EVALUATION OF LOCATION MEASUREMENTS FROM AN ENVIRONMENTAL SENSOR FOR A MOTOR VEHICLE

Description

[0001] The invention relates to a method for evaluating location measurements from an environmental sensor for a motor vehicle, in which location measurements and associated measurement uncertainties of the location measurements are obtained from the environmental sensor.

[0002] In driver assistance systems for motor vehicles, for example in automatic distance control systems or collision warning systems, radar sensors are frequently used to detect the traffic environment.

[0003] Recently, there has been increasing interest in radar sensors with ever higher angular resolution and angular separation capability, which can detect a large number of radar reflections in an extended object depending on its size and distance.

[0004] Roos et al., "Reliable Orientation Estimation of Vehicles in High-Resolution Radar Images", IEEE Transactions on Microwave Theory and Techniques 64.9 (2016), 2986-2993, describe a clustering algorithm that assigns radar measurements to an object independently of any prior knowledge about the object. However, this can lead to jumps in object parameters such as orientation or size.

[0005] However, assigning each individual measurement associated with an object to a corresponding point in a detailed object model would require a high level of model knowledge regarding object type and object extent, which is not always reliably available.

[0006] DE 102006019474 A1 describes an electrical system with a main system that has a main filter configured as a probability filter, several subsystems connected to the main system, each comprising at least one data source, wherein data can be supplied to the main system from the subsystems, and wherein each subsystem has a subfilter configured as a probability filter, and wherein data can be supplied to the subsystems from the main system. In one example, a first of the subsystems comprises a relative sensor array as a data source, a second of the subsystems comprises several transponders as data sources, and the main filter and the subfilters comprise unscented particle filters.

[0007] DE 10 2011 017 323 A1 describes a method for determining the probability of a collision between a vehicle and an object within the vehicle itself, in which, starting from a state vector based on measured values ​​that describes the current position and motion state of the vehicle and a detected object, a minimum object distance between the vehicle and the detected object expected within a predictive time, as well as the statistical distribution of the values ​​of the minimum object distance, are predictively determined.

[0008] US 2014 / 0072233 A1 concerns updating a predicted location of an object in a multidimensional space.

[0009] WO 2018 / 085107 A1 describes a method for measuring distances when forming a vehicle convoy. It describes the identification of the rear of a specific vehicle and the tracking of that rear.

[0010] The object of the invention is to provide a method for evaluating location measurements of an environmental sensor for a motor vehicle, which makes it possible to efficiently improve the state estimation of an object.

[0011] This problem is solved according to the invention with the features specified in the independent claims. Advantageous further developments and embodiments of the invention are specified in the dependent claims. The environmental sensor is a radar sensor.

[0012] Estimating the object's current state parameters is achieved by adjusting these parameters to the location measurements associated with the object. Estimating current state parameters while considering weights can also be referred to as weighted estimation, weighted fitting, or weighted fitting.

[0013] The core idea of ​​the solution is that for new measurements associated with an existing object, an association probability is determined, and these association probabilities are incorporated into weights for a weighted fit of the object's model parameters (state parameters). For example, the fit can be performed for a line segment describing an edge of an object, using the state parameters position, orientation, and minimum and maximum values ​​of a valid range (endpoints of the line segment), thus re-estimating the line segment.

[0014] By incorporating multiple measurements into a parameter estimation of an object's state parameters, the data volume for a single object can be reduced, enabling a more robust description of a large number of objects. Simultaneously, the state estimator can iteratively update the state of objects. Compared to the conventional assignment of individual measurements to defined points in a model, this results in greater robustness. In particular, weighting the parameter adjustments of the state parameters can increase the robustness of the method against systematic deviations between localization measurements and the model that are not included in the model. Examples of such systematic deviations include, for instance, systematic variations in the radial velocity of reflections from rotating wheels, or deviations of real object edge profiles from a modeled straight line.

[0015] A positioning measurement includes measurements of the distance r, the radial velocity vr, and / or an object angle phi, for example, an azimuth angle. With a radar sensor, the positioning measurements correspond to radar reflections from an object.

[0016] An object state can, for example, include state parameters of a rectangular box or edge in a neighborhood of the environment sensor, and is described by a Cartesian position, extent, and / or velocity in Cartesian coordinates.

[0017] The fact that a determination or calculation depends on a quantity means here that the quantity is an input variable of the determination or calculation, and that the result of the determination or calculation depends on it. In particular, an input variable of the calculation comprises different values, for which different results are obtained.

