A point cloud background filtering method and system for intelligent roadside unit

Abstract

The present invention provides a point cloud background filtering method for intelligent roadside equipment, comprising the following steps: acquiring point cloud data; preprocessing the point cloud data to obtain preprocessed point cloud data; performing background filtering on the preprocessed point cloud data to obtain a filtered foreground point set. The point cloud background filtering method for intelligent roadside equipment of the present invention can effectively and efficiently filter out background points in the background point cloud in real time.

Description

Technical Field

[0001] The present invention belongs to the field of point cloud data processing, and in particular relates to a point cloud background filtering method and system for an intelligent roadside unit. Background Art

[0002] Vehicle-road collaboration takes vehicle information on the road as the perception target, and uses the new generation of information and communication technology to achieve network connections between vehicles, vehicles and people, vehicles and roads, and vehicles and service platforms, improve the overall driving level of vehicles, and provide users with safe, comfortable, intelligent, and efficient driving experience and traffic services. At the same time, it improves traffic operation efficiency and the intelligence level of social traffic services. In the vehicle-road collaboration system, the intelligent roadside unit is an indispensable component, which is generally deployed at road intersections and is responsible for the all-round perception of complex scenes and the role of blind spots. The roadside unit is mainly composed of a sensor kit, a roadside perception algorithm module, and a communication module. Among them, the sensors are generally laser radars and cameras. The laser radar of the roadside unit generally only needs to detect dynamic targets in the area, and does not care about static targets. 90-95% of the point clouds collected by the roadside laser radar are background points, including the ground, trees, and buildings. Background filtering can filter out these background points in advance, which can greatly reduce the workload of three-dimensional detection on the one hand, and reduce the false detection rate of the detection algorithm on the other hand. Therefore, background filtering plays a very important role in the three-dimensional target detection of roadside units and is an indispensable key step.

[0003] Point cloud background filtering needs to distinguish the point cloud of dynamic targets from the background point cloud. The existing algorithm implementation is generally divided into two steps. The first step is background file production, which collects point cloud data of multiple frames of static scenes, and then stores the location information or feature information of background points to produce background files; the second step is background filtering. For new point cloud frames, each point is traversed and then the background file is queried to see if the point is a background point. There are two technical difficulties in background filtering. One is how to produce a background file that accurately records background information. In actual scenes, due to weather or human factors, background jitter often occurs. Some of this jitter can return to normal in a short time, but some will change permanently over time, which brings challenges to background file production. The other is how to efficiently query whether a point is a background point. The time consumption of existing distance matching or density matching algorithms will increase significantly with the increase in the number of point clouds, which is very unfriendly to embedded devices.

[0004] For example, patent CN111652936A proposes a three-dimensional perception and stacking planning method and system for open container loading, which uses background filtering technology to complete three-dimensional perception. The background filtering technology first collects the three-dimensional point cloud data of the background, then calculates the vertical normal vector and curvature of each point, and then uses the region growing algorithm to divide the background point cloud into multiple facets, each facet has a similar normal vector and curvature; then the algorithm will store the center point, normal direction and maximum radius of each facet as features in the background model configuration file in sequence to complete the background modeling. In the actual point cloud filtering process, the method uses the distance from each point to the background facet as a reference to set a threshold to filter the background point cloud. This method needs to calculate the normal vector and curvature of each point, which consumes huge computing resources and is difficult to meet real-time requirements. Patent CN111999741A proposes a roadside laser radar target detection method, which also includes a point cloud background filtering technology. The background filtering method first collects multiple frames of point clouds in different time periods, then rasterizes each frame of point cloud within the selected three-dimensional detection range, counts the number of point clouds, average height and maximum height difference in each grid, and obtains the background grid statistics table. In the filtering stage, the method first rasterizes the input point cloud frame, then subtracts it from the background grid statistics table, and uses the specified threshold to filter the background points. This method does not consider the sparsity of the point cloud when rasterizing the point cloud, which will bring a lot of unnecessary computing resources and storage space waste. In addition, this method is not suitable for situations where the background is complex and changeable. When the background changes, the background filtering algorithm will fail.

