CN111487643A - A building detection method based on LiDAR point cloud and near-infrared images - Google Patents

Abstract

The invention discloses a building detection method based on a laser radar point cloud and a near-infrared image. For the normalized vegetation index NDVI of each laser point, vegetation identification is completed based on NDVI and support vector machine classifier; for non-vegetation laser points, the building identification is completed through the nearest neighbor search algorithm and height threshold; for building laser points, the roof is extracted. Seed points and candidate elevation points, and obtain the roof point group based on individual buildings based on the roof seed points; estimate the vertical elevation of the building based on the roof point group of each building; estimate the vertical elevation based on the candidate elevation points. The precise extraction of the elevation points is carried out; the detection of 3D buildings is completed through the roof points and the elevation points obtained by the precise extraction. The invention effectively improves the detection accuracy of the building, and can ensure a higher detail level of the building model.

Description

A building detection method based on LiDAR point cloud and near-infrared images

technical field

The invention relates to the technical field of three-dimensional extraction of remote sensing ground objects, in particular to a building detection method based on laser radar point clouds and near-infrared images.

Background technique

Building detection using remote sensing data has good applications in various fields such as urban planning, real estate and land use analysis. Conceptually, building detection is a classification problem that needs to separate buildings from other objects such as vehicles, ground (roads, lawns), and vegetation (trees and shrubs). Existing building detection algorithms can be divided into three categories according to the different input data.

The first is the aerial image-based method, which adopts pixel-based and object-based image segmentation methods to extract buildings. Heterogeneous pixel-based methods, whether using supervised or unsupervised learning, have the advantage that the complexity is always low, but due to the high complexity and variability of spectra and textures in urban scenes, pixel-based methods are difficult to use high-resolution data. Initial object-based methods mainly use the spectral similarity and homogeneity of pixels to identify objects, which have better performance than pixel-based methods, but it is difficult to determine the optimal scale level. In addition, the main limitation of using images for building detection is that only two-dimensional outlines of buildings can be identified.

The second is to use the original ALS (Airborne Laser Scannerals) point cloud to detect buildings. Due to their high pulse frequency, high vertical resolution, and the ability to directly provide spatial shape information, ALS point clouds have been widely used for building detection. However, such methods only use spatial and intensity features, ignoring texture and spectrum. Considerations that lead to such methods tend to confuse trees and buildings with smooth canopy.

The third category is the fusion of multi-source data from point clouds and aerial imaging for building detection. Rottensteiner et al. in "Rottensteiner, F; Trinder, J.; Clode, S.; Kubik, K. Using the Dempster–Shafermethod for the fusion of LIDAR data and multi-spectral images for building detection. Inf. Fusion 2005, 6, 283-300 .” In the D-S method and the laser point cloud, the interpolation dsm is derived, and the pixel-based building area is obtained by combining with the multispectral image. This method improves the overall accuracy of the result by fusing the image data. The main limitation lies in the interpolation dsm data source. Reduced the resolution of the laser point cloud, the output is a building area instead of a marked point cloud. Awrangjeb et al. in "Awrangjeb, M.; Ravanbakhsh, M.; Fraser, C.S. Automatic detection of residential buildings using LIDAR data and multispectralimagery. ISPRS J. Photogramm. Remote Sens. 2010, 65, 457-467." Two sets of building masks are extracted from point clouds and multispectral images, and the resulting masks are then used to classify possible building objects. This object-based classification method results in a very high misclassification rate between buildings and trees. , when the trees surrounding the building merge with the building's mask, there is no way to separate it.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a building detection method based on lidar point cloud and near-infrared image, so as to solve the problem of high misclassification rate between buildings and trees in high vegetation coverage area and lack of building scanning in the prior art The problem of high-precision detection methods at the point level can effectively improve the detection accuracy of buildings.