[0018] The association probability is the probability that the location measurement accurately represents the location of a real object to which the estimated object state refers. The association probability thus indicates the probability that the location measurement in question belongs to the object described by the estimated object state. In other words, the association probability indicates the probability that the location measurement was obtained from the real object described by the estimated object state.

[0019] When determining the association probability, the estimated object state is taken into account, whereby increasing association probabilities are determined for increasing spatial similarities of a measured location of a positioning measurement with the estimated object state, and / or whereby increasing association probabilities are determined for increasing similarities of a measured velocity of a positioning measurement with the estimated object state.

[0020] Different association probabilities are accounted for by different weights. The association probability is determined as one of several values ​​that encompass multiple intermediate values ​​between a lower bound (e.g., zero) and an upper bound (e.g., one). For example, the association probability can be determined from a set of discrete values ​​or from a range of values.

[0021] When adjusting the state parameters to the location measurements associated with the object, the weighting of each location measurement depends on the specific probability of association with that particular measurement. Increasing association probabilities result in higher weightings. Furthermore, these association probabilities are used as weightings for the location measurements when adjusting the state parameters to the location measurements associated with the object.

[0022] In one or more embodiments, when determining the association probability for associating a location measurement with the object, the association probability is determined taking into account the measurement uncertainty of the location measurement. For example, the measurement uncertainty of the location measurement can be taken into account by reducing the relevant association probability, with larger reductions possible for increasing measurement uncertainties.

[0023] In one or more embodiments, when determining the association probability for associating a location measurement with the object, the association probability is determined taking into account an uncertainty in the estimated object state. For example, the uncertainty in the estimated object state can be taken into account by reducing the relevant association probability, with larger reductions possible for increasing uncertainties in the estimated object state.

[0024] In one or more embodiments, the association probability for associating a location measurement with the object is determined taking into account a model uncertainty of the estimated object state. For example, the model uncertainty can correspond to an expected deviation of real objects from an underlying model. In particular, the model uncertainty can include, for example, a shape uncertainty. For instance, in the case of an object based on a geometric model, the model uncertainty can describe the deviations from the model expected due to the real shapes of cars and trucks differing from the geometric model.For example, model uncertainty for an object based on a model of a rectangular box (or a straight edge) can describe the deviations from the model that are expected due to the real shapes of cars and trucks, which differ from the box shape (or the shape of a straight edge).

[0025] Preferably, in the step of associating received location measurements with the object described by an estimated object state, a location measurement is only associated with the object described by the estimated object state if the association probability reaches a minimum value. This is also known as gating. Thus, the location measurements on which the estimation of the object's current state parameters is based are included in the estimation with respective weights, where the respective weight depends on the specified association probability. Above the minimum value, different association probabilities are taken into account through different weights. Therefore, it is not simply a matter of filtering out location measurements and giving equal consideration to the remaining location measurements.

[0026] Associating received location measurements with an object described by an estimated object state can include associating received location measurements with one or more sub-objects of an object described by an estimated object state. An association with a sub-object of an object also constitutes an association with the object itself.

[0027] The current state parameters estimated in the step of estimating the object's current state parameters can be, for example, state parameters of one or more sub-objects of the object. For instance, the state parameters of a side edge of a rectangular box can be estimated based on location measurements associated with that side edge.

[0028] The method preferably comprises: updating the estimated object state by the state estimator, based on the object's previous estimated state and the object's estimated current state parameters. This update can thus be iterative. Particularly preferably, the method comprises: generating a prediction of a new object state by the state estimator based on the object's previous estimated state; and correcting the prediction of the new object state by the state estimator based on the object's estimated current state parameters.

[0029] For the iterative updating of the estimated object state, uncertainties of the estimated current state parameters of the object can also be taken into account, which are estimated based on the measurement uncertainties of the measurements.

[0030] The state estimator is preferably configured to update the estimated state of the object based on the estimated current state parameters of the object. In other words, the state estimator is configured to determine a new state of the object based on the previous estimated state and the estimated current state parameters of the object. The state estimator is particularly preferably configured to generate a prediction of a new state of the object based on a previous estimated state and to correct the prediction using the estimated current state parameters of the object to update the state. It could, for example, be a Kalman filter, an extended Kalman filter, an unscented Kalman filter, or other well-known state estimators.

[0031] The method preferably comprises the following steps: obtaining measurement uncertainties of the positioning measurements from the environmental sensor; and Estimating uncertainties of the estimated current state parameters of the object based on the obtained measurement uncertainties of the location measurements associated with the object, whereby the estimated current state parameters of the object and the estimated uncertainties are passed to the state estimator to update the estimated object state of the object.

[0032] The procedure can, for example, further include: updating an estimated uncertainty of the estimated object state of the object by the state estimator.