[0005] In view of this, there is an urgent need to propose a point cloud background filtering method and system for intelligent roadside units. Summary of the invention

[0006] To this end, the present invention provides a point cloud background filtering method and system for an intelligent roadside unit.

[0007] The point cloud background filtering method and system for an intelligent roadside unit of the present invention comprises the following steps:

[0008] S1. Obtain point cloud data;

[0009] S2. preprocessing the point cloud data to obtain preprocessed point cloud data;

[0010] S3, performing background filtering on the pre-processed point cloud data to obtain a filtered foreground point set

[0011] In step S3, background filtering of the pre-processed point cloud specifically includes the following steps:

[0012] S31, calculating the point cloud density and hash value of the pre-processed point cloud data in the three-dimensional voxel grid, and generating a point cloud background hash table;

[0013] S32, momentum updates the point cloud background hash table;

[0014] S33, traversing all points in the point cloud data, and generating hash values ​​corresponding to the points;

[0015] S34, determine the type of point, the specific process is as follows:

[0016] If the hash value of a point is not in the background table, the point is a foreground point;

[0017] If the hash value of the point is in the point cloud background hash table, compare whether the point cloud density corresponding to the hash value is greater than the predetermined threshold:

[0018] If it is greater than the preset threshold, the point is a background point.

[0019] If it is less than the predetermined threshold, the point is a foreground point;

[0020] S35, filtering all background points to obtain a foreground point set.

[0021] Furthermore, the specific process of step S31 is as follows:

[0022] S31a, continuously collecting frame data of pre-processed point cloud data;

[0023] S31b, calculating the voxel coordinates of the points in the frame data in the three-dimensional voxel grid, the calculation formula is as follows:

[0024]

[0025] in,

[0026] P(p x , p y , p z ) are the three-dimensional coordinates of the point;

[0027] is the boundary point and its coordinates with the smallest coordinate value of the three-dimensional voxel grid;

[0028] r is the resolution of the three-dimensional voxel grid;

[0029] S31c, calculate the hash value of each point; the hash value represents the position index of the point in the three-dimensional voxel grid, and the calculation formula is as follows:

[0030] hash=Id x +Id y ×L+Id z ×L×W

[0031] Where L and W represent the length and width of the three-dimensional voxel grid, respectively;

[0032] S31d. Calculate the point cloud density of each non-empty 3D voxel grid:

[0033] Point cloud density = the number of points in all frames of the background package accumulated in the voxel grid / the total number of frames of the background package;

[0034] The background package is a collection of frame data of the collected background;

[0035] S31e. Generate a point cloud background hash table corresponding to the hash value and the point cloud density.

[0036] Furthermore, in step S32, momentum updating the point cloud background hash table includes the following steps:

[0037] S32a, after the background area changes, continuously collect frame data of the point cloud data;

[0038] S32b, generating an updated point cloud background hash table corresponding to the frame data according to steps S31b-S31e;

[0039] S32c, momentum updates the point cloud background hash table, and the momentum update formula is as follows:

[0040] D=(1-γ)D+γD new

[0041] in,

[0042] D is the point cloud density in the point cloud background hash table;

[0043] D new To update the point cloud density in the point cloud background hash table;

[0044] γ is an update parameter, and its value range is [0,1].

[0045] Furthermore, the frequency of continuous collection in step S32a is lower than the frequency of continuous collection in step S31a.

[0046] Furthermore, step S1 obtains point cloud data from a device driver of an intelligent roadside device.

[0047] Furthermore, in step S2, preprocessing of the point cloud data includes invalid point removal, outlier point removal, voxel filtering, and region of interest filtering.

