A method for rod extraction and multi-scale recognition in road point clouds

Abstract

The present invention discloses a method for extracting and multi-scale identifying rod-shaped objects from road point clouds, comprising the following steps: (1) acquiring point clouds of roads and roadside features using a vehicle-mounted laser scanner, and outputting point cloud data in PCD format after computation; (2) preprocessing the point cloud data using a cloth simulation filter and downsampling algorithm on a Cloud Compare platform to improve subsequent processing efficiency; (3) fully automatic segmentation of the rod-shaped object point cloud; and (4) rod-shaped object classification by fusing multi-scale classification results. The present invention can solve the problems of low rod-shaped object extraction efficiency and poor recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a method for extracting and multi-scale identifying rod-shaped objects from road point clouds. Background Art

[0002] Current methods for extracting rods from road scene point clouds can be categorized into three main categories: methods based on rod structural features, clustering-then-recognition methods, and template matching methods. Methods based on rod structural features primarily separate rods from point clouds based on their three-dimensional structure. Common methods utilize geometric features such as rod height, echo intensity, and thickness to extract rods. These extraction methods place high demands on the geometric characteristics of rods, requiring the extracted rods to be relatively regular in shape and significantly different from surrounding objects. Clustering-then-recognition methods first process the discrete point cloud, dividing it into classes or clusters based on spatial distance. Rods are then extracted and identified based on the class or cluster characteristics. Template matching methods leverage the fact that identical rods share common geometric properties to match objects in the original point cloud, extracting rods based on pairwise similarities between objects. This method places high demands on the robustness of matching features and suffers from computational inefficiency.

[0003] Current methods for extracting rods from road scenes have two main shortcomings. First, extraction methods fail to account for the diversity and structural complexity of rod types, resulting in most methods being unable to accurately extract rods or requiring high quality preconditions for successful extraction. Second, they place high demands on the quality of the raw data. For the road point cloud to be extracted, the rod point cloud itself must be relatively pure and free of redundant data, significantly reducing the practicality of these extraction methods.

[0004] After extracting pole point clouds, they often need to be identified for road resource surveys and smart transportation development. Currently, the main approaches include semantic-based recognition, machine learning point-by-point classification, and deep learning. Semantic segmentation methods rely on manually formulating a set of semantic rules tailored to specific road environments. These pre-defined rules are often ineffective for identifying poles in different environments, forcing researchers to develop different rules for each specific road scenario, significantly reducing the method's generalizability. Machine learning point-by-point recognition methods often only consider the features of points within a specific neighborhood, failing to account for the interplay between features at different scales. While they can achieve good recognition results for point cloud data with significant differences within a local neighborhood, they often struggle with rods with similar features within that neighborhood. Deep learning methods require a large number of training samples and are not universally applicable across different scenarios. These algorithmic limitations make them unsuitable for widespread use. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for extracting and multi-scale identifying rods from road point clouds, which can solve the problems of low rod extraction efficiency and poor recognition accuracy.

[0006] To solve the above technical problems, the present invention provides a method for extracting and multi-scale identifying rod-shaped objects from road point clouds, comprising the following steps:

[0007] (1) Use a vehicle-mounted laser scanner to obtain point clouds of roads and roadside features, and output point cloud data in PCD format after solving;

[0008] (2) Point cloud data preprocessing: using cloth simulation filtering and downsampling algorithms on the Cloud Compare platform to preprocess the point cloud and improve the efficiency of subsequent processing;

[0009] (3) Automatic segmentation of rod-shaped object point cloud;

[0010] (4) Rod classification by integrating multi-scale classification results.

[0011] Preferably, in step (3), the fully automatic segmentation of the rod point cloud is specifically as follows: the rod is retained according to three segmentation ideas: the rod has the characteristic of longitudinal continuity, the non-rod part of the rod generally extends perpendicularly to the rod part, and when the one-way double encoding strategy is used for the detailed part of the rod, the difference between the encodings of the two is often large.

[0012] Preferably, the rod-shaped object has the characteristic of longitudinal continuity, and the longitudinal continuity is retained; first, the road scene point cloud data is voxel-divided, and the point cloud data of the large scene is divided into 0.3*0.3*0.3 voxel blocks; secondly, the voxel blocks divided in the previous step are traversed, and the traversal order is traversed from bottom to top according to each column of voxel blocks; finally, the voxel blocks containing point clouds in this column are recorded. If the number of voxel blocks containing point clouds in the voxel block is greater than a certain threshold, it is considered to be a potential rod-shaped object and the column is retained.

