Point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes

Abstract

This invention discloses a point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes, comprising: 1. Using a laser radar to collect point cloud data from a clear-sky scene and sample it to obtain a clear-sky point cloud dataset, which is then input into a foggy simulation network to generate a foggy point cloud dataset; 2. Encoding the point cloud in the foggy point cloud dataset using a non-uniform cylindrical voxel method; 3. Inputting the encoded point cloud voxel dataset into a neural network for feature extraction, and outputting a point cloud recognition result. This invention can adaptively adjust the algorithm's adaptability to fog concentration, reducing the impact of foggy conditions on the laser radar point cloud, thereby ensuring the accuracy of perceived environmental assessments.

Description

Technical Field

[0001] The present invention belongs to the field of environmental perception technology and is a method for processing laser radar point cloud data. Specifically, it is a method for detecting point cloud targets based on non-uniform cylindrical voxel coding in foggy scenes. Background Art

[0002] In foggy conditions, the light pulses emitted by LiDAR are subject to two main effects when passing through water molecules in the air: extinction and backscattering. Extinction refers to the significant attenuation of the light pulse's energy during propagation, resulting in an overall decrease in the received signal power. This also prevents the light signals emitted by objects within the intended measurement range from being received, reducing the LiDAR's effective sensing range. Backscattering refers to the generation of received signal peaks within the incorrect range, resulting in a large number of false points in the generated LiDAR point cloud.

[0003] Compared to raw point clouds, voxels exhibit greater stability against data corruption, and the size of voxels has a decisive influence on their robustness. For example, larger voxels are more stable against damage caused by motion blur, while smaller voxels are more resilient to data corruption caused by fog, snow, and crosstalk. Therefore, to address the damage caused to point cloud data by foggy weather, it is necessary to divide the voxels as small as possible. However, in order to ensure a sufficient detection range, a reduction in voxel size will inevitably lead to an increase in the number of voxels, resulting in a corresponding increase in the computing power and time required for deep learning networks to process voxels, which is not conducive to real-time detection under foggy conditions. Therefore, it is necessary to design a special voxel encoding method to adapt to the impact of foggy weather environments on lidar point clouds. Summary of the Invention

[0004] In order to address the shortcomings of the above-mentioned existing technologies, the present invention proposes a point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes, so that the volume of cylindrical voxels divided by higher fog concentrations will be smaller in the near distance, thereby adaptively adjusting the algorithm's adaptability to fog and improving detection robustness.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes of the present invention is characterized in that it includes the following steps:

[0007] A. After using LiDAR to collect point cloud data of the target scene under sunny scenes and sampling, the dimension is Sunny sky point cloud dataset ;in, Represents a sunny point cloud dataset The i-th clear sky point cloud in , and The coordinates of , Represents a sunny point cloud dataset The number of point clouds, W represents the i-th sunny point cloud Number of messages;

[0008] B. Using foggy weather simulation network to analyze sunny point cloud dataset Processing is performed to generate the dimension Fog point cloud dataset ;in, Represents a foggy point cloud dataset The A foggy point cloud, The coordinates of ;

[0009] C. Foggy point cloud dataset Perform non-uniform cylindrical voxel encoding to obtain the encoded point cloud voxel dataset in the foggy scene ;

[0010] C1. Establish a cylindrical coordinate system with the laser radar as the coordinate origin, the vehicle's forward direction as the positive direction of the X axis, the direction perpendicular to the ground and upward as the positive direction of the Z axis, and the direction pointing to the driver's left as the positive direction of the Y axis. Use formula (1) to convert the Cartesian coordinate system into Coordinates Perform coordinate transformation to obtain the cylindrical coordinate system Coordinate encoding :

[0011] (1)

[0012] In formula (1), express The Euclidean distance from the projection on the XY plane in the Cartesian coordinate system to the origin of the coordinate system in the cylindrical coordinate system, express The angle between the projection of the XY plane in the Cartesian coordinate system to the origin of the cylindrical coordinate system and the positive direction of the X axis in the Cartesian coordinate system, express The coordinate of the Z axis in the Cartesian coordinate system;

[0013] C2. In the cylindrical coordinate system, select On the axis within the range and Axis unit step is the voxel division angle, On the axis within the scope of Axis unit step The height of the voxel division is used to obtain the foggy point cloud dataset after preliminary division. On the axis Within the range, the foggy point cloud dataset after preliminary division is divided to obtain the final foggy point cloud dataset ;in, and 、 and 、 and Respectively represent the coordinates of the foggy point cloud in the cylindrical coordinate system axis, axis, Minimum and maximum values ​​on the axis;