[0048] The present invention also provides a point cloud background filtering system for an intelligent roadside unit, comprising:

[0049] The raw data acquisition module obtains point cloud data from the device driver of the intelligent roadside device;

[0050] A data preprocessing module receives the point cloud data transmitted by the original data acquisition module and preprocesses the point cloud data into a preprocessed point cloud;

[0051] A background filtering module receives the preprocessed point cloud processed by the data preprocessing module and performs background filtering;

[0052] The three-dimensional detection module receives the data after background filtering by the background filtering module and detects the information of the point cloud.

[0053] Furthermore, the background filtering module includes:

[0054] Background table making module, which makes the pre-processed point cloud into a point cloud hash background table;

[0055] A background table updating module, which momentum updates the point cloud hash background table;

[0056] The background point filtering module determines the type of points in the point cloud and filters out background points.

[0057] The above technical solution of the present invention has the following advantages compared with the prior art:

[0058] The present invention introduces a three-dimensional voxel table to quantify point cloud data into a point cloud background hash table, retrieves background points and foreground points from the point cloud background hash table, and filters out background points, thus achieving efficient background filtering.

[0059] The present invention also monitors background changes in real time by means of momentum update to update the point cloud background hash table to meet the complex and changeable environment and realize real-time background filtering. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a schematic diagram of module connections of a system provided by an embodiment of the present invention;

[0061] Figure 2 The point cloud data in the method provided by the embodiment of the present invention is converted into a point cloud background hash table. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] The point cloud background filtering method and system for an intelligent roadside unit of the present invention comprises the following steps:

[0064] S1. Obtain point cloud data.

[0065] Specifically, this embodiment acquires point cloud data from a device driver of an intelligent roadside device.

[0066] S2. Preprocess the point cloud data to obtain preprocessed point cloud data.

[0067] Specifically, the preprocessing of point cloud data includes invalid point removal, outlier point removal, voxel filtering, and region of interest filtering. Invalid point removal is to filter out illegal point clouds generated by the driver; discrete point removal is to filter out external points in the point cloud distribution by statistical analysis. After that, the point cloud data will be downsampled through voxel filtering to reduce the amount of calculation of the method of this embodiment. The region of interest filtering can be a simple box filtering, or it can be filtered according to the lane information provided by the high-precision map to filter out the point cloud outside the area, further improving the processing rate of the method of this embodiment.

[0068] S3. Perform background filtering on the pre-processed point cloud data to obtain a filtered foreground point point set.

[0069] In step S3, background filtering of the pre-processed point cloud specifically includes the following steps:

[0070] S31, calculating the point cloud density and hash value of the pre-processed point cloud data in the three-dimensional voxel grid, and generating a point cloud background hash table.

[0071] The specific process of step S31 is as follows:

[0072] S31a, continuously collecting frame data of pre-processed point cloud data;

[0073] S31b, calculating the voxel coordinates of the points in the frame data in the three-dimensional voxel grid, the calculation formula is as follows:

[0074]

[0075] in,

[0076] P(p x , p y , p z ) are the three-dimensional coordinates of the point;

[0077] is the boundary point and its coordinates with the smallest coordinate value of the three-dimensional voxel grid;

[0078] r is the resolution of the three-dimensional voxel grid;

[0079] S31c, calculate the hash value of each point; the hash value represents the position index of the point in the three-dimensional voxel grid, and the calculation formula is as follows:

[0080] hash=Id x +Id y ×L+Idz ×L×W

[0081] Where L and W represent the length and width of the three-dimensional voxel grid, respectively;

[0082] S31d. Calculate the point cloud density of each non-empty 3D voxel grid:

[0083] Point cloud density = number of points in the non-empty voxel grid / number of all points in the frame data;

[0084] The background package is a collection of frame data of the collected background;

[0085] S31e. Generate a point cloud background hash table corresponding to the hash value and the point cloud density.

