DE102020001346A1 - Method for the detection of rod-shaped objects using lidar data - Google Patents

Abstract

The invention relates to a method for the detection of rod-shaped objects (O1 to On) using lidar data, a point cloud recorded by means of a lidar (2) as an input variable using a convolutional neural network in which an object class for rod-shaped objects (01 to On) is provided, is semantically labeled. According to the invention, in a clustering, the point cloud is divided into several object clusters (C1 to Cm) with object candidates (K1 to Km) for rod-shaped objects (O1 to On) and in a subsequent parameter estimation, a three-dimensional one is made for each object candidate (K1 to Km) using a principal component analysis Position and a radius (R) are determined. The invention further relates to a use of such a method for vehicle self-localization or robot self-localization.

Description

The invention relates to a method for detecting rod-shaped objects by means of lidar data according to the preamble of claim 1.

The invention also relates to a use of such a method.

In " Landa et. al .: Detection of pole-like objects from LIDAR data; 19th International Conference Enterprise and Competitive Environment 2016, ECE 2016, 10-11 March 2016, Brno, Czech Republic “Describes an approach to the detection of rod-shaped objects from lidar data. A directional vector is used to detect rod-shaped structures in disordered point clouds.

The invention is based on the object of specifying a method for detecting rod-shaped objects by means of lidar data, which is improved compared to the prior art, and a use of such a method.

The object is achieved according to the invention by a method which has the features specified in claim 1 and by a use which has the features specified in claim 4.

Advantageous embodiments of the invention are the subject of the subclaims.

In a method for detecting rod-shaped objects by means of lidar data, a point cloud captured by means of a lidar is semantically labeled as an input variable by means of a convolutional neural network in which an object class is provided for rod-shaped objects.

According to the invention, the point cloud is divided into several object clusters with object candidates for rod-shaped objects in a clustering and a three-dimensional position and a radius are determined for each object candidate in a subsequent parameter estimation by means of a principal component analysis. Automated, in particular highly automated or autonomous, operated vehicles and robots have to locate themselves in relation to a known map in order to enable safe navigation in their surroundings. For this purpose, it is necessary that the vehicle's own or robot's own sensors recognize and measure known localization features in the environment. In urban environments in particular, there are many at least essentially rod-shaped or cylindrical objects, also known as poles, such as traffic signs, traffic lights, street lamps, flag posts, tree poles, etc. Due to their defined geometric structure, such poles can be used as a good basis for precise lateral and longitudinal Localization of the vehicle or robot are used.

Poles are reliably identified and measured using the present method. The parameter estimation based on the principal component analysis, also known as principal component analysis, PCA for short, leads to a very robust and precise estimation of the three-dimensional position and its radii of the poles, as there are only very few free parameters and no strict modeling requirements. Thus, the method enables robust vehicle self-localization and robot self-localization.

Embodiments of the invention are explained in more detail below with reference to drawings.

• 1 schematisch ein Fahrzeug und einen Ausschnitt einer Umgebung des Fahrzeugs,
• 2 schematisch in einem Clustering aus einer Punktwolke eines Lidars erzeugte Objektcluster mit Objektkandidaten für stabförmige Objekte und
• 3 schematisch ein Bild einer Umgebung eines Fahrzeugs, ein Lidar sowie einen Objektcluster mit einem Objektkandidaten während einer Parameterschätzung.

Show:

• 1 schematically a vehicle and a section of the surroundings of the vehicle,
• 2 object clusters with object candidates for rod-shaped objects and schematically generated in a clustering from a point cloud of a lidar
• 3 schematically an image of the surroundings of a vehicle, a lidar and an object cluster with an object candidate during a parameter estimation.

Corresponding parts are provided with the same reference symbols in all figures.

In 1 are a vehicle 1 and a section of the surroundings of the vehicle 1 shown. The vehicle 1 includes a lidar 2 to an environment detection. In the vicinity of the vehicle 1 is a plurality of rod-shaped or cylindrical objects O1 to On, also known as Poles, are present. Such objects O1 to On are, for example, traffic signs, traffic lights, street lamps, flag posts, tree trunks.

The vehicle 1 is designed for an automated, in particular highly automated or autonomous driving operation. For such a driving operation it is necessary that the vehicle 1 localized with respect to a known digital map in order to allow safe navigation in its surroundings. For this it is necessary that the vehicle's own sensors, in the present embodiment the lidar 2 Recognize and measure known localization features in the environment. The rod-shaped objects are suitable as such localization features, in particular landmarks 01 to On because of their defined geometric structure.

To such rod-shaped objects O1 to On in the vicinity of the vehicle 1 to be detected is provided, one by means of the lidar 2 recorded point cloud as an input variable by means of a convolutional neural network, in which an object class for rod-shaped objects O1 until On is planned to be semantically labeled. That is, the point cloud is semantically marked by means of a convolution network, including an object class for rod-shaped objects O1 until On. This is done, for example, according to the methods described in “Piewak et. al .: Boosting LiDAR-based Semantic Labeling by Cross-Modal Training Data Generation "and" Wu et. al .: SqueezeSeg - Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud “.

Clustering and then parameter estimation are also carried out.

In such a clustering from a point cloud of the lidar 2 generated object clusters C1 up to Cm with object candidates for rod-shaped objects O1 until On shows 2 .