[0014] C3. Using the preset extinction coefficient , and the maximum sensing distance of the laser radar is obtained by formula (2) :

[0015] (2)

[0016] In formula (2), n is the detectable background noise, g is the adaptive laser gain of the lidar, is the intensity of the lidar point cloud under clean weather conditions; where n, g, and is the intrinsic parameter of the lidar;

[0017] C4. Using the maximum sensing distance of LiDAR ,Will exist The length of the axis is set in sections, so that exist The voxels in the y are evenly divided into S parts, exist arrive The voxels in the foggy scene are evenly divided into T parts, thereby obtaining the encoded point cloud voxel dataset , where S>T;

[0018] D. will Input into the neural network for feature extraction, and then output the final point cloud recognition result.

[0019] The point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes according to the present invention is also characterized in that step B comprises:

[0020] B1. Targets struck by the laser are classified into hard targets and soft targets based on the spatial pulse response of the laser radar. A hard target is one whose temporal signature remains unchanged after the laser hits and reflects. A soft target is one that causes distributed scattering of the laser pulse.

[0021] B2. When laser radar reaches Evenly increase point cloud, thus obtaining a dimension of Point cloud dataset , and calculated by formula (3) Middle Point Cloud Distance to LiDAR ;in, express The number of point clouds in That is Middle Sunny day with a few clouds :

[0022] (3)

[0023] In formula (3), for coordinates;

[0024] B3, obtained from formula (4) Middle Point Cloud The spatial impulse response of And as the spatial impulse response of the hard target, it is obtained by formula (5) The spatial impulse response of And as the spatial impulse response of the k-th soft target:

[0025] (4)

[0026] (5)

[0027] In formula (4) and formula (5), represents the scattering coefficient of the hard target, represents the scattering coefficient of the soft target, is the Dirac function, is a unit step function;

[0028] B4. Calculate the signal receiving power of the hard target using equations (6) and (7) respectively and the signal receiving power of the kth soft target :

[0029] (6)

[0030] (7)

[0031] In formula (6) and formula (7), is the half-power pulse width of the laser radar, represents the peak pulse power of the laser radar, c is the speed of light, t is the time, is the system constant of the lidar, represents the cross function;

[0032] B5. Calculate the maximum received signal power of the soft target ,in, Indicates the point cloud corresponding to the maximum signal receiving power Distance to the lidar;

[0033] like , then retain the point cloud corresponding to the hard target As the i-th fog point cloud Otherwise, As the i-th fog point cloud ;

[0034] B6. After assigning i+1 to i, return to step B2 and execute sequentially to finally generate the foggy point cloud dataset. .

[0035] The electronic device of the present invention includes a memory and a processor, the characteristics of which are that the memory is used to store a target detection program that supports the processor to execute the non-uniform cylindrical voxel encoding of the laser radar point cloud, and the processor is configured to execute the program stored in the memory.

[0036] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, executes the target detection step of the non-uniform cylindrical voxel encoding of the laser radar point cloud.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. In response to the problem of a significant decrease in lidar perception accuracy under foggy conditions, the present invention proposes a point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes. The voxels of the lidar point cloud exhibit better stability in the face of data damage, and non-uniform cylindrical voxel encoding is used when dividing the point cloud data into voxels. The method achieves that the higher the concentration of fog, the smaller the volume of the cylindrical voxels divided by the method of the present invention is in the near distance, thereby being able to adaptively adjust the algorithm's adaptability to fog, which is beneficial to improving the environmental perception capability of intelligent vehicles.

