DE102022000849A1 - Method for generating an environment representation for a vehicle - Google Patents

Abstract

The invention relates to a method for generating an environment representation for a vehicle (1) using data recorded by means of at least one environment sensor, the vehicle environment being divided into grid cells (G1 to Gn). According to the invention, distribution functions are determined for individual two-dimensional grid cells (G1 to Gn). Height information (HI1 to Hlm) is modeled using the distribution functions and occupancy probabilities are determined for each specific distribution function and reduced adaptively as a function of driving dynamics, a roadway profile and/or properties of detected environmental information.

Description

The invention relates to a method for generating an environment representation for a vehicle according to the preamble of claim 1.

From the DE 10 2019 203 274 B3 a computer-implemented method for creating an environment model for a vehicle that can be operated automatically is known. The surroundings of the vehicle are divided into cells and scanned with rays from a surroundings detection sensor, with the scanning of objects being stored in the associated cells. A maximum number of target points that can be stored and a maximum number of free space points that can be stored are specified for each cell. An average elevation, an elevation range, and a logit value for the average elevation are respectively stored for each of the samples in the target points and clearance points. The target points have a logit value greater than or equal to zero and the free space points have a logit value less than zero. A target point or a clearance point of a new scan is merged with a corresponding target point or clearance point of a previous scan if the respective height ranges overlap, or stored as a new target point or clearance point. If a new scan reaches its maximum number, two target points or two free space points from previous scans are merged and the new target point or free space point is stored in the respective cell.

Furthermore, from the DE 10 2013 210 263 A1 an occupancy map for a vehicle is known. The occupancy map has a plurality of cells arranged in a grid-like manner, with the cells of the occupancy map being adapted to the driving situation as a function of a driving situation of the vehicle. The cells are arranged in a Cartesian coordinate system or a polar coordinate system.

The invention is based on the object of specifying a novel method for generating an environment representation for a vehicle.

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

Advantageous configurations of the invention are the subject matter of the dependent claims.

In a method for generating an environment representation for a vehicle based on data recorded by means of at least one environment sensor, a vehicle environment is divided into grid cells.

According to the invention, distribution functions, for example normal distributions, are determined for individual two-dimensional grid cells. Height information is modeled on the basis of the distribution functions, with occupancy probabilities being determined for each specific distribution function and being reduced adaptively as a function of driving dynamics, a roadway profile and/or properties of detected environmental information.

Robust representations of a vehicle environment are required for the operation of driver assistance systems and automated, in particular highly automated or autonomous, vehicles. Known methods for a grid-based representation of the vehicle environment, for example a static vehicle environment, including for example buildings, curbs, crash barriers, etc., are either limited to two spatial dimensions or are very memory-intensive if height information is taken into account.

By means of the present method for environment representation, on the other hand, it is possible to reduce memory requirements in a profitable manner in comparison to existing methods with the same or improved performance. Due to the reduction in the storage requirement for displaying the height information, use on computer units of modern driver assistance systems is possible in a simple manner.

The environment representation generated by the method can be used for (self-) localization, trajectory planning, obstacle avoidance, creation of maps of the environment, control of vehicle actuators and/or as a basis for further algorithms, such as classification into semantic classes, e.g. curb, vehicle , house etc., can be used. In particular, the present method generalizes to different road topologies and allows a faster reaction to changed vehicle environments than known approaches. For example, the method is used to generate an efficient representation of a static environment of a vehicle in that the vehicle environment is detected by active sensors, for example radar, ultrasonic or lidar sensors, or passive sensors, for example a camera. In this case, the environment representation can first be combined by a sensor fusion unit with environment representations from other sensors to form an overall representation of the vehicle environment and then used for the aforementioned purposes. Results determined by means of the sensor fusion unit can, for example, be used by control units for operating the respective driver assistance system or the Vehicle will use them for trajectory planning, self-localization and generally for solving a driving task. The height information that flows into the representation enables decisions to be made about whether obstacles can be driven over or under.

In this case, the environment representation can be generated in a particularly advantageous manner without using very memory-intensive free space points.