[0013] Preferably, the non-rod portion of the rod-shaped object generally extends perpendicularly to the rod portion, and the non-rod portion of the rod-shaped object is retained; the non-rod portion is mostly perpendicular to the extracted rod portion, and a voxel growth strategy perpendicular to the rod portion is used to retain the non-rod portion of the rod-shaped object; the voxels of the rod-shaped object retained in the first step are used as initial seed voxels to perform voxel-based region growth until there are no more continuous voxels with point clouds in the horizontal direction, and the growth is stopped.

[0014] Preferably, for the details of the rod-shaped object, when using a one-way double coding strategy, the difference between the two codes is large, and the one-way double coding strategy retains the details of the rod-shaped object; for the part that is not directly connected to the rod part of the rod-shaped object, a one-way double coding strategy is used to perform secondary retention on it; taking each vertical column of the voxel grid as the research object, one of the double codings is to start coding from the lowest point voxel of the voxel grid to the highest point voxel; the other coding is to use the voxel where the point cloud appears as the initial voxel for coding, and only encode when encountering a voxel containing a point cloud. The above-mentioned unretained point cloud data is retained by using the characteristic that the difference between the double codes is greater than a certain threshold. At the same time, low vegetation is filtered out due to the characteristic that the two codes are similar.

[0015] Preferably, in step (4), the rod-shaped object classification by fusing the multi-scale classification results specifically includes the following steps:

[0016] (41) Obtaining local point cloud features of the rod-shaped object;

[0017] (42) Obtain the global point cloud features of the rod-shaped object,

[0018] (43) Fusion of classification results at different scales.

[0019] Preferably, in step (41), obtaining the local point cloud features of the rod-shaped object is specifically as follows: first, a 14-dimensional point cloud feature vector is constructed according to the structural characteristics of the rod-shaped object, including: intensity, height difference, height difference variance, anisotropy, surface feature, spherical feature, total variance, line feature, number of points in the cylinder, cylinder height difference, density, volume density, curvature and roughness, different rod-shaped objects are given different labels in turn, and then the feature vector and the label are combined and a random forest classifier is used to classify the different rod-shaped object point clouds.

[0020] Preferably, in step (42), the segmented rod-shaped object point cloud is clustered, and different point clouds are clustered into different point cloud clusters. Since they are all rod-shaped object point clouds, different point cloud clusters must exist. Different point cloud clusters are sliced and clustered, and landing points of the point cloud clusters are obtained according to the fact that different rod-shaped objects have different rod bodies. After the landing points are obtained, rod-shaped objects with a distance between landing points less than a certain threshold are extracted, and the two are regarded as overlapping rod-shaped objects. The supervoxel generation results of two different restriction conditions are fused to generate supervoxels for the overlapping area. The two are divided according to the supervoxels and used as a single rod-shaped object point cloud. The features of the above-mentioned single rod-shaped objects are obtained; wherein the global features include: a global viewpoint feature histogram VFH, an outer bounding box geometric size, a voxel ratio and an average intensity. The feature vector composed of the calculated global features is put into a random forest classifier, and the trained model is used to classify the scene to be identified.

[0021] Preferably, in step (43), the recognition effects at different scales are integrated according to the recognition progress, and the parts with better performance in the two recognition processes are merged to improve the recognition accuracy at a single scale.

[0022] The beneficial effects of the present invention are as follows: the present invention takes the point cloud obtained by the vehicle-mounted three-dimensional laser scanning system as the entry point, and makes certain improvements to the extraction and recognition methods of rod-shaped point clouds commonly seen in road scenes according to the characteristics of road scene point cloud data: in terms of segmentation, the present invention uses three constraints and growth conditions to segment rod-shaped objects. Compared with the traditional method of retaining rod-shaped objects only based on their geometric structure, this method retains rod-shaped objects more completely, is not restricted by the complexity of the scene, and is more universal in extraction methods; in the recognition of rod-shaped objects, the present invention uses classification results at two different scales for fusion, which solves the limitation of traditional recognition schemes that only recognize features at a certain neighborhood scale, and improves the recognition accuracy to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the method flow of the present invention

[0024] Figure 2 Schematic diagram of the acquisition device of the present invention.

[0025] Figure 3 The front view is reserved for the present invention based on a longitudinally continuous rod.

[0026] FIG4( a ) is a top view of the voxel growth coordinate system of the present invention.

[0027] FIG4( b ) is a schematic diagram of local voxel growth according to the present invention.

[0028] Figure 5 This is a schematic diagram of the rod-shaped object point cloud extracted by the present invention.