[0086] Furthermore, through 3D voxel encoding, all points of each frame of the point cloud can be mapped to a unique voxel grid, and then the point cloud density of the voxel grid can be calculated by counting the number of point clouds in each non-empty voxel grid. Since the background points exist in almost every frame, the point cloud density of the voxel grid corresponding to the background points is basically 1; and the foreground points only appear in a fixed voxel grid in a few frames, so the point cloud density of the voxel grid corresponding to the foreground points is very small. By recording the voxels and their corresponding point cloud density values, a point cloud background hash table can be constructed, which encodes the background information well.

[0087] The resolution and length, width and height of the three-dimensional voxel grid can be set according to different scenes. A high grid resolution can achieve fine encoding of the background, but it is not conducive to the algorithm's adaptation to background jitter. It should also be noted that the three-dimensional voxel grid is introduced in this embodiment only for the convenience of understanding. In the method of this embodiment, it is not necessary to maintain the three-dimensional voxel grid. It is sufficient to calculate the hash value through the point cloud coordinate information and record the density information.

[0088] S32: Momentum update the point cloud background hash table.

[0089] In order to adapt to the situation where the background changes, when a certain area of ​​the background changes, the corresponding voxel point cloud density information in the pre-made background table is no longer reliable, and it is necessary to re-record the point cloud frame and calculate the density value of the area for update. Therefore, the method of this embodiment introduces a background table update mechanism (momentum update).

[0090] Momentum updating of the point cloud background hash table includes the following steps:

[0091] S32a, after the background area changes, continuously collect frame data of the point cloud data;

[0092] S32b, generating an updated point cloud background hash table corresponding to the frame data according to steps S31b-S31e;

[0093] S32c, momentum updates the point cloud background hash table, and the momentum update formula is as follows:

[0094] D=(1-γ)D+γD new

[0095] in,

[0096] D is the point cloud density in the point cloud background hash table;

[0097] D new To update the point cloud density in the point cloud background hash table;

[0098] γ is an update parameter, and its value range is [0,1].

[0099] When the value of γ is close to 0, it means that the density update is more conservative and trusts the historical value more; when the value of γ is close to 1, it means that the algorithm is more inclined to the new density value. According to the characteristics of the background, the value of γ in different areas can be different. For fixed buildings and other background areas that basically do not change, the value of γ can be close to 0, while for trees, roads and other background areas that are easily disturbed by the environment, the value of γ can be close to 1. Through this update method, a good update effect can be achieved with a small computational cost.

[0100] S33, traverse all points in the point cloud data and generate hash values ​​corresponding to the points.

[0101] S34, determine the type of point, the specific process is as follows:

[0102] If the hash value of a point is not in the background table, the point is a foreground point;

[0103] If the hash value of the point is in the point cloud background hash table, compare whether the point cloud density corresponding to the hash value is greater than the predetermined threshold:

[0104] If it is greater than the preset threshold, the point is a background point.

[0105] If it is less than the predetermined threshold, the point is a foreground point;

[0106] S35, filtering all background points to obtain a foreground point set.

[0107] The entire judgment process in the method of this embodiment only requires one query and one comparison with the density threshold, and the time complexity is O(1), thereby achieving efficient filtering.

[0108] Furthermore, the frequency of continuous collection in step S32a is lower than the frequency of continuous collection in step S31a.

[0109] This embodiment also provides a point cloud background filtering system for an intelligent roadside unit, such as Figure 2As shown, including:

[0110] The raw data acquisition module obtains point cloud data from the device driver of the intelligent roadside device;

[0111] A data preprocessing module receives the point cloud data transmitted by the original data acquisition module and preprocesses the point cloud data into a preprocessed point cloud;

[0112] A background filtering module receives the preprocessed point cloud processed by the data preprocessing module and performs background filtering;

[0113] The three-dimensional detection module receives the data after background filtering by the background filtering module and detects the information of the point cloud.

[0114] Furthermore, the background filtering module includes:

[0115] Background table making module, which makes the pre-processed point cloud into a point cloud hash background table;

[0116] A background table updating module, which momentum updates the point cloud hash background table;

[0117] The background point filtering module determines the type of points in the point cloud and filters out background points.

