When the target object is segmented, the judgment of discrete 3D data point clustering is based on whether the distance between points is close. In lidar point cloud data, a large part of data belongs to ground point data, and the ground point cloud has texture, which interferes with the classification and recognition of subsequent obstacle point clouds. If these ground point data are not removed, the segmentation algorithm will fail when segmenting the target object, so it is necessary to filter first. Therefore, removing ground point cloud data is an effective means to reduce the amount of data and improve the accuracy of segmentation algorithm.
Therefore, in order to improve the accuracy and robustness of ground point cloud removal algorithm, many scholars have proposed a large number of research methods, mainly including the following two categories: ground removal research based on grid diagram method and ground removal research based on original scan line data of 3D lidar.
The point cloud scanned by lidar contains most ground points, and the commonly used grid map methods for filtering point clouds on the ground include grid height difference method, normal vector method and height method. The grid height method firstly generates grids according to the grid size, calculates the height difference between the lowest point and the highest point of each grid, compares H with the preset height difference threshold size, and classifies the grids. Finally, the points in the grid are classified according to the grid. The normal vector method assumes that the calculated ground normal vector is vertically downward or upward, that is, the value of the ground point normal vector is (0,0, 1) or (0,0,-1). The method process is to calculate the normal vector of points and set the threshold of normal vector of point classification. Height method is the most commonly used and least time-consuming method to remove ground point clouds. According to the installation position and attitude of lidar, point clouds can be directly divided into ground points and obstacle points according to the set threshold.
The point cloud processing method based on raster map is to project three-dimensional point cloud data to the ground, establish multiple grid units, and cluster the targets by using connected region labeling algorithm or neighborhood expansion strategy. This method is widely used in three-dimensional modeling of lidar. Firstly, the projection process from 3D point cloud to 2D plane greatly reduces the amount of data, which can improve the real-time performance of algorithm processing. Secondly, due to the mapping from point cloud to raster map, the complex 3D point cloud processing problem is transformed into image processing problem, and mature image processing related algorithms can be used to improve the time efficiency of algorithm processing. The grid method is simple, reliable and efficient. However, the grid cell parameters are fixed and usually determined by experience. When the distant target point cloud is sparse, it will be over-segmented, while when the nearby target point cloud is dense, it will be under-segmented. The algorithm relies heavily on the selection of threshold parameters, and often needs frame-by-frame analysis, which will inevitably lose some real-time performance.
Because the original 3D data of 3D lidar contains detailed spatial information, it can also be used to process related point cloud data. The angle values formed by laser scanning lines on the ground and obstacles are significantly different, which can be used as an important basis for separating ground points. The multi-beam laser in lidar horizontally scans the objects in the surrounding environment, and the angle formed between two adjacent objects is very small, while the angle value of the same object is very large. This enlightens us that we can make full use of this feature. When it is larger than the angle threshold, these two points can be considered as the same object, which solves the problem of under-segmentation of adjacent targets. By dividing the non-ground point cloud into different objects, and then identifying the target object, more detailed information of obstacles such as vehicles and pedestrians can be provided for unmanned vehicles, so as to avoid collision with different types of obstacles and avoid them in time. Insufficient segmentation of ground point clouds will lead to missed detection of targets, while excessive segmentation will affect subsequent operations such as identification. Based on the geometric features of point clouds generated by lidar, researchers have proposed a variety of feature construction methods, and the ground removal research based on the original scanning line data of 3D lidar is one of the more commonly used methods.
In the process of collecting three-dimensional point cloud data, lidar will be affected by various factors, so there will be some noise when obtaining data. In fact, in practical work, in addition to the error of self-measurement, it will also be affected by the external environment, such as the target being blocked, obstacles, the surface material of the measured target and so on. In addition, some local large-scale noises cannot be filtered by the same method because they are far away from the target point cloud.
Noise is a point that has nothing to do with the description of target information, and it is useless for the subsequent reconstruction of the whole 3D scene. However, in the actual point cloud data processing algorithm, it is not easy to distinguish noise points from target points with characteristic information, and some characteristic information will inevitably be lost due to many external factors in the denoising process. A good point cloud filtering algorithm not only requires high real-time performance, but also preserves the feature information of the model while denoising [88]. It is necessary to deeply study the noise characteristics of point cloud data in order to propose a better denoising algorithm.