In order to achieve the above purpose, the present invention provides the following solutions: The present invention provides a building detection method based on a lidar point cloud and a near-infrared image, comprising the following steps:

Acquire orthophotos and lidar point clouds with near-infrared bands, and register and fuse orthophotos and lidar point clouds;

Calculate the normalized vegetation index NDVI of each laser point in the lidar point cloud after fusion, and complete the rough identification of vegetation in the lidar point cloud based on NDVI and support vector machine SVM classifier;

For the non-vegetation laser points in the lidar point cloud, the ground, low objects and buildings are divided by the nearest neighbor search algorithm and height threshold;

For building laser points in the lidar point cloud, extract roof seed points and candidate facade points, and obtain roof point groups based on individual buildings based on the roof seed points;

Based on the roof point group of each building, the vertical elevation of the building is estimated; based on the candidate elevation points and the estimated vertical elevation, the distance threshold method is used to extract the elevation points; Face points complete the detection of 3D buildings.

Preferably, the specific method for registration and fusion of orthophoto and lidar point cloud includes:

For each laser point in the lidar point cloud, the pixel values of the near-infrared band, red band, and green band of the nearest pixel in the orthophoto are allocated to the laser point, and the orthophoto and lidar point cloud are completed. fusion.

Preferably, the coarse identification method of vegetation in the lidar point cloud includes:

Calculate the NDVI of each laser point according to the near-infrared band spectral value NIR and red band spectral value R of each laser point in the fused lidar point cloud;

Select two types of training samples of vegetation and non-vegetation in the lidar point cloud, and calculate the NDVI of each sample, and train the SVM classifier based on NDVI;

The NDVI of each laser point in the fused lidar point cloud is input to the trained SVM classifier, and the recognition result of vegetation in the lidar point cloud is obtained.

Preferably, the method for dividing the ground, low objects and buildings includes:

Use a nearest neighbor search algorithm to determine the ground points adjacent to each non-vegetated laser point;

Calculate the relative height rh of each non-ground point;

Based on the relative height rh of each non-ground point, the division of buildings and other low-rise objects is done through a preset height threshold.

Preferably, the method for obtaining candidate elevation points and individual building-based roof point groups includes:

Based on the surface curvature c of the laser point of the building, the direction value Nz of the normal vector z and the nearest echo information Echo, a threshold classification method is used to extract reliable roof seed points and candidate elevation points;

Based on the roof seed points, the region growing algorithm and the image segmentation algorithm are used to obtain the roof point group belonging to each individual building.

Preferably, the precise extraction method of building facade points includes:

Based on the roof point group of each building, the roof boundary tracking algorithm and regularization algorithm are used to estimate the vertical elevation; If it is greater than the preset threshold, the candidate elevation point is marked as a false elevation point, otherwise, the candidate elevation point is marked as the real elevation point of the current roof segment, and the precise extraction of the building elevation point is completed.

The present invention discloses the following technical effects:

(1) The present invention adopts the NDVI index to effectively separate the building from the surrounding dense vegetation, which greatly improves the segmentation accuracy of the building and the vegetation. The experimental results on the test data set show that the correctness and completeness of the present invention can reach more than 92% whether it is at the point cloud level or the building object level when it detects buildings.

(2) Compared with the existing building detection methods, the present invention detects the roof points and facade points of each building by integrating laser point cloud data and multi-band orthophotos with near-infrared wavebands, and can detect the roof points and facade points of each building Each point is assigned an object class label, enabling the geometric reconstruction of a single building model to a relatively high level of detail.

(3) The present invention has a high degree of automation in building detection, and is a very practical and promising method for extracting buildings in areas covered by high vegetation.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

Fig. 1 is the flow chart of the building detection method based on lidar point cloud and near-infrared image of the present invention;

FIG. 2 is an intermediate result diagram of verifying the building detection method of the present invention through a test data set in an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Referring to FIG. 1 , this embodiment provides a building detection method based on a lidar point cloud and a near-infrared image, including the following steps:

Step S1, acquiring orthophotos and lidar point clouds with near-infrared bands, and registering and merging the orthophotos and lidar point clouds; specifically including:

For each laser point in the lidar point cloud, the pixel values of the near-infrared band, red band, and green band of the nearest pixel in the orthophoto are allocated to the laser point, and the orthophoto and lidar point cloud are completed. In the fusion result, each laser point includes 9 attributes: x-coordinate value, y-coordinate value, z-coordinate value, echo intensity, the number of echoes, the total number of echoes, the near-infrared band spectral value NIR, The spectral value R in the red band and the spectral value G in the green band.