Furthermore, the present method takes into account potentially moving objects in the vehicle environment and avoids long-lasting errors in an occupancy map and can be used independently of a selected coordinate system, for example a polar, Cartesian or any curvilinear coordinate system.

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

• 1 schematisch eine Belegungskarte mit einer Draufsicht eines Fahrzeugs und einer Fahrzeugumgebung,
• 2 schematisch eine Belegungskarte mit einer Draufsicht eines Fahrzeugs und einer Fahrzeugumgebung,
• 3 schematisch eine Belegungskarte mit einer Darstellung eines Fahrzeugs und einer Fahrzeugumgebung in einem Koordinatensystem mit Gitterzellen und Höheninformationen,
• 4 schematisch eine Belegungskarte einer Szene eines Fahrzeugs in einer Fahrzeugumgebung mit Gitterzellen und Höheninformationen sowie ein Kamerabild der Szene,
• 5 schematisch eine Belegungskarte einer weiteren Szene eines Fahrzeugs in einer Fahrzeugumgebung mit Gitterzellen und Höheninformationen sowie ein Kamerabild der Szene,
• 6 schematisch eine Belegungskarte eines fokussierten Bereichs der Szene gemäß 5 sowie das Kamerabild der Szene,
• 7 schematisch eine Belegungskarte einer Draufsicht der Szene gemäß 5 sowie das Kamerabild der Szene und
• 8 schematisch eine Belegungskarte bei einer Ansicht der Szene gemäß 5 aus einer Perspektive eines Fahrers des Fahrzeugs sowie das Kamerabild der Szene.

show:

• 1 schematically an occupancy map with a top view of a vehicle and a vehicle environment,
• 2 schematically an occupancy map with a top view of a vehicle and a vehicle environment,
• 3 schematically an occupancy map with a representation of a vehicle and a vehicle environment in a coordinate system with grid cells and height information,
• 4 schematically an occupancy map of a scene of a vehicle in a vehicle environment with grid cells and height information and a camera image of the scene,
• 5 schematically an occupancy map of another scene of a vehicle in a vehicle environment with grid cells and height information and a camera image of the scene,
• 6 schematically an occupancy map of a focused area of the scene according to FIG 5 as well as the camera image of the scene,
• 7 schematically an occupancy map according to a top view of the scene 5 as well as the camera image of the scene and
• 8th schematically an occupancy map in a view of the scene according to FIG 5 from a perspective of a driver of the vehicle and the camera image of the scene.

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

In 1 an occupancy map BK, also referred to as an occupancy grid, is shown schematically with a plan view of a vehicle 1 and a vehicle environment. The vehicle 1 is designed, for example, for automated operation, for example highly automated or autonomous driving operation.

Approaches for a representation of a static and a dynamic vehicle environment are required for a perception of the vehicle environment. The static vehicle environment includes, for example, curbs, fences, parked vehicles, in particular cars, trucks, buses, bicycles, motorcycles, etc., traffic lights, lanterns, crash barriers, house walls, bridges, tunnel walls, tunnel ceilings, buildings, etc. The dynamic vehicle environment is present, for example, moving subjects and / or in the 4 until 8th objects O1 to Oz shown in more detail, for example pedestrians and vehicles, are counted.

The method described below determines the environment representation for the static vehicle environment. Objects O1 to Oz of the static vehicle environment can potentially be dynamic objects O1 to Oz, for example parked vehicles or waiting vehicles at a red traffic light. However, as long as these objects 01 to 0z do not move, they are counted as part of the static vehicle environment. On the other hand, as soon as these objects O1 to Oz move, they belong to the dynamic vehicle environment.

There are generally so-called and in the 4 until 8th shown occupancy maps BK, also referred to as occupancy grid maps known. These occupancy maps BK have a large number of grid cells G1 to Gn. For example, an occupancy map BK with a size of 100 mx 100 m with a size of a grid cell G1 to Gn of 10 cm×10 cm can have a total of 1000×1000 grid cells G1 to Gn, ie one million grid cells G1 to Gn. In the simplest case, a two-dimensional spatial area around one's own vehicle 1 is covered by the occupancy map BK.