In the clustering, the lidar is used 2 recorded point cloud in several object clusters C1 up to cm with object candidates K1 up to Km for rod-shaped objects 01 divided up to On. For example, a three-dimensional coordinate system (not shown in more detail) with the axes X, Y and Z, which form a sensor-oriented right-handed coordinate system, is used here. The X-axis is oriented in the longitudinal direction to the front, the Y-axis in the transverse direction to the left and the Z-axis in the vertical direction upwards. It is assumed that the lidar 2 so on the vehicle 1 is mounted so that the Z-axis is aligned at least substantially in the direction of gravity.

During the clustering, the entire captured point cloud is examined and all of the object candidates K1 belonging to km and as a rod-shaped object O1 Points marked to On are projected into a projection plane. In particular, a plane is selected as the projection plane which extends parallel to a floor plane or along the floor plane. The projection plane is spanned in particular by the X-axis and the Y-axis. Because the Z-axis is at least essentially aligned in the direction of gravity and most of the features of the rod-shaped objects 01 until On are also generally aligned in the direction of gravity, they form a rod-shaped object O1 to On corresponding points of a single object O1 to On typically a small, localized cluster of objects C1 up to Cm in the projection plane, which can easily be separated from other clusters. For this purpose, a kd-tree-based clustering algorithm is used, for example, two-dimensional neighborhood searches around each point with a fixed, parameterizable radius ε performs. All points that are in this radius ε belong to the same object cluster C1 up to cm.

An image B of a surroundings of a vehicle 1 , a lidar 2 as well as an object cluster C1 with an object candidate K1 shows during a parameter estimation performed after clustering 3 .

In the parameter estimation, a principal component analysis, also referred to as a principal component analysis, PCA for short, is used for each object candidate K1 up to Km a three-dimensional position and a radius R. certainly.

It becomes an object candidate K1 up to Km first through a complete three-dimensional point cloud W. of as rod-shaped objects 01 to On recognized points which relate to a specific object cluster C1 belong to Cm, shown. Then a center of gravity and a covariance matrix of the point cloud W. calculated. The covariance matrix is then diagonalized to perform the principal component analysis. An axis of symmetry of the rod-shaped objects 01 to On is typically the longest axis with the greatest eigenvalue, also referred to as the Z 'axis. The shortest axis with the smallest eigenvalue is typically the axis that is parallel to that of the lidar 2 emitted lidar beam is also referred to as the X'-axis. The remaining orthogonal axis is referred to as the Y 'axis.

When performing the parameter estimation, for example, the radius is determined R. of the respective rod-shaped object O1 until On. The radius R. scales linearly with an extent σ the point cloud W. of the object cluster C1 in the Y 'direction, whereby the free parameter can be estimated from real world data.

Because the lidar 2 only half of the respective rod-shaped object O1 until On can capture, the focus of the object cluster is C1 to Cm not on an axis of symmetry of the same. This is the case when calculating the three-dimensional position of a lower or upper point of the rod-shaped object 01 to consider until On becomes a midpoint M. of the respective rod-shaped object O1 until On determines which one is along the X 'axis from the lidar 2 away with a shift d proportional to the estimated radius R. of the rod-shaped object O1 until On is shifted.

$$\text{Y}\text{'}=\text{p}\_\left\{\text{radius}-\text{to}-\text{shift}\right\}*\text{R},$$

$$\text{Z}\text{'}=\text{min}\_\text{i}\left(\text{Z}\text{'}\_\text{i}\right).$$

This results in the coordinates X ', Y', Z 'for a lower point of the rod-shaped object O1 to On according to $$\text{X '}=\mathrm{0,}$$

$$\text{Y}\text{'}=\text{p}\_\left\{\text{radius}-\text{to}-\text{shift}\right\}*\text{R.},$$

$$\text{Z}\text{'}=\text{min}\_\text{i}\left(\text{Z}\text{'}\_\text{i}\right).$$

Here p_ {radius-to-shift} is the second and last free parameter, which can be estimated from the real distance data. The Z 'coordinate of the lower point is determined by taking the minimum of all Z' coordinates of the point cloud W. is determined and chosen. About the three-dimensional position of the rod-shaped object 01 to On, the determined coordinates X ', Y', Z 'of the coordinate system used in the principal component analysis are transformed into the coordinates X, Y, Z.

List of reference symbols

1

vehicle

2

Lidar

C1 to Cm

Object cluster

d

shift

K1 to Km

Object candidate

M.

Focus

O1 to On

object

R.

radius

W.

Point cloud

ε

radius

σ

expansion

QUOTES INCLUDED IN THE DESCRIPTION

This list of the documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent literature cited

• Landa et. al .: Detection of pole-like objects from LIDAR data; 19th International Conference Enterprise and Competitive Environment 2016, ECE 2016, 10-11 March 2016, Brno, Czech Republic [0003]

Claims (4)

Method for the detection of rod-shaped objects (01 to On) by means of lidar data, whereby a point cloud recorded by means of a lidar (2) is semantically labeled as an input variable by means of a convolutional neural network in which an object class for rod-shaped objects (O1 to On) is provided, characterized in that - in a clustering, the point cloud is divided into several object clusters (C1 to Cm) with object candidates (K1 to Km) for rod-shaped objects (01 to On) and - in a subsequent parameter estimation by means of a principal component analysis for each object candidate (K1 to Km) a three-dimensional position and a radius (R) can be determined. Procedure according to Claim 1 , characterized in that in the clustering, all of the points of the point cloud belonging to the object candidates (K1 to Km) are projected into a projection plane. Procedure according to Claim 2 , characterized in that a plane is selected as the projection plane which extends parallel to a floor plane or along the floor plane. Use of a method according to one of the preceding claims for vehicle self-localization or robot self-localization.

Images