[0033] According to a further development of the invention, the uncertainties of the fitted model parameters are determined using an unscented transform (UT).

[0034] For example, estimating uncertainties of the estimated current state parameters of the object can be done using an unscented transformation, comprising the following steps: Calculating sigma points of the unscented transformation, wherein a vector comprising the position measurements associated with the object is used as a mean estimate for a distribution of the sigma points to be calculated, and wherein a matrix comprising the measurement uncertainties of the position measurements associated with the object is used as a covariance matrix for the distribution of the sigma points to be calculated; for each sigma point of the unscented transformation, estimating current state parameters of the object, comprising fitting the state parameters to the respective sigma point; determining a variance of a distribution of the current state parameters of the object estimated for the sigma points as an estimate of the uncertainties of the estimated current state parameters of the object.

[0035] The input for the unscented transformation is a vector of position measurements associated with the object, for which the uncertainty can be determined by the measurement uncertainties of the position measurements and, if applicable, their covariances. The sigma points are chosen according to the unscented transformation procedure such that their distribution exhibits a mean and a covariance that correspond to a given mean in the form of the vector of position measurements associated with the object and a given covariance in the form of the matrix encompassing the measurement uncertainties of the position measurements associated with the object. Thus, the distribution of sigma points and its covariance represent the position measurements associated with the object and their measurement uncertainties.

[0036] The sigma points are transformed by estimating the object's current state parameters for each sigma point. Each sigma point is thus transformed into an estimated current state parameter of the object.

[0037] The (co-)variance of the transformed sigma points then gives an estimated uncertainty of the estimated current state parameters of the object.

[0038] The unscented transformation thus makes it possible to obtain an estimate of the uncertainties of the estimated current state parameters by transforming the sigma points into respective estimates of the current state parameters of the object.

[0039] Thus, the unscented transformation can provide an estimate of the uncertainties of the estimated current state parameters of the object. This is particularly advantageous because parameter fitting is usually a non-linear mapping, and measurement uncertainties would be distorted or biased by such a mapping. Even when transforming radial coordinates from position measurements into Cartesian model coordinates of an object, improved uncertainty estimates can be obtained by determining the uncertainties using the unscented transformation.

[0040] Preferably, when adjusting the state parameters to the respective sigma point, weights for the sigma point are taken into account, which weights depend on the covariance matrix.

[0041] For example, when adjusting the state parameters to the respective sigma point, the same weights can be used as when adjusting the state parameters to the location measurements associated with the object (in the step of estimating the object's current state parameters).

[0042] The invention also relates to a sensor system with an environment sensor for motor vehicles, in which one of the methods described above is implemented.

[0043] The following section explains an exemplary embodiment in more detail with reference to the drawing.

[0044] They show: Fig. 1 a schematic representation of the surroundings of a motor vehicle; Fig. 2 a schematic diagram of a sensor system for a motor vehicle with an environmental sensor; and Fig. 3 a schematic flowchart of an unscented transformation for determining uncertainties of estimated current state parameters of an object.

[0045] Fig. 1 Figure 1 schematically shows the environment of a motor vehicle's ambient sensor with several location detections within the sensor's detection range, subsequently referred to as location measurements 10 and 12. Also shown are objects 14, representing real-world objects in the form of motor vehicles and modeled as rectangular boxes. Each object in the model is assigned a position in Cartesian coordinates X, Y, an extent, an orientation, and a velocity in the form of a velocity vector 16. A single location measurement 10, 12, for example, includes a radial distance r, a radial relative velocity vr, and a direction angle in the form of an azimuth angle phi, each relative to the self-position of the ambient sensor or the vehicle itself at x = 0, y = 0.

[0046] In Fig. 1 Location measurements 10 that were associated with an object 14 are represented by filled symbols. Location measurements 12 that were not associated with any of the objects 14 are represented by an open symbol.

[0047] Fig. 2 Figure 1 schematically shows a sensor system with an environment sensor 20 in the form of a radar sensor with an antenna system 22. Figure 22 schematically shows a location measurement 10 corresponding to a located radar reflection.

[0048] An evaluation unit of the sensor system comprises an association unit 24 for associating location measurements 10 with an object 14, an adaptation unit 26 for performing a fit to adapt state parameters to location measurements 10 associated with an object 14, an unscented transformation unit 28 for estimating uncertainties of the adapted current state parameters, and a state estimator 30 for iteratively updating an estimated object state of the object 14.

[0049] The association unit 24 comprises a first unit 32 for calculating association probabilities and a second unit 34 for selecting and associating location measurements 10 to an object 14.