Each grid cell G1 to Gn of the occupancy map BK then corresponds to a small section of the real vehicle environment. In the simplest case, an occupancy probability is stored for each grid cell G1 to Gn, which is calculated from measurements carried out using environmental sensors of the vehicle 1 . In each time step in which the environmental sensors take measurements, measurement points become Updating the grid cells G1 to Gn used, which are in the vicinity of the respective measurement point. This creates a two-dimensional representation of the vehicle environment. Dynamic objects O1 to Oz are often not modeled, so that only the static environment is mapped.

This is exemplified by the representation according to FIG 1 , which shows an occupancy map BK of a top view of a scene in the area of a freeway entrance A in a highly simplified manner. Unhatched grid cells G1 to Gn in the areas B1 represent free space, single-hatched grid cells G1 to Gn in the areas B2 represent unknown occupancies and cross-hatched grid cells G1 to Gn in the areas B3 represent occupied spaces.

2 shows another occupancy map BK with a plan view of a vehicle 1 and a vehicle environment, with single-hatched grid cells G1 to Gn in areas B2 representing unknown occupancies and cross-hatched grid cells G1 to Gn in areas B3 representing occupied spaces.

In 3 an occupancy map BK with a vehicle 1 and a vehicle environment is shown in a coordinate system with grid cells G1 to Gn and height information HI1 to Hlm, with the height information HI1 to Hlm being plotted on a vertical axis z in the coordinate system as a function of a longitudinal axis x or a transverse axis y is.

In other words, in addition to simply specifying whether a grid cell G1 to Gn is occupied or not, there is also provision for modeling height information HI1 to Hlm. This is required in particular for the implementation of an automated driving function of the vehicle 1 in order to determine whether obstacles can be driven over and under.

In addition to an occupancy probability, a number of distribution functions, in particular normal distributions, are stored in each grid cell G1 to Gn as height information HI1 to Hlm.

Three values are calculated and saved for each distribution function. These values include a mean value, which indicates a mean position of the respective distribution function, and a standard deviation, which indicates a width of the respective distribution function and corresponds to the height of a respective object O1 to Oz or obstacle. Furthermore, the values include an existence probability, which indicates how certain it is that the distribution function actually exists and did not arise due to ghost measurements.

A possible exemplary embodiment of a method for generating an environment representation for a vehicle 1 with generation of an occupancy map BK with height information HI1 to Hlm is described below.

First, a selection is made as to whether the height information HI1 to Hlm is to be determined as a function of the longitudinal axis x or the transverse axis y of the vehicle 1 . The determination is carried out in the same way for both axes. The occupancy map BK is designed as a regular grid or has different resolutions, ie different sizes of grid cells G1 to Gn.

In the direction of the vertical axis z, no lattice is formed or there are no layers.

The distribution functions are not sorted into layers, but are defined by the data points themselves.

Each grid cell G1 to Gn has a number M of Gaussian distributions as elements N according to $$N\left({\mu }_i,{\sigma }_i\right)miti=\mathrm{1.2,}...,M.$$

These distributions indicate a probability for a height extension of objects 01 to Oz, not an occupancy probability of a grid cell G1 to Gn.

An adaptation of the occupancy map BK due to changes in a height coordinate of the vehicle 1 is not necessary since the distribution functions are parameterized in sequence coordinates and are therefore independent of a height position of the vehicle 1 .

As shown, the longitudinal axis x or the transverse axis y is represented on the horizontal axis of the coordinate system. It doesn't matter which of the two axes is displayed, since both are treated the same. Lines running vertically mark the individual grid cells G1 to Gn, which are the usual two-dimensional grid cells G1 to Gn in the x-y plane. Each of the grid cells G1 to Gn contains a series of Gaussian distributions which model the distribution functions of height values, ie the height information HI1 to Hlm, within the individual grid cells G1 to Gn. The number M of Gaussian distributions per grid cell G1 to Gn is an adjustable parameter, which can also be a function of the distance, for example.

In order to include newly detected objects O1 to Oz in the occupancy map BK, three steps are carried out: These include a two-dimensional integration, dealing with a height value in the direction of the height axis z and combining or merging the distributions.