Step S2: Calculate the NDVI (Normalized Vegetation Index) of each laser point in the lidar point cloud after fusion, and complete the rough identification of vegetation in the lidar point cloud based on the NDVI and the support vector machine SVM classifier; specifically include:

Calculate the NDVI of each laser point according to the near-infrared band spectral value NIR and red band spectral value R of each laser point in the fused lidar point cloud, as shown in formula (1):

NDVI=(NIR-R)/(NIR+R)………………………………(1)

Select two types of training samples of vegetation and non-vegetation in the lidar point cloud, calculate the NDVI of each sample according to formula (1), and train the SVM classifier based on NDVI;

The NDVI of each laser point in the fused lidar point cloud is input to the trained SVM classifier, and the recognition result of vegetation in the lidar point cloud is obtained to complete the rough identification of vegetation.

Step S3, for the non-vegetation laser points in the lidar point cloud, complete the division of the ground, low objects and buildings through the nearest neighbor search algorithm and height threshold; specifically include:

First, use the nearest neighbor search algorithm to determine the ground points adjacent to each non-vegetation laser point;

Second, calculate the relative height rh of each non-ground point, as shown in formula (2);

是当前非地面点的海拔，

是当前非地面点的第j个邻近地面点的海拔，N是当前非地面点的邻近地面点数目，j＝1，2，…，N；in

is the elevation of the current non-ground point,

is the altitude of the jth adjacent ground point of the current non-ground point, N is the number of adjacent ground points of the current non-ground point, j=1, 2,...,N;

Thirdly, based on the relative height rh of each non-ground point, the building and other low-rise objects are divided by a preset height threshold, which is set to 2.5 meters in this embodiment.

Step S4: For the building laser points in the lidar point cloud, extract roof seed points and candidate facade points, and obtain roof point groups based on individual buildings based on the roof seed points; specifically including:

For a point cloud set that only contains building laser points, first, based on the surface curvature c of the building laser points, the direction value Nz of the normal vector z and the nearest echo information Echo, a threshold classification method is used to extract reliable roof seed points and candidates. Facade points; then, based on the roof seed points, the region growing algorithm and the image segmentation algorithm are used to obtain the roof point group belonging to each individual building.

Step S5, based on the roof point group of each building, estimate the vertical elevation of the building; based on the candidate elevation points and the estimated vertical elevation, use the distance threshold method to accurately extract the elevation points; The obtained elevation points complete the detection of 3D buildings. Specifically include:

Based on the roof point group of each building, the roof boundary tracking algorithm and regularization algorithm are used to estimate the vertical elevation; If it is greater than the preset threshold, the candidate elevation point is marked as a false elevation point, otherwise, the candidate elevation point is marked as the real elevation point of the current roof segment; the 3D building is completed by the roof point and the real elevation point detection.

In order to further verify the accuracy of the building detection method of the present invention, the building detection method of the present invention is verified by the test data set, and the schematic diagram of the results of each step is shown in Figure 2; wherein, Figure (a) is the original lidar point cloud plane projection Figure, points of similar heights have similar colors; Figure (b) is a 3D image of the lidar point cloud colored according to the spectral value of the orthophoto; Figure (c) is the lidar point cloud colored according to the NDVI value 3D map; Figure (d) is the mask for rough vegetation extraction using SVM; Figure (e) is the ground and low targets extracted by using the nearest neighbor search algorithm and height threshold technology; Figure 2 (f) is the use of surface curvature c , the direction value Nz of the normal vector z and the reliable roof seed points classified by the nearest echo information; Figure (g) is the result of clustering and filtering the candidate objects of the building roof by using the region growing and region threshold segmentation algorithm; Figure (g) (h) is the extracted building roof. Through the verification of the test data set, the building detection method based on the lidar point cloud and the near-infrared image of the present invention achieves more than 92% of the correctness and completeness of the building detection, and makes the geometric reconstruction of the single building model reach 92%. A relatively high level of detail.