[0118] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A point cloud background filtering method for intelligent roadside equipment, characterized in that: The steps include: S1. Obtain point cloud data; S2. preprocessing the point cloud data to obtain preprocessed point cloud data; S3, performing background filtering on the pre-processed point cloud data to obtain a filtered foreground point set; in step S3, performing background filtering on the pre-processed point cloud specifically includes the following steps: S31, calculating the point cloud density and hash value of the pre-processed point cloud data in the three-dimensional voxel grid, and generating a point cloud background hash table; S32, momentum updates the point cloud background hash table; S33, traversing all points in the point cloud data, and generating hash values ​​corresponding to the points; S34, determine the type of point, the specific process is as follows: If the hash value of a point is not in the background table, the point is a foreground point; If the hash value of the point is in the point cloud background hash table, compare whether the point cloud density corresponding to the hash value is greater than the predetermined threshold: If it is greater than the preset threshold, the point is a background point. If it is less than a predetermined threshold, the point is a foreground point; S35, filter all background points to obtain a foreground point set; the specific process of step S31 is as follows: S31a, continuously collecting frame data of pre-processed point cloud data; S31b, calculating the voxel coordinates of the points in the frame data in the three-dimensional voxel grid, the calculation formula is as follows: in, P(p x , p y , p z ) are the three-dimensional coordinates of the point; is the boundary point and its coordinates with the smallest coordinate value of the three-dimensional voxel grid; r is the resolution of the three-dimensional voxel grid; S31c, calculate the hash value of each point; the hash value represents the position index of the point in the three-dimensional voxel grid, and the calculation formula is as follows: hash=Id x +Id y ×L+Id z ×L×W Where L and W represent the length and width of the three-dimensional voxel grid, respectively; S31d. Calculate the point cloud density of each non-empty 3D voxel grid: Point cloud density = the number of points in all frames of the background package accumulated in the voxel grid / the total number of frames of the background package; S31e, generating a point cloud background hash table corresponding to the hash value and the point cloud density; in step S32, momentum updating the point cloud background hash table includes the following steps: S32a, after the background area changes, continuously collect frame data of the point cloud data; S32b, generating an updated point cloud background hash table corresponding to the frame data according to steps S31b-S31e; S32c, momentum updates the point cloud background hash table, and the momentum update formula is as follows: D=(1-γ)D+γD new in, D is the point cloud density in the point cloud background hash table; D new To update the point cloud density in the point cloud background hash table; Y is the update parameter, and its value range is [0, 1].

2. The point cloud background filtering method for intelligent roadside equipment according to claim 1, characterized in that: The frequency of continuous acquisition in step S32a is lower than the frequency of continuous acquisition in step S31a.

3. The point cloud background filtering method for intelligent roadside equipment according to claim 2, characterized in that: Step S1 obtains point cloud data from the device driver of the intelligent roadside device.

4. The point cloud background filtering method for intelligent roadside equipment according to claim 1, characterized in that: In step S2, the preprocessing of the point cloud data includes invalid point removal, outlier point removal, voxel filtering, and region of interest filtering.

5. A point cloud background filtering system for intelligent roadside equipment, characterized in that: include: The raw data acquisition module obtains point cloud data from the device driver of the intelligent roadside device; A data preprocessing module receives the point cloud data transmitted by the original data acquisition module and preprocesses the point cloud data into a preprocessed point cloud; A background filtering module receives the preprocessed point cloud processed by the data preprocessing module and performs background filtering; The three-dimensional detection module receives the data after background filtering by the background filtering module and detects the information of the point cloud.

6. The point cloud background filtering method for intelligent roadside equipment according to claim 5, characterized in that: The background filtering module comprises: Background table making module, which makes the pre-processed point cloud into a point cloud hash background table; A background table updating module, which momentum updates the point cloud hash background table; The background point filtering module determines the type of points in the point cloud and filters out background points.

Images