[0029] Figure 6 Schematic diagram of the rod-shaped object point cloud after purification according to the present invention.

[0030] Figure 7 Schematic diagram of the overlapping area segmentation result of the present invention. DETAILED DESCRIPTION

[0031] A method for extracting and multi-scale identifying rod-shaped objects from a road point cloud comprises the following steps:

[0032] (1) Use a vehicle-mounted laser scanner to obtain point clouds of roads and roadside features, and output point cloud data in PCD format after solving;

[0033] (2) Point cloud data preprocessing: using cloth simulation filtering and downsampling algorithms on the Cloud Compare platform to preprocess the point cloud and improve the efficiency of subsequent processing;

[0034] (3) Automatic segmentation of rod-shaped object point cloud;

[0035] (4) Rod classification by integrating multi-scale classification results.

[0036] This paper uses the Alpha3D vehicle-mounted mobile measurement system to collect data on a certain urban road. The system has a horizontal accuracy of <0.030m RMS, a vertical accuracy of <0.025m RMS, a laser point frequency of up to 1,000,000 points / second, and a measurement accuracy of up to 2mm. A total of about 3km of experimental data was collected, with a total of 60,244,135 points, including natural rods, artificial rods, vehicles, and low vegetation. Scanning equipment such as Figure 2 shown.

[0037] First, the scene point cloud of the experimental road is divided into voxel blocks with a size of 0.3*0.3*0.3. The division method is shown in the following equations 1 and 2.

[0038] M x =ceil((x max -xmin ) / S)

[0039] N y =ceil((y max -y min ) / S)

[0040] H z =ceil((z max -z min ) / S) (1)

[0041] M=ceil((point.xx min ) / S)

[0042] N=ceil((point.yy min ) / S)

[0043] H=ceil((point.zz min ) / S) (2)

[0044] x in the formula max 、x min 、y max 、y min 、z max 、z min , point.x, point.y, and point.z are the maximum and minimum coordinate values of the preprocessed point cloud and the specific x, y, and z values of a point, respectively. x 、N y 、H z Represents the number of voxel grids in the X-axis direction, Y-axis direction and Z-axis direction respectively, and M, N, and H represent the row and column number and height position of the point in the voxel grid respectively. The point cloud data divided into voxels is searched by column. If the number of voxel blocks containing point clouds in a column meets a certain threshold, the column is retained. The retained point cloud is the rod part of the rod-shaped object. The front view of the retention principle diagram is as follows Figure 3 shown.

[0045] Region growing and dual encoding strategies for horizontal grids. After preserving the rod-like components based on longitudinal continuity, only the rod-like components on both sides of the road are retained, leaving behind the non-rod components. To reduce computational effort while preserving the non-rod components, a horizontal grid region growing algorithm is employed. The key idea behind this algorithm is to establish a local spatial coordinate system using the grid of the vertically retained rod-like components as a reference, and then diverge queries into the four quadrants. Based on the morphological connection between the non-rod components and the rod-like components, if a continuously diverging horizontal grid appears around the reference grid, it is considered to be the non-rod component of the rod and is retained. Schematic diagrams of the algorithm are shown in Figures 4(a) and 4(b) below. Even after this step, some non-rod components are still omitted from the preservation. Portions that protrude above the rod, such as the lamp caps of streetlights, and portions not directly connected to the rod, such as the lower half of traffic lights, are not fully preserved. Therefore, this patent employs a one-way dual encoding strategy to perform a secondary preservation of these components. The mathematical model of the one-way dual coding method is expressed as D(n1,n2), where n1 and n2 represent the number of voxel codes from bottom to top in a certain column and the number of voxel codes containing point clouds, respectively. The detailed algorithm idea is: taking the voxel grid of each vertical column as the research object, one of the dual coding is to start coding from the lowest point grid of the voxel grid to the highest point grid. The other coding is to use the voxel where the point cloud appears as the initial voxel for coding, and only encode when encountering a voxel containing a point cloud. For the above-mentioned point cloud data that has not been retained, the difference between the dual codes is greater than a certain threshold to retain it. At the same time, the low vegetation has two similar codes and can also be filtered out to a certain extent. The rod-shaped object point cloud after these two steps is as follows Figure 5 shown.

[0046] In order to obtain purer rod-shaped object point cloud data, the present invention uses a method combining Euclidean distance clustering with projection height difference to purify the rod-shaped object point cloud. First, Euclidean distance clustering is used to divide points with similar distances into the same point cluster. When the number of point clusters is too small, the point is considered to be a noise point, and the point cloud is filtered in clusters. This operation can effectively remove some outliers. After removing the outliers, the three-dimensional points are projected two-dimensionally according to the Z axis, and the projected point cloud is divided into a grid that is finer than the above-mentioned one. When the maximum height of the point in the grid and the height difference within the grid are less than the threshold, it is considered to be low vegetation and is also filtered out. The purified point cloud is as follows Figure 6 shown.