[0039] 2. This paper proposes a strategy for generating a foggy point cloud dataset. Compared to previous foggy datasets, this method first collects a clear-sky LiDAR point cloud dataset, then builds a physical model of foggy LiDAR point clouds to batch-generate the fog that might be present in foggy scenes. Finally, it uses a point cloud fog simulation network to create a foggy point cloud dataset. This solves the problems of high cost, limited data volume, and low generalization of real-world datasets, improves the adaptability of target detection algorithms to adverse weather conditions, and facilitates the environmental perception capabilities of intelligent vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is an overall flow chart of the point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes of the present invention;

[0041] Figure 2 This is a flow chart of the point cloud fog simulation of the present invention;

[0042] Figure 3 This is a structural diagram of the point cloud non-uniform cylindrical voxel encoding of the present invention. DETAILED DESCRIPTION

[0043] In this embodiment, a point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes is described. Figure 1 As shown, the following steps are included:

[0044] A. After using LiDAR to collect point cloud data of the target scene under sunny scene and sampling, the sampling process includes: recording the number of point clouds collected by LiDAR as , and specify the number of point clouds after sampling ,like , then the point cloud collected by the lidar is randomly downsampled until the number of point clouds is Otherwise, the point cloud collected by the lidar is randomly copied until the number of point clouds reaches , and the dimension is Sunny sky point cloud dataset ;in, Represents a sunny point cloud dataset The i-th clear sky point cloud in , and The coordinates of , Representing a point cloud dataset The number of sunny point clouds, W represents the i-th sunny point cloud The number of messages.

[0045] B. Using foggy weather simulation network to analyze sunny point cloud dataset Processing is performed to generate the dimension Fog point cloud dataset ;in, Represents a foggy point cloud dataset The A foggy point cloud, The coordinates of ;Point cloud fog simulation process is as follows Figure 2 As shown;

[0046] B1. Based on the spatial pulse response of the laser radar, the targets hit by the laser are divided into hard targets and soft targets. Among them, hard targets refer to targets whose temporal characteristics of the laser radar do not change when hitting the target and reflecting; soft targets refer to fog molecules that cause distributed scattering of the light pulse. Among them, the clear sky point cloud dataset The point cloud in the image corresponds to the hard target, while the soft target simulates the possible fog points.

[0047] B2. When laser radar reaches Evenly increase point cloud, thus obtaining a dimension of Point cloud dataset , and calculated by formula (3) Middle Point Cloud Distance to LiDAR ;in, express The number of point clouds in That is Middle Sunny day with a few clouds :

[0048] (3)

[0049] In formula (3), for The coordinates of the point cloud dataset The points in the dataset are arranged in order. The first point in the dataset is is the point closest to the lidar, and then add point clouds at the same interval, so the point cloud dataset in That is Middle Sunny day with a few clouds .

[0050] B3, obtained from formula (4) Middle Point Cloud The spatial impulse response of And as the spatial impulse response of the hard target, it is obtained by formula (5) The spatial impulse response of And as the spatial impulse response of the k-th soft target:

[0051] (4)

[0052] (5)

[0053] In formula (4) and formula (5), represents the scattering coefficient of the hard target, represents the scattering coefficient of the soft target, is the Dirac function, is a unit step function.

[0054] B4. According to electronic engineering and signal processing theory, the output of any linear system can be expressed as the convolution of the input signal and the system function (system impulse response). Therefore, assuming that light propagates in fog by inelastic scattering, that is, all reflected light is not reflected by other fog molecules and no energy is converted to other wavelengths, the received signal power of the lidar is , use equations (6) and (7) to calculate the signal receiving power of the hard target respectively and the signal receiving power of the kth soft target :

[0055] (6)

[0056] (7)

[0057] In formula (6) and formula (7), is the half-power pulse width of the laser radar, represents the peak pulse power of the laser radar, c is the speed of light, t is the time, is the system constant of the laser radar, and its value is ,in, is the loss in the receiver optics, represents the sensor optical aperture area, Represents the cross function, describing the area illuminated by the emitter The area observed by the receiver (sensor optical aperture area), its value is

[0058]

[0059] B5. Calculate the maximum received signal power of the soft target ,in, Indicates the point cloud corresponding to the maximum signal receiving power Distance to the lidar.

[0060] In an environment where both hard and soft targets exist, the LiDAR signal receiving power of the hard target is usually greater than that of the soft target. At this time, the LiDAR can detect the object normally, but it will attenuate as the distance between the object and the LiDAR increases. Therefore, during the propagation of light, it is very likely that the signal receiving power of the soft target will exceed that of the hard target. At this time, fog will "block" the object and generate a virtual point. Therefore, the signal receiving power strength of the soft target and the hard target is compared. If , then retain the point cloud corresponding to the hard target As the i-th fog point cloud Otherwise, As the i-th fog point cloud ;

[0061] B6. After assigning i+1 to i, return to step B2 and execute sequentially to finally generate the foggy point cloud dataset. , thus realizing fog simulation.