In the description of the present invention, it should be understood that the terms "portrait", "horizontal", "upper", "lower", "front", "rear", "left", "right", "vertical", The orientation or positional relationship indicated by "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention, rather than indicating or It is implied that the device or element referred to must have a particular orientation, be constructed and operate in a particular orientation, and therefore should not be construed as limiting the invention.

The above-mentioned embodiments are only to describe the preferred modes of the present invention, but not to limit the scope of the present invention. Without departing from the design spirit of the present invention, those of ordinary skill in the art can make various modifications to the technical solutions of the present invention. Variations and improvements should fall within the protection scope determined by the claims of the present invention.

Claims (6)

1. A building detection method based on laser radar point cloud and near-infrared images is characterized by comprising the following steps:

acquiring an ortho-image and a laser radar point cloud with near-infrared bands, and registering and fusing the ortho-image and the laser radar point cloud;

calculating a normalized vegetation index NDVI of each laser point in the fused laser radar point cloud, and completing coarse identification of vegetation in the laser radar point cloud based on the NDVI and a Support Vector Machine (SVM) classifier;

for non-vegetation laser points in the laser radar point cloud, the division of the ground, low and short buildings and buildings is completed through a neighbor search algorithm and a height threshold value;

for building laser points in the laser radar point cloud, extracting roof seed points and candidate facade points, and acquiring a roof point group based on an individual building based on the roof seed points;

estimating a vertical facade of the building based on the group of roof points for each building; based on the candidate vertical face points and the estimated vertical face, adopting a distance threshold value method to carry out fine extraction on the vertical face points; and finishing the detection of the three-dimensional building through the roof points and the elevation points obtained by fine extraction.

2. The building detection method based on the lidar point cloud and the near-infrared image as claimed in claim 1, wherein the specific method for registering and fusing the orthoimage and the lidar point cloud comprises:

and aiming at each laser point in the laser radar point cloud, respectively distributing the pixel values of the near infrared band, the red band and the green band of the nearest pixel in the ortho-image to the laser point, and finishing the fusion of the ortho-image and the laser radar point cloud.

3. The building detection method based on the lidar point cloud and the near-infrared image according to claim 1, wherein the rough identification method of vegetation in the lidar point cloud comprises:

calculating the NDVI of each laser point according to the near infrared band spectral value NIR and the red band spectral value R of each laser point in the fused laser radar point cloud;

selecting two types of training samples of vegetation and non-vegetation in the laser radar point cloud, calculating the NDVI of each sample, and training an SVM classifier based on the NDVI;

and inputting the NDVI of each laser point in the fused laser radar point cloud into a trained SVM classifier to obtain a recognition result of the vegetation in the laser radar point cloud.

4. The building detection method based on the lidar point cloud and the near-infrared image as claimed in claim 1, wherein the method for dividing the ground, the low and short buildings and the building comprises:

determining a ground point adjacent to each non-vegetation laser point by using a neighbor search algorithm;

calculating the relative height rh of each non-ground point;

and completing the division of buildings and other low buildings through a preset height threshold value based on the relative height rh of each non-ground point.

5. The lidar point cloud and near-infrared image-based building detection method of claim 1, wherein the candidate facade points and the individual building-based roof point group acquisition method comprise:

extracting reliable roof seed points and candidate facade points by adopting a threshold classification method based on the surface curvature c of the building laser point, the direction value Nz of the normal vector z and the nearest Echo information Echo;

based on the roof seed points, a region growing algorithm and an image segmentation algorithm are adopted to obtain a roof point group belonging to each single building.

6. The building detection method based on the lidar point cloud and the near-infrared image as claimed in claim 1, wherein the method for finely extracting the building facade points comprises the following steps:

based on the roof point group of each building, adopting a roof boundary tracking algorithm and a regularization algorithm to estimate a vertical facade; and calculating the maximum normal distance from the candidate facade point to the estimated vertical facade closest to the candidate facade point, if the maximum normal distance is greater than a preset threshold value, marking the candidate facade point as a false facade point, otherwise, marking the candidate facade point as a real facade point of the current roof segment, and finishing the fine extraction of the facade point of the building.

Images