[0050] Location measurements 10, 12 are obtained from the environmental sensor 20 with, for example, the parameters r, vr, phi and associated measurement uncertainties Δr, Δvr, Δphi.

[0051] Initially, objects 14 are initialized in a known manner via an independent mechanism, such as clustering location measurements 10, 12 not previously associated with objects 14 and correspondingly initializing the object states of the newly created objects 14.

[0052] The first unit 32 calculates association probabilities p for each object 14 for associating the respective location measurements 10, 12 with the object 14 in question. The second unit 34 selects, for each object 14, the location measurements 10 to be associated with the object 14 based on their association probability p and associates them with the object 14 in question. Location measurements 12 that can only be associated with an object 14 with a low association probability below a minimum value pmin are filtered out, i.e., they are not associated. This is also called gating. In the Fig. 1 The representation shown depicts the location measurements associated with the respective objects (14) with filled symbols.

[0053] In the example shown, the Fig. 2 The association for a sub-object 36 of object 14 in the form of a front edge of object 14 is shown as an example.

[0054] Determining the association probability p for the association of a location measurement 10 to an object 14 (or the sub-object 36) can, for example, be done taking into account: the measurement uncertainty Δr, Δvr, Δphi, for example in the form of a covariance in the measurement parameter space; a prediction uncertainty of object 14 with respect to the parameters of the predicted object state of object 14 such as position, velocity, extent, orientation; and / or a model uncertainty, for example in the form of a shape uncertainty of object 14.

[0055] For example, as a form uncertainty of a sub-object 36 in the form of an edge or side edge of an object 14, a deviation in the range of + / -30 cm of individual positioning measurements 10 from an exactly straight course of the edge of the model of the sub-object 36 can be taken into account.

[0056] Based on the determined association probabilities p, a parameter estimation in the form of a weighted model fit is then performed for sub-object 36 or object 14. For example, fitting unit 26 performs a weighted straight-line fit in the form of a weighted least-squares straight-line fit or a weighted principal component analysis. The input variables for the fit in this case are the positions of the location measurements 10. A range of validity for the fitted straight line, corresponding to the extent of sub-object 36, can also be estimated. A weighted least-squares model fit can also be performed to estimate a current state parameter in the form of a velocity vx, vy of object 14. For example, assuming a yaw rate of zero for object 14, fitting unit 26 can estimate the current state parameters vx and vy in the model. vr = vx * cos phi + vy * sin phi treasure.

[0057] The current state parameters of an object 14 (or its sub-objects 36) estimated by the adaptation unit 26 are passed to the state estimator 30.

[0058] To provide the state estimator 30 with the uncertainties associated with the estimated current state parameters of an object 14, the uncertainties are estimated as follows by performing an unscented transformation. Since both the straight-line fit and the fit using the velocity equation vr = vx * cos(phi) + vy * sin(phi) involve a nonlinear relationship, an unscented transformation is used.

[0059] All position measurements 10 associated with an object 14 are combined into a single vector Y. To do this, the individual position measurements 10, each with a measurement vector dimension n, and of which I position measurements 10 are available, are combined into a vector of length N = n * I. A covariance matrix K of the combined measurement vector Y is generated, for example, by combining the covariances of the individual position measurements 10 in a block diagonal matrix. The covariance matrix has a size of N x N. If covariances between the position measurements are known, they can be recorded in subsidiary block diagonals.

[0060] The implementation of the unscented transformation (UT) is in Fig. 3This is presented in the form of a schematic flowchart. In step S10, sigma points are calculated based on the combined measurement vector Y and its combined covariance K, as is known for unscented transformations. For a total measurement vector Y of dimension N, 2N + 1 sigma points are calculated. For each of the sigma points, the weighted model fit (step S12) is performed. This corresponds to the fitting performed by the fitting unit 26. The resulting 2N + 1 fitting results represent a distribution whose variance (covariance matrix) is an estimate of the uncertainties of the current state parameters of object 14 as estimated by the fitting unit 26. The variance (covariance matrix) is accordingly calculated from the distribution of the 2N + 1 results (step S14). It is passed to the state estimator 30 together with the estimated current state parameters.

[0061] The state estimator 30 iteratively updates the estimated object state of object 14 based on the estimated current state parameters passed by the fitting unit 26 and the associated estimate of the uncertainties of the estimated current state parameters of object 14 passed by the unscented transformation unit 28.

[0062] The estimated current state parameters of the object represent a pseudo-measurement P. The state estimator 30 receives the estimated uncertainties of these estimated current state parameters, i.e., the pseudo-measurement P, in the form of the variance ΔP. Considering the measurement uncertainties ΔP of the pseudo-measurement P thus increases the reliability of the update of the estimated object state by the state estimator 30.