[0062] C. Foggy point cloud dataset Perform non-uniform cylindrical voxel encoding to obtain the encoded point cloud voxel dataset in the foggy scene ;

[0063] C1. Establish a cylindrical coordinate system with the laser radar as the coordinate origin, the vehicle's forward direction as the positive direction of the X axis, the direction perpendicular to the ground and upward as the positive direction of the Z axis, and the direction pointing to the driver's left as the positive direction of the Y axis. Use formula (1) to convert the Cartesian coordinate system into Coordinates Perform coordinate transformation to obtain the cylindrical coordinate system Coordinate encoding :

[0064] (1)

[0065] In formula (1), express The Euclidean distance from the projection on the XY plane in the Cartesian coordinate system to the origin of the coordinate system in the cylindrical coordinate system, express The angle between the projection of the XY plane in the Cartesian coordinate system to the origin of the cylindrical coordinate system and the X axis in the Cartesian coordinate system, express The coordinate of the Z axis in the Cartesian coordinate system.

[0066] C2. In the cylindrical coordinate system, select On the axis within the range and Axis unit step is the voxel division angle, On the axis within the scope of Axis unit step is the height of the voxel division, and the foggy point cloud dataset after preliminary division is obtained. In order to make the point cloud detection algorithm adapt to the influence of foggy conditions with different concentrations (i.e., different extinction coefficients), A non-uniform cylindrical voxel encoding is proposed. On the axis Within the range, the foggy point cloud dataset after preliminary division is divided to obtain the final foggy point cloud dataset ;in, and 、 and 、 and Respectively represent the coordinates of the foggy point cloud in the cylindrical coordinate system axis, axis, The minimum and maximum values ​​on the axis.

[0067] C3. Using the preset extinction coefficient , and the maximum sensing distance of the laser radar is obtained by formula (2) :

[0068] (2)

[0069] In formula (2), n is the detectable background noise, g is the adaptive laser gain of the lidar, is the intensity of the lidar point cloud under clean weather conditions; where n, g, and is the intrinsic parameter of the lidar;

[0070] C4, such as Figure 3 As shown, the maximum perception distance of the laser radar is ,Will exist The length of the axis is set in sections, so that exist The voxels in the y are evenly divided into S parts, exist arrive The voxels in the foggy scene are evenly divided into T parts, thereby obtaining the encoded point cloud voxel dataset , where S>T; due to the maximum perception distance With the extinction coefficient Therefore, for higher fog concentrations, the volume of the cylindrical voxels divided by this method is smaller in the near distance, thereby realizing non-uniform cylindrical voxel encoding, which can adaptively adjust the adaptability of the algorithm to fog and improve the robustness of the detection method;

[0071] D. will Input the feature extraction into the neural network to output the final point cloud recognition result, which includes the following steps: the encoded point cloud voxel dataset in the foggy scene First, through a 3D convolution layer and The activation function extracts preliminary features, and then reduces the size of the feature map through the 3D maximum pooling layer. This process is then repeated to further extract and compress features, ultimately obtaining a high-level feature map. The region proposal network is then used to Generate candidate regions, finally fuse the features of the candidate regions with the global features, and use the detection head to output the final detection results. The 3D convolutional network includes: three 3D convolutional layers (32, 64, 128 channels), each convolutional layer is followed by The activation function and the maximum pooling layer have convolution steps of 1, 2, and 3D convolution layers, and the maximum pooling layer has a step size of 2.

[0072] In this embodiment, an electronic device is designed, including a memory and a processor. The memory is used to store a target detection program that supports the processor to execute the non-uniform cylindrical voxel encoding of the lidar point cloud. The processor is configured to execute the program stored in the memory.

[0073] In this embodiment, a computer-readable storage medium is designed, on which a computer program is stored. When the computer program is run by a processor, the target detection step of the non-uniform cylindrical voxel encoding of the lidar point cloud is executed.

