Method for Determining an Object's Position Using Different Items of Sensor Information

Abstract

A method for determining an object's position using different items of sensor information includes: a) reading the sensor information into a Kalman filter, b) merging the sensor information with the Kalman filter, with the Kalman filter supplying as a result estimated values for states and information associated with the estimated values regarding the accuracy of the estimates, c) monitoring the results of the Kalman filter by estimating a probability of accuracy, with which the estimation error lies within an error band, with the probability of accuracy being determined on the basis of a plurality of conditional probabilities, the conditions for which each relate to estimation errors from at least one earlier series.

Claims

1. A method for determining a position of an object using different sensor information items, comprising: reading the sensor information items into a Kalman filter; fusing the sensor information items using the Kalman filter, the Kalman filter delivering as results estimated values for states and information, associated with the estimated values (x.sub.estimate), concerning an accuracy of the estimated values; and monitoring the results of the Kalman filter by estimating an accuracy probability with which an estimation error lies within an error band, the accuracy probability established using a plurality of conditional probabilities, conditions of the conditional probabilities respectively relate to other estimation errors from at least one earlier time step.

2. The method as claimed in claim 1, wherein monitoring the results further comprises: establishing bounds for the conditional probabilities.

3. The method as claimed in claim 2, wherein monitoring the results further comprises: using the Bienaymé-Tchebycheff inequality and/or the Berge inequality to establish the bounds.

4. The method as claimed in claim 1, wherein monitoring the results further comprises: establishing a bound for the accuracy probability with the following formula: $\underset{i=0}{\overset{n-2}{.Math.}}\frac{{\sigma }_{t-i}^2+{\sigma }_{t-i-1}^2+2{\sigma }_{t-i}{\sigma }_{t-i-1}}{\left(1-\gamma \right)\max \left(2{\epsilon }^2,{\max \left(0.2\epsilon -{\sigma }_{t-i}-{\sigma }_{t-i-1}\right)}^2\right)}.Math.\frac{{