[0047] Based on the segmentation of the rod point cloud, the classification and attribute output are performed. First, a complete rod point cloud after segmentation is selected as the standard model, and the R neighborhood features of this standard model are obtained. These include attribute features and geometric features, totaling 14 dimensions. Specifically, these include: strength, height difference, height variance, anisotropy, surface features, spherical features, total variance, line features, cylinder height difference, number of points within the cylinder, density, volume density, curvature, and roughness. These are defined as a 14-dimensional feature vector.

[0048] The feature vector is input into the random forest model for training to obtain a classifier model of local neighborhood features. Then, the features of the extracted rods in the overall R neighborhood are obtained to form a feature vector and the trained classifier is used to classify rod-shaped objects in the local neighborhood.

[0049] Relying solely on local neighborhoods for recognition is not robust because it only considers the features exhibited within the local neighborhood of the point cloud. Therefore, this method combines the global features of the point cloud to classify rod-shaped objects. The steps are as follows: First, the extracted rod-shaped objects are Euclidean clustered, and the point cloud of the large scene is clustered into point cloud clusters. Then, the landing points are obtained by slice clustering, and the distance between the two landing points is calculated. If the distance between the two landing points is too small, it is considered that there is overlap between the two rod-shaped entities, and they are extracted. The extracted overlapping parts are generated into two supervoxels under different restrictions. The two supervoxels are merged, and then the rod-shaped supervoxels in the supervoxels are extracted. Then, the spherical supervoxels are classified to the nearest rod-shaped object according to the distance. After the spherical supervoxel growth is completed, the spherical supervoxels are also classified according to this scheme. The above method can separate the overlapping rod-shaped objects, and the overlapping ground object segmentation effect is as follows. Figure 7 shown.

[0050] The clustered point cloud is processed according to the above conditions to obtain non-overlapping rods. A relatively complete portion of the rod point cloud is selected from the collected data for global feature extraction. These global features include the geometric dimensions of the rod's outer bounding box, the average rod intensity, and the varying proportions of supervoxels. This global feature is then fed into a random forest model using a local training method. This trains a rod classifier based on global features. The segmented point cloud is then subjected to global feature extraction using the same method and fed into the trained random forest model as a prediction sample.

[0051] The prediction results made under the local neighborhood features are more robust to missing situations because they only take into account the point cloud features in a certain neighborhood. However, for some similar rod-shaped objects, only considering the neighborhood features cannot identify the two. Compared with local features, the recognition results of global features are more robust for similar rod-shaped objects, but because it considers the features of the point cloud globally, this recognition result is not ideal for missing trees or trees that are relatively short in spatial morphology. Therefore, this method fuses the two recognition results. Experimental results show that compared with the recognition results under a single scale, this method effectively improves the recognition accuracy. The recognition results of the fusion of the two are shown in Table 1 below. In terms of accuracy assessment, the present invention uses the judgment method shown in Formula 3.

[0052]

[0053]

[0054] Where RDP represents the correct extraction rate, ADP represents the complete extraction rate, a represents the number of correctly extracted data, b represents the number of incorrectly extracted data, and c represents the actual number. The recognition accuracy of the present invention has been verified to be 96%, which is indeed higher than that of previous methods.

[0055] Table 1 Recognition results of the fusion of the two

[0056] type Actual quantity (c) Correct quantity (a) Number of missed extractions (n) Wrong amount (b) RDP (%) ADP (%) Signage 4 4 0 0 100 100 Low signage 2 1 1 0 100 50 Low traffic light 2 1 1 1 50 50 traffic light 4 3 1 2 60 75 monitor 7 6 1 1 85.7 85.7 street lights 26 24 2 0 100 92.3 trees 154 152 2 0 100 98.7

[0057] The present invention is based on vehicle-mounted three-dimensional laser point cloud data, follows the prior knowledge that road poles are continuous in the Z direction and that the non-pole part extends perpendicular to the pole part, and combines the one-way dual coding criterion to effectively extract the road pole point cloud. In the segmentation of overlapping areas, the present invention uses a method of combining supervoxels under different rule constraints to effectively segment the overlapping areas. The concise and intuitive extraction rules effectively improve the segmentation efficiency, and the complete extraction of poles brings great convenience to the subsequent classification. In the pole classification part, a method of fusing recognition results at different scales is used to effectively reduce the limitations of point cloud recognition under single-scale features and improve the recognition accuracy. In addition, the recognition results are more stable, more universal and efficient.