Point cloud data is an unstructured data format. The point cloud data scanned by lidar is affected by the distance between the object and the radar, and its distribution is uneven. The point cloud data of objects near the radar are densely distributed, while the point cloud data of objects far away from the radar are sparsely distributed. In addition, point cloud data has the characteristics of disorder and asymmetry, which leads to the lack of clear and unified data structure in data representation and aggravates the difficulty of subsequent point cloud segmentation and recognition. As an end-to-end network structure, neural network often deals with conventional input data, such as sequences, images, videos and 3D data, and cannot directly deal with unordered data such as point sets. When processing point cloud data with convolution operation, convolution directly discards the shape information of point cloud and only keeps the sequence information of point cloud.
Point cloud filtering is a research hotspot in the field of three-dimensional reconstruction technology, and it is also a crucial step in data processing of many unmanned applications. Three-dimensional point cloud filtering methods are mainly divided into the following three categories, including statistical filtering, neighborhood filtering and projection filtering.
Because the concept of statistics accords with the characteristics of point cloud, many scholars at home and abroad have introduced statistical methods into point cloud filtering technology. Kalogerakis and others applied the robust statistical model framework to point cloud filtering and achieved very good filtering results. Under the framework of this statistical model, the curvature tensor is estimated by the least square iteration method, and the weight is assigned to the samples according to the area around each point in each iteration, thus refining each neighborhood around each point. Then the calculated curvature and statistical weight are used to re-correct the normal distribution. By calculating curvature and normal to minimize global energy and remove outliers, the texture features of point clouds can be well maintained.
Neighborhood-based point cloud filtering technology uses a similarity measurement method to measure the similarity between points and other neighboring points that have great influence on filtering efficiency and results, so as to determine the location of filtering points. Generally speaking, the similarity of point clouds can be measured by their positions, normals and regional positions. In 1998, Tomasi et al. extended the bilateral filter to 3D model denoising. Because bilateral filtering has the characteristic of keeping the edge smooth, it can remove the noise of point cloud data and keep the details well. However, because this method denoises through the grid generation process, extra noise will be introduced in the grid generation process. Compared with data structures such as regular grid and irregular triangulation, voxel is a real 3D structure, and it implies neighborhood relations, which can effectively prevent noise from being introduced in the process of grid generation. However, this method needs to set more parameters, which can not meet the actual industrial application.
Based on projection filtering technology, the position of each point in the point cloud is adjusted by different projection measurements, so as to remove the noise point cloud. However, if the input point cloud is particularly uneven, the point cloud will become more scattered after local optimal projection processing. Sun Du and others proposed a filtering algorithm based on multiple echoes and Fisher discriminant. Firstly, the point cloud is meshed according to the idea of meshing. In each grid, the sample to be determined is divided by the echo frequency and intensity of point cloud data. Secondly, the Fisher discriminant analysis method is used to project the point cloud into a one-dimensional space, and the vegetation point cloud is separated from the ground point cloud by judging the critical value, so as to filter the point cloud on the steep slope. However, this method is only effective for point clouds with echo intensity, but not for point clouds without echo intensity information.
In order to make up for the disorder, asymmetry, unstructured and lack of information of point cloud itself, it is necessary to transform the data form of point cloud before feature recognition and semantic segmentation. The commonly used point cloud morphological transformation methods include gridded point cloud, voxelized point cloud, point cloud spherical mapping and so on.
Voxelization is to keep the feature points on the surface of point cloud and filter out redundant point cloud data without features. Because the common normal vector calculation depends on the number of adjacent points, and the resolution of two cloud points is also different. Therefore, voxelization with the same voxel size is to generate equivalent local regions in two point clouds with different resolutions. In the actual process of point cloud registration algorithm, because the source point cloud data and target point cloud data used for registration are huge and dense, and these original point cloud data contain many points that have no influence on describing the shape characteristics of objects, if the source point cloud and target point cloud are directly configured by the algorithm, the whole process will take a lot of time, so it is necessary to downsample the point cloud and still retain the part in the point cloud that can reflect the shape outline characteristics.
Due to the sparsity, disorder and uneven distribution of point clouds, end-to-end convolutional neural network can not directly operate on disordered point clouds when using deep full convolution neural network structure to semantically segment lidar point clouds. In order to make the end-to-end neural network universal on unordered point cloud data, it is necessary to map the point cloud data first. Common projection methods include plane-based projection, cylindrical projection and spherical point cloud projection.
Reference:
Zhou Tiantian et al. (Research on semantic segmentation algorithm of lidar point cloud based on deep neural network)
Li Hongyu (Research on Point Cloud Data Processing of Lidar)
Fan Xiaohui (Pedestrian Target Detection and Recognition Based on Lidar)