A position of detection d _j is transformed into sequence coordinates and is given by X _seq , Y _seq , Z _seq .

The two-dimensional integration is almost identical to the creation of a two-dimensional occupancy map BK. Starting from a two-dimensional position of the object 01 to Oz, given by x _seq and y _seq , and a two-dimensional covariance matrix Σ _x,y , all relevant grid cells G1 to Gn in the two-dimensional occupancy map BK are selected and their respective likelihood ratio is calculated by integrating a two-dimensional Gaussian distribution centered at x _seq and y _seq with the covariance matrix over the
Grid cells G1 to Gn calculated. For each of these K cells that are relevant to the current detection d _j , the integration yields a probability value p _n (n=1, 2,...,K).

In the step in which the height value is handled in the direction of the height axis z, the following update step is carried out for each of these K cells. Depending on the number of existing distributions within a grid cell G1 to Gn and an exact position of the detection, different cases can be distinguished.

In a first case, no distributions exist in the respective grid cell G1 to Gn, or the height value of the detection in the direction of the height axis z is far removed from all distributions. This case can be defined by using z _seq of the detection and µ _i , σ _i of all M-distributions, for example by means of a so-called Mahalanobis distance or the like.

In this case, a new distribution is created: $$N\left({\mu }_i={\mathrm{e.g}}_{seq},{\sigma }_i={\sigma }_{\mathrm{e.g},seq}\right)$$

Thus, a new Gaussian distribution is created with mean z _seq and standard deviation σ _z,seq . Mean value and standard deviation are thus only used by the detection itself. In addition to the mean value and the standard deviation, the probability value p _n is also stored for this distribution. Instead of the probability value p _n log-odds p̃ can be stored. Log odds, also referred to as logit, are determined by converting a probability p (with values between 0 and 1) into a logarithmic scale using the calculation rule logodds=log(p/(1-p)). It calculates the logarithm of the odds (“odds”), where the odds are defined as p/(1-p).

In a second case, the detection is close to at least one existing distribution. If the detection is in the vicinity of at least one existing distribution, in particular a plurality of distributions, the distribution is selected which has the shortest distance to the detection. Again, a metric for the distance between detection and distribution needs to be developed. For example, the so-called Mahalanobis distance, a so-called Bhattacharyya distance or the so-called 3-sigma range can be used.

The task now is to update internal parameters of the distribution. It is thereby assumed that the distribution N with the parameters μ _i , σ _i and p̃ was selected as the next distribution for detection. The following two methods do not take into account the standard deviation of detection, but only use the measured z-position for the update.

und $${\sigma }_i\to {\tilde{\sigma }}_i=\frac{n_i\cdot {\sigma }_i+\left(z_{seq}-{\mu }_i\right)\cdot \left(z_{seq}-{\tilde{\mu }}_i\right)}{n_i+1}$$

In a first method, a number of the proofs ni, which was previously used to form the distribution, is used to weight a new or old mean and the variance according to: $${\mu }_i\to {\tilde{\mu }}_i=\frac{n_i\cdot {\mu }_i+{\mathrm{e.g}}_{seq}}{n_i+1}$$

and $${\sigma }_i\to {\tilde{\sigma }}_i=\frac{n_i\cdot {\sigma }_i+\left({\mathrm{e.g}}_{seq}-{\mu }_i\right)\cdot \left({\mathrm{e.g}}_{seq}-{\tilde{\mu }}_i\right)}{n_i+1}$$

In a second method, instead of using the number of detections, the log-odd values are used as weights. The standard deviations reported by the respective environmental sensor are also taken into account.