Claims

1. A point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes, characterized by: The steps include: A. After using LiDAR to collect point cloud data of the target scene under sunny scenes and sampling, the dimension is Sunny sky point cloud dataset ;in, Represents a sunny point cloud dataset The i-th clear sky point cloud in , and The coordinates of , Represents a sunny point cloud dataset The number of point clouds, W represents the i-th sunny point cloud Number of messages; B. Using foggy weather simulation network to analyze sunny point cloud dataset Processing is performed to generate the dimension Fog point cloud dataset ;in, Represents a foggy point cloud dataset The A foggy point cloud, The coordinates of ; C. Foggy point cloud dataset Perform non-uniform cylindrical voxel encoding to obtain the encoded point cloud voxel dataset in the foggy scene ; C1. Establish a cylindrical coordinate system with the laser radar as the coordinate origin, the vehicle's forward direction as the positive direction of the X axis, the direction perpendicular to the ground and upward as the positive direction of the Z axis, and the direction pointing to the driver's left as the positive direction of the Y axis. Use formula (1) to convert the Cartesian coordinate system into Coordinates Perform coordinate transformation to obtain the cylindrical coordinate system Coordinate encoding : (1) In formula (1), express The Euclidean distance from the projection on the XY plane in the Cartesian coordinate system to the origin of the coordinate system in the cylindrical coordinate system, express The angle between the projection of the XY plane in the Cartesian coordinate system to the origin of the cylindrical coordinate system and the positive direction of the X axis in the Cartesian coordinate system, express The coordinate of the Z axis in the Cartesian coordinate system; C2. In the cylindrical coordinate system, select On the axis within the range and Axis unit step is the voxel division angle, On the axis within the scope of Axis unit step The height of the voxel division is used to obtain the foggy point cloud dataset after preliminary division. On the axis Within the range, the foggy point cloud dataset after preliminary division is divided to obtain the final foggy point cloud dataset ;in, and 、 and 、 and Respectively represent the coordinates of the foggy point cloud in the cylindrical coordinate system axis, axis, Minimum and maximum values ​​on the axis; C3. Using the preset extinction coefficient , and the maximum sensing distance of the laser radar is obtained by formula (2) : (2) In formula (2), n is the detectable background noise, g is the adaptive laser gain of the lidar, is the intensity of the lidar point cloud under clean weather conditions; where n, g, and is the intrinsic parameter of the lidar; C4. Using the maximum sensing distance of LiDAR ,Will exist The length of the axis is set in sections, so that exist The voxels in the y are evenly divided into S parts, exist arrive The voxels in the foggy scene are evenly divided into T parts, thereby obtaining the encoded point cloud voxel dataset , where S>T; D. will Input into the neural network for feature extraction, and then output the final point cloud recognition result.

2. The point cloud target detection method based on non-uniform cylindrical voxel encoding in foggy scenes according to claim 1 is characterized in that: The step B comprises: B1. Targets struck by the laser are classified into hard targets and soft targets based on the spatial pulse response of the laser radar. A hard target is one whose temporal signature remains unchanged after the laser hits and reflects. A soft target is one that causes distributed scattering of the laser pulse. B2. When laser radar reaches Evenly increase point cloud, thus obtaining a dimension of Point cloud dataset , and calculated by formula (3) Middle Point Cloud Distance to LiDAR ;in, express The number of point clouds in That is Middle Sunny day with a few clouds : (3) In formula (3), for coordinates of B3, obtained from formula (4) Middle Point Cloud The spatial impulse response of And as the spatial impulse response of the hard target, it is obtained by formula (5) The spatial impulse response of And as the spatial impulse response of the k-th soft target: (4) (5) In formula (4) and formula (5), represents the scattering coefficient of the hard target, represents the scattering coefficient of the soft target, is the Dirac function, is a unit step function; B4. Calculate the signal receiving power of the hard target using equations (6) and (7) respectively and the signal receiving power of the kth soft target : (6) (7) In formula (6) and formula (7), is the half-power pulse width of the laser radar, represents the peak pulse power of the laser radar, c is the speed of light, t is the time, is the system constant of the lidar, represents the cross function; B5. Calculate the maximum received signal power of the soft target ,in, Indicates the point cloud corresponding to the maximum signal receiving power Distance to the lidar; like , then retain the point cloud corresponding to the hard target As the i-th fog point cloud Otherwise, As the i-th fog point cloud ; B6. After assigning i+1 to i, return to step B2 and execute sequentially to finally generate the foggy point cloud dataset. .

3. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a target detection program that supports the processor to execute the non-uniform cylindrical voxel encoding of the lidar point cloud according to claim 1 or 2, and the processor is configured to execute the program stored in the memory.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the target detection step of the non-uniform cylindrical voxel encoding of the lidar point cloud according to claim 1 or 2 is performed.

Images