stored instructions and data even after the computer is powered off. In some embodiments, the permanent storage device uses a large-capacity storage device (such as a magnetic or optical disk, flash memory) as a permanent storage device. In other embodiments, the permanent storage device may be a removable storage device (such as a floppy disk, optical drive). The system memory may be a readable and writable storage device or a volatile readable and writable storage device, such as a dynamic random access memory. The system memory may store some or all instructions and data required by the processor at run time. In addition, the memory 1910 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (such as DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and/or optical disks may also be used. In some embodiments, the memory 1910 may include a readable and/or writable removable storage device, such as a laser disc (CD), a read-only digital versatile disc (e.g., DVD-ROM, double-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. The computer-readable storage medium does not include carrier waves and transient electronic signals transmitted wirelessly or wired.

The memory 1910 stores executable codes, and when the executable codes are processed by the processor 1920 , the processor 1920 may execute part or all of the methods described above.

In addition, the method according to the present application can also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing some or all of the steps in the above-mentioned method of the present application.

Alternatively, the present application can also be implemented as a computer-readable storage medium (or non-transitory machine-readable storage medium or machine-readable storage medium) on which executable code (or computer program or computer instruction code) is stored. When the executable code (or computer program or computer instruction code) is executed by a processor of an electronic device (or server, etc.), the processor executes part or all of the steps of the above-mentioned method according to the present application.

The embodiments of the present application have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

A method for processing a point cloud, comprising:

Obtaining a point cloud, wherein points in the point cloud have reflectivity;

In response to the point cloud, dividing the point cloud into a first type of points and/or a second type of points based on reflectivity of points in the point cloud, the reflectivity of the first type of points being higher than a first reflectivity threshold and the reflectivity of the second type of points being lower than a second reflectivity threshold;

If the second-type point is connected to the first-type point and the second-type point satisfies a deletion condition, the second-type point is deleted, and the deletion condition indicates that the second-type point belongs to expansion noise.
The method according to claim 1, characterized in that if the second type of point is connected to the first type of point and the second type of point satisfies a deletion condition, deleting the second type of point comprises:

Obtaining at least one first connected domain corresponding to the first type of points by projection, and/or obtaining at least one second connected domain corresponding to the second type of points by projection;

If the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, then the points in the point cloud corresponding to the second connected domain are deleted, and the deletion condition indicates that the points corresponding to the second connected domain belong to expansion noise.
The method according to claim 2, characterized in that:

The first connected domain is a connected domain in the projected binary graph corresponding to the first type of points;

The second connected region is a connected region in the projected binary graph corresponding to the second type of points.
The method according to claim 3 is characterized in that, in the XYZ three-dimensional coordinate system, the projected binary image includes: a first projected binary image in the XY plane, a second projected binary image in the XZ plane, or a third projected binary image in the YZ plane.
The method according to claim 4, further comprising:

If the second connected domain is separated from the first connected domain in at least one of the first projected binary image, the second projected binary image, or the third projected binary image, the points in the second connected domain are retained.
The method according to claim 2, further comprising:

Performing morphological processing on the first connected domain, wherein the morphological processing includes: at least one of a morphological closing operation or a morphological dilation operation;

If the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, deleting points in the point cloud corresponding to the second connected domain comprises:

If the second connected domain is connected to the first connected domain after morphological processing, and the second connected domain satisfies a deletion condition, then the points in the point cloud corresponding to the second connected domain are deleted.
The method according to claim 2, further comprising:

The first connected domain is split to obtain a plurality of sub-connected domains, and the plurality of sub-connected domains respectively correspond to different objects.
The method according to claim 2, wherein the deletion condition includes at least one of the following:

The distance between a point in the point cloud corresponding to the second connected domain and the first connected domain is less than or equal to a distance threshold;

A ratio of an area of the second connected domain to an area of the first connected domain is less than or equal to a ratio threshold.
The method according to claim 2, further comprising:

The second type of points are selected according to a selection criterion based on the height information of the second type of points and a preset height range.
The method according to claim 9 is characterized in that the preset height range includes at least two height ranges, and the rejection criteria corresponding to each of the at least two height ranges are different.
The method according to claim 9, characterized in that after the second type of points are selected according to the selection criteria based on the height information of the second type of points and the preset height range, the method further comprises:

For the retained second connected domain, extract the burr portion of the contour of the second connected domain;

The burr portion in the second connected domain is removed.
The method according to claim 9, characterized in that after the second type of points are selected according to the selection criteria based on the height information of the second type of points and the preset height range, the method further comprises:

For the retained second connected domain, based on the reflectivity gradient change information of the second connected domain, determining the edge of the second connected domain;

If the ratio of the area occupied by the edge of the second connected domain to the area of the second connected domain is smaller than a preset area ratio, the second connected domain is deleted.
The method according to claim 9, characterized in that after the second type of points are selected according to the selection criteria based on the height information of the second type of points and the preset height range, the method further comprises:

For the retained second connected domain, if the reflectivity consistency of the second connected domain is higher than a preset consistency threshold, or the reflectivity of the second connected domain changes monotonically along a specific direction, the second connected domain is deleted.
The method of claim 13, wherein the reflectivity consistency is determined based on reflectivity variance.
The method according to any one of claims 1 to 14 is characterized in that the second reflectivity threshold is determined based on an average reflectivity of the first type of points.
A device for processing point clouds, comprising:

a point cloud classification module, configured to, in response to obtaining a point cloud, classify the point cloud into first-category points and/or second-category points based on reflectivity of points in the point cloud, wherein the reflectivity of the first-category points is higher than a first reflectivity threshold, and the reflectivity of the second-category points is lower than a second reflectivity threshold;

A connected domain acquisition module, used for obtaining at least one first connected domain corresponding to the first type of points by means of projection, and/or obtaining at least one second connected domain corresponding to the second type of points by means of projection;

The noise removal module is used to delete the points in the point cloud corresponding to the second connected domain if the second connected domain is connected to the first connected domain and the second connected domain satisfies a deletion condition, wherein the deletion condition indicates that the points in the second connected domain belong to expansion noise.
A radar comprising the device for processing point clouds as claimed in claim 16.
An electronic device, comprising:

Processor; and

A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to execute the method according to any one of claims 1 to 15.