The standard deviations of detections give information from the environment sensor about the certainty that the detection is really at the reported z-position. However, the standard deviations of the distributions are a measure of the height of an object O1 to Oz from which the detections originate. Therefore, the standard deviation of detection and distribution are not mixed. That is, interpretations of the standard deviations vary widely.

$$w_2=\frac{{\tilde{\rho }}_i}{{\sigma }_i}$$

To use the standard deviation as another weight of a detection, it is assumed that the standard deviation of a detection gives a measure of the certainty that the detection is in the reported position. A detection with a low standard deviation, i.e. high confidence that the reported position is correct, is therefore treated with a higher impact on the distribution than a detection with a high standard deviation. If only the occupancy values were used for weighting, two detections with the same x/y/z position but very different standard deviations in the z direction would affect the mean value of the resulting distribution in the same way. This is certainly not desirable, since a greater influence of those detections that have small standard deviations is desired. For example, instead of simply performing the merging between detection and distribution using the weights constructed from the occupancy values, the following values can be used as weights, for example: $${\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1=\frac{{\tilde{\rho }}_n}{{\widehat{\sigma }}_{\mathrm{e.g}}}$$

$$w_2=\frac{{\tilde{\rho }}_i}{{\sigma }_i}$$

$${\sigma }_i^2\to {\tilde{\sigma }}_{}^2{}_i=\frac{w_2\cdot \left({\sigma }_i^2+{\mu }_i^2\right)+w_1{z}_{seq}^2}{w_1+w_2}-{\tilde{\mu }}_i^2$$

$${\tilde{\rho }}_i\to {\tilde{\rho }}_i+{\tilde{\rho }}_n$$

Here σ̂ _z is the transformed standard deviation for the detection in direction z _seq . In an update step, the weights are used as follows: $${\mu }_i\to {\tilde{\mu }}_i=\frac{w_2\cdot {\mu }_i+{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1\cdot {\mathrm{e.g}}_{se\mathrm{<figure-callout\; id="1"\; label="q\; w"\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">q</figure-callout>}}}{w_{}}$$

$${\sigma }_i^2\to {\tilde{\sigma }}_{}^2{}_i=\frac{w_2\cdot \left({\sigma }_i^2+{\mu }_i^2\right)+{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1{\mathrm{e.g}}_{seq}^2}{{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1+w_2}-{\tilde{\mu }}_i^2$$

$${\tilde{\rho }}_i\to {\tilde{\rho }}_i+{\tilde{\rho }}_n$$

An update sequence must be observed. It is assumed that multiple detections exist in the same grid cell G1 to Gn. In this context, different strategies can be applied to update the distributions.

For example, all detections can be iterated over and the distributions are updated immediately. This implies that a second detection is checked against different distributions than the first. The order in which the detections are inserted plays a role here and can change the result. However, this is not desired.

In order to obtain deterministic results that should not depend on the order of insertions, it is first iterated over all detections and then it is determined which detections are merged into a distribution based on old parameters of the distribution. After that, the update is carried out. A sorting order of the detections does not matter here to determine which detections are merged. However, it may be that detections that were initially discarded for merging would now be eligible for merging with the updated distribution. It does not matter whether a first detection is first used to update the distribution and then a second detection is used to update the new distribution, or vice versa. The resulting mean and standard deviations are the same.

This strategy may be more beneficial for getting consistent results. On the other hand, a recursive update to merge detections that were not considered in the first step can lead to a difficult estimation of a worst-case behavior.

In a third case, all distributions are in use and the detection does not belong to any of them.

If all M distributions are already in use and the new detection does not belong to any of them, there are two possibilities. According to a first possibility, the new detection is closer to the vehicle 1 in the z-direction than any other distribution. This detection is then considered more relevant and is therefore chosen as the basis of a new distribution. The distribution farthest from z _ego is removed to make room for the new one. In a second possibility, a second detection is far away from all distributions and also farther away from vehicle 1 than all existing distributions. The detection is then discarded. In scenarios with large changes in road grade, such as uphill or downhill, other methods can be used.