Claims

1. A method for extracting and multi-scale identifying rod-shaped objects from road point clouds, characterized in that: The steps include: (1) Use a vehicle-mounted laser scanner to obtain point clouds of roads and roadside features, and output point cloud data in PCD format after solving; (2) Point cloud data preprocessing: using cloth simulation filtering and downsampling algorithms on the Cloud Compare platform to preprocess the point cloud and improve the efficiency of subsequent processing; (3) Fully automatic segmentation of rod-shaped point clouds: The full automatic segmentation of rod-shaped point clouds follows three segmentation ideas to preserve the rod-shaped objects: the rod-shaped objects have the characteristics of longitudinal continuity, and the non-rod part of the rod-shaped objects extends perpendicularly to the rod part; Rod-shaped objects are characterized by longitudinal continuity, so they are retained. First, the road scene point cloud data is divided into voxel blocks, and the point cloud data of the large scene is divided into 0.3*0.3*0.3 voxel blocks. Then, the voxel blocks divided in the previous step are traversed from bottom to top in the order of each column of voxel blocks. Finally, the voxel blocks containing point clouds in this column are recorded. If the number of voxel blocks containing point clouds in a voxel block is greater than a certain threshold, it is considered a potential rod-shaped object and the column is retained. The non-rod portion of the rod extends perpendicularly to the rod portion, and the non-rod portion of the rod is retained. The non-rod portion is perpendicular to the extracted rod portion, and a voxel growth strategy perpendicular to the rod portion is used to retain the non-rod portion of the rod. The voxels of the rod retained in the first step are used as initial seed voxels for voxel-based region growth until there are no more continuous voxels with point clouds in the horizontal direction, at which point the growth stops. A one-way dual encoding strategy preserves the details of the rod-shaped objects; a one-way dual encoding strategy is used to retain the parts not directly connected to the rod-shaped parts. Taking each vertical column of voxel grid as the research object, one of the dual encodings starts from the lowest voxel in the voxel grid and continues to the highest voxel. The other encoding process starts with the voxel where the point cloud appears as the initial voxel for encoding, and only encodes the voxel containing the point cloud. The unretained point cloud data is retained by taking advantage of the fact that the difference between the dual encodings is greater than a certain threshold. At the same time, low vegetation is filtered out because the two codes are similar. (4) Rod classification by integrating multi-scale classification results.

2. The method for extracting and multi-scale identifying rod-shaped objects from road point clouds according to claim 1, wherein: In step (4), the rod classification based on the fusion of multi-scale classification results specifically includes the following steps: (41) Obtaining local point cloud features of the rod-shaped object; (42) Obtain the global point cloud features of the rod-shaped object, (43) Fusion of classification results at different scales.

3. The method for extracting and multi-scale identifying rod-shaped objects from road point clouds according to claim 2, wherein: In step (41), the local point cloud features of the rod-shaped object are obtained as follows: first, a 14-dimensional point cloud feature vector is constructed according to the structural characteristics of the rod-shaped object, which includes: intensity, height difference, height difference variance, anisotropy, surface feature, spherical feature, total variance, line feature, number of points in the cylinder, cylinder height difference, density, volume density, curvature and roughness, and different rod-shaped objects are given different labels in turn. Then, the feature vector and the label are combined and the random forest classifier is used to classify the different rod-shaped object point clouds.

4. The method for extracting and multi-scale identifying rod-shaped objects from road point clouds according to claim 2, wherein: In step (42), the segmented rod-shaped object point cloud is clustered, and different point clouds are clustered into different point cloud clusters. Then, the different point cloud clusters are sliced and clustered. According to the fact that different rod-shaped objects have different rod bodies, the landing points of the point cloud clusters are obtained. After the landing points are obtained, the rod-shaped objects whose landing points have a distance difference less than a certain threshold are extracted, and the two are regarded as overlapping rod-shaped objects. Two different restriction conditions are used to generate supervoxels in the overlapping area, and the two supervoxel results are fused to obtain the final supervoxel point cloud in the overlapping area; the two are divided according to the supervoxels and used as a single rod-shaped object point cloud, and the features of the above-mentioned single rod-shaped object are obtained; The global features include: global view feature histogram VFH, outer bounding box geometry, voxel ratio and average intensity. The composed feature vector is put into the random forest classifier, and the trained model is used to classify the scene to be identified.

Images