[0063] The state estimator 30 can also estimate and update the dimensions of object 14 using the estimated current state parameters of object 14 (or its sub-objects 36). This allows, for example, even very large objects, such as trucks, to be fully captured step by step in terms of their dimensions. A more accurate estimate of the dimensions of object 14 can then, in a subsequent iteration (corresponding to a subsequent measurement cycle), lead to a better association of new location measurements 10 by the association unit 24. The orientation of object 14 can also be updated accordingly.

[0064] The described steps can be carried out accordingly for all objects 14 and associated location measurements 10.

Claims

1. Method for evaluating location measurements of an environment sensor in the form of a radar sensor for a motor vehicle, comprising the steps of: obtaining location measurements (10) in radial coordinates for radar reflections from an object from the environment sensor (20); associating obtained location measurements (10) with an already created object (14) described by an estimated object state, wherein the estimated object state is described by state parameters in the form of a Cartesian position, extent and / or velocity in Cartesian coordinates, and wherein an association probability (p) for the association of the location measurement (10) with the already created object (14) is determined for each of the location measurements (10) in consideration of the estimated object state of the object (14); estimating current state parameters (P) of the object (14), comprising customizing the state parameters to the location measurements (10) associated with the object (14), the customization giving consideration to weightings of the location measurements (10) associated with the object (14), the weighting for each of the location measurements (10) depending on the determined association probability (p) for the association of the relevant location measurement (10) with the object (14); transferring the estimated current state parameters (P) of the object (14) to a state estimator (30) to update the estimated object state of the object (14).

2. Method according to Claim 1, in which the determination of the association probability (p) for the association of a location measurement (10) with the object (14) involves the association probability (p) representing a probability of the location measurement (10) being an exact location of a real object to which the estimated object state relates.

3. Method according to Claim 1 or 2, in which the determination of the association probability (p) for the association of a location measurement with the object (14) involves the association probability (p) being determined in consideration of the measurement uncertainty of the location measurement (10).

4. Method according to one of the preceding claims, in which the determination of the association probability (p) for the association of a location measurement (10) with the object (14) involves the association probability (p) being determined in consideration of an uncertainty of the estimated object state of the object (14).

5. Method according to one of the preceding claims, in which the determination of the association probability (p) for the association of a location measurement (10) with the object (14) involves the association probability (p) being determined in consideration of a model uncertainty of the estimated object state of the object (14).

6. Method according to one of the preceding claims, the method comprising the steps of: obtaining measurement uncertainties of the location measurements (10) from the environment sensor (20); and estimating uncertainties (ΔP) of the estimated current state parameters (P) of the object (14) on the basis of the obtained measurement uncertainties of the location measurements (10) associated with the object (14), wherein the estimated current state parameters (P) of the object (14) and the estimated uncertainties (ΔP) are transferred to the state estimator (30) to update the estimated object state of the object (14).

7. Method according to Claim 6, wherein the estimation of uncertainties (ΔP) of the estimated current state parameters of the object (14) is carried out by means of an unscented transformation, wherein the location measurements (10) associated with the object (14) and the obtained measurement uncertainties of the location measurements (10) associated with the object (14) are used as input variables for the unscented transformation.

8. Method according to Claim 6 or 7, in which the estimation of uncertainties (ΔP) of the estimated current state parameters (P) of the object (14) is carried out by means of an unscented transformation, comprising the steps of: computing (S10) sigma points of the unscented transformation, wherein a vector (Y) comprising the location measurements (10) associated with the object (14) is used as a mean estimate for a distribution of the sigma points to be computed, and wherein a matrix comprising the measurement uncertainties of the location measurements (10) associated with the object (14) is used as a covariance matrix (K) for the distribution of the sigma points to be computed; estimating (S12) current state parameters of the object (14) for each sigma point of the unscented transformation, comprising customizing the state parameters to the respective sigma point; determining (S14) a variance of a distribution of the current state parameters of the object (14) that are estimated for the sigma points as an estimate of the uncertainties (ΔP) of the estimated current state parameters (P) of the object (14).

9. Sensor system for a motor vehicle, comprising a radar environment sensor (20) and an evaluation unit (24, 26, 28, 30) for evaluating location measurements of the environment sensor (20), the evaluation unit (24, 26, 28, 30) comprising a state estimator (30) for iteratively updating estimated object states of respective objects (14), the evaluation unit (24, 26, 28, 30) being configured to carry out the method according to one of the preceding claims.