After a distribution has been updated within a grid cell G1 to Gn, the merging of the distributions can take place. If, for example, two distributions have moved closer together due to the insertion of the new detection and are very close together, then they can be combined to form a common Gaussian distribution. Again, the merging can be done using the log odds for weighting. To define the distance between two very closely spaced distributions, one is required mathematical definition, for example by using the Bhattacharyya distance.

$${\widehat{\sigma }}^2=\frac{{\tilde{\rho }}_1\left({\sigma }_1^2+{\mu }_1^2\right)+{\tilde{\rho }}_2\left({\sigma }_2^2+{\mu }_2^2\right)}{{\tilde{\rho }}_1+{\tilde{\rho }}_2}-{\widehat{\mu }}^2$$

$$\tilde{\rho }={\tilde{\rho }}_1+{\tilde{\rho }}_2$$

The merging or merging takes place, for example, according to: $$\widehat{\mu }=\frac{{\tilde{\rho }}_1{\mathrm{<figure-callout\; id="1"\; label="\mu "\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">\mu </figure-callout>}}_1+{\tilde{\rho }}_2{\mu }_2}{{\tilde{\rho }}_1+{\tilde{\rho }}_2}$$

$${\widehat{\sigma }}^2=\frac{{\tilde{\rho }}_1\left({\mathrm{<figure-callout\; id="1"\; label="\sigma "\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}_1^2+{\mathrm{<figure-callout\; id="1"\; label="\mu "\; filenames="DE102022000849A1\_0015.png"\; state="\{\{state\}\}">\mu </figure-callout>}}_1^2\right)+{\tilde{\rho }}_2\left({\sigma }_2^2+{\mu }_2^2\right)}{{\tilde{\rho }}_1+{\tilde{\rho }}_2}-{\widehat{\mu }}^2$$

$$\tilde{\rho }={\tilde{\rho }}_1+{\tilde{\rho }}_2$$

Furthermore, distributions can also be divided into two distributions each with a smaller standard deviation if the standard deviation of one of the distributions exceeds a predetermined limit value.

After the distributions are created and updated if necessary, elevation values are extracted for each distribution. For this purpose, it is taken into account how the standard deviation of the distributions and the height of the true objects O1 to Oz are related. In particular, there is not just one constant factor that links the standard deviation and the height values. For example, it is possible that a different factor is used for smaller objects O1 to Oz than for larger ones. That is, for distributions with smaller standard deviation, a smaller factor is used to calculate a true height of the object O1 to Oz compared to distributions with larger standard deviation. However, the reported standard deviation values for the detections themselves influence the distributions, at least via the weights, so that a distance dependency may also have to be integrated.

If fading is to be used, then the stored ρ̃ values of each Gaussian can be reduced. This has the same effect as reducing the log-odds in a normal two-dimensional occupancy map BK.

aus der Summe aller p̃.A total occupancy of a grid cell G1 to Gn results according to $$\widehat{\rho }={\displaystyle \sum _{i=1}^M{\tilde{\rho }}_i}$$

from the sum of all p̃.

Here, too, M is the number of (active) distributions in the grid cell G1 to Gn. This provides a direct way to compress an elevation raster into a simple occupancy raster. Here p̂ is calculated for each grid cell G1 to Gn and reported as the occupancy value of the grid cell G1 to Gn. This value is identical to the case where height is not taken into account at all, i.e. identical to a standard two-dimensional occupancy map BK with the same data and neglected z-component.

The Gaussian distributions are stored for each grid cell G1 to Gn. Each Gaussian distribution has three parameters. The parameters are an average, a variance and the number of detections or occupancies. If the map is used for weighting and updating, a separate two-dimensional map layer is not needed. Thus, N different values are stored. N is obtained according to $$N=M\times 3\times N_x\times N_y$$

In this case, N _x and N _y are a respective number of grid cells G1 to Gn in the direction of the longitudinal axis x and in the direction of the transverse axis y.

The method described above makes it possible to reduce the memory requirement since, in contrast to the methods known from the prior art, free space cells can be dispensed with. Furthermore, in the present method, in contrast to known methods, logits originate from the integration in the x, y direction. As a result, the x,y position of a detection is included in an update step of a grid cell G1 to Gn. A detection thus has a higher weight when it is close to the grid cell G1 to Gn in which the distributions are being updated, and a lower weight when it is further away.

In addition, the measurement uncertainty of the detection is included in the calculation. Furthermore, the method ensures that measurement uncertainties of the detections are correctly taken into account and that the standard deviation of the distribution approximates the actual extent of the objects O1 to Oz. With the Bhattacharyya distance, a suitable measure is also used to decide which distributions within a grid cell G1 to Gn are updated by a new measurement point, ie a new detection.

• - Verteilungen, die durch falsche Messpunkte erzeugt wurden, werden in nachfolgenden Schritten automatisch weniger gewichtet, da ihre Existenzwahrscheinlichkeit reduziert wird.
• - Erscheint eine Gitterzelle G1 bis Gn als belegt, das heißt enthält sie mindestens eine Verteilung mit hoher Existenzwahrscheinlichkeit, und ist diese Belegung durch ein potenziell bewegtes Objekt O1 bis Oz entstanden, beispielsweise durch ein stehendes Auto, und beginnt sich das Objekt O1 bis Oz zu bewegen, so sollte die Existenzwahrscheinlichkeit der passenden Verteilungsfunktion reduziert werden. Durch adaptives Fading kann die Existenzwahrscheinlichkeit von denjenigen Verteilungsfunktionen reduziert werden, welche in Gitterzellen G1 bis Gn liegen, in denen auch ein dynamisches Objekt O1 bis Oz gemessen wird. Beispielsweise kann mit einem Radarsensor oder einem so genannten FMCW-Lidar nicht nur die Position, sondern auch eine Dopplergeschwindigkeit, das heißt eine radiale Geschwindigkeit, eines Objekts O1 bis Oz gemessen werden. Somit kann in denjenigen Gitterzellen G1 bis Gn, in denen eine Detektion mit einen Schwellwert übersteigenden Dopplergeschwindigkeit gemessen wurde, die Existenzwahrscheinlichkeit der passenden Verteilungsfunktion stärker reduziert werden als in Gitterzellen G1 bis Gn, in denen keine dynamische Detektion gemessen wurde. Alternativ können auch Informationen aus einem System zum Tracken von dynamischen Objekten O1 bis Oz genutzt werden, um an den Orten, wo sich laut dem anderen System dynamische Objekte 01 bis Oz befinden, die Existenzwahrscheinlichkeiten zu reduzieren.
• - Das Fading kann adaptiv an die Fahrsituation angepasst werden. Gibt es beispielsweise starke Änderungen im Neigungswinkel, so können gezielt Existenzwahrscheinlichkeiten von denjenigen Verteilungen reduziert werden, welche basierend auf dem aktuellen Neigungswinkel weit oberhalb oder weit unterhalb des Fahrzeugs 1 liegen.

Furthermore, the probability of existence of a distribution is reduced by the fading in each time step. This has three positive effects/areas of application:

• - Distributions that were generated by incorrect measurement points are automatically weighted less in subsequent steps because their probability of existence is reduced.
• - If a grid cell G1 to Gn appears to be occupied, i.e. it contains at least one distribution with a high probability of existence, and this occupancy is caused by a potentially moving object O1 to Oz, for example a stationary car, and the object O1 to Oz begins to move move, the probability of existence of the appropriate distribution function should be reduced. Adaptive fading can reduce the probability of existence of those distribution functions that lie in grid cells G1 to Gn in which a dynamic object O1 to Oz is also measured. For example, not only the position but also a Doppler speed, ie a radial speed, of an object O1 to Oz can be measured with a radar sensor or a so-called FMCW lidar. Thus, in those grid cells G1 to Gn in which a detection with a Doppler velocity exceeding a threshold value was measured, the probability of existence of the appropriate distribution function can be reduced more than in grid cells G1 to Gn in which no dynamic detection was measured. Alternatively, information from one system for tracking dynamic objects O1 to Oz can also be used in order to reduce the existence probabilities at the locations where dynamic objects 01 to Oz are located according to the other system.
• - The fading can be adapted to the driving situation. If, for example, there are major changes in the angle of inclination, the probability of existence of those distributions that are far above or far below vehicle 1 based on the current angle of inclination can be reduced in a targeted manner.

Already existing distributions can become “free” again through fading. This is the case, for example, when the probability of existence falls below a threshold value, so that in the next cycle the memory area of this distribution function can be used to generate a new distribution function from new measurement points. Also, by using fading, no use of free space points is required.

The method also means that there is no need to be restricted to a Cartesian coordinate system.

A current vehicle pose depending on a road topology can also be included in the update step. In doing so, taking into account the angle of inclination, it is checked which of the existing distributions are still relevant in the future. Those distributions are classified as less relevant which, based on the current road gradient, will be at a greater distance from vehicle 1 in the future. The distribution with the highest distance is then overwritten by a new distribution generated from a new measurement point. In particular, this only happens when the new measuring point is also closer to vehicle 1 than the distribution function then overwritten. As a variation, high-precision map material could also be used to estimate the future course of the road and thus make the decision on the relevance of the distributions more robust. This avoids marking areas as occupied that are actually only framed by two obstacles, such as a curb on the ground under a bridge above. This is also avoided when a new measurement arrives and does not match the existing distributions, such as a pedestrian on the bridge.

A possibility can also be provided for dividing a very broad distribution function into two distribution functions with a smaller standard deviation. This can be useful, for example, when two objects O1 to Oz exist at different heights, two distribution functions initially to describe the
Objects O1 to Oz are generated and these are connected to a large distribution by incorrect or inaccurate measurements in one cycle. Such objects O1 to Oz can be, for example, curbs close to the ground and a bridge at a height of 5 m.

To avoid that this error cannot be corrected later, merging of distributions is not allowed or distributions with a certain width are split after a few cycles.

Furthermore, the method makes it possible to use a grid in which the
Grid cells G1 to Gn are of different sizes depending on the driving situation. A larger cell size could be used at a greater distance, the cell size could be adapted to a speed of vehicle 1 and/or the cell size could be adaptively adapted to other driving maneuvers, for example cornering. Unlike when using a two-dimensional occupancy map BK, the two-dimensional grid including the distribution functions is adjusted here. For example, a different number M of distributions per grid cell G1 to Gn depending on the distance voltage range, speed, etc. can be used.

The method also allows a position of the vehicle 1 not to be in the middle of the occupancy map BK. It is conceivable that the position of the vehicle 1 is moved further to the edge of the occupancy map BK depending on the speed. Thus, the vehicle 1 is in the middle of the occupancy map BK for low speeds, but close to the lower edge of the occupancy map BK for high speeds, so that the largest possible area in the direction of travel is covered.

In 4 an occupancy map BK of a scene of a vehicle 1 in a vehicle environment with grid cells G1 to Gn and height information HI1 to Hlm generated according to the described method or embodiments thereof, as well as an image P of the scene are shown. Different hatchings represent different height values of objects O1 to Oz shown in image P.

the 5 until 8th show occupancy maps BK of a scene of a vehicle 1 shown in an image P from different perspectives, with an object O1 being an overhanging tree and an object Oz being a bridge.

while showing 6 an occupancy map BK of an area focused on the bridge according to the scene 5 , 7 an occupancy map BK according to a plan view of the bridge of the scene 5 such as 8th an occupancy map BK in a view of the scene according to 5 from a perspective of a driver of the vehicle 1.

Also in the 5 until 8th different hatchings represent different height values of objects O1 to Oz shown in the image P.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited 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.

Patent Literature Cited

• DE 102019203274 B3 [0002]
• DE 102013210263 A1 [0003]

Claims (3)

Method for generating an environment representation for a vehicle (1) using data recorded by means of at least one environment sensor, the vehicle environment being divided into grid cells (G1 to Gn), characterized in that - distribution functions are determined for individual two-dimensional grid cells (G1 to Gn). - is modeled using the distribution functions height information (HI1 to Hlm) and - for each specific distribution function occupancy probabilities are determined and reduced adaptively depending on driving dynamics, a roadway profile and/or properties of detected environmental information. procedure after claim 1 , characterized in that normal distributions are determined as distribution functions. procedure after claim 1 or 2 , characterized in that in addition to the occupancy probabilities for each specific distribution function - a mean value, which indicates a mean position of the respective distribution function, and - a standard deviation, which indicates a width of the respective distribution function and with a height of a value in the associated grid cell (G1 to Gn) located object (O1 to Oz) corresponds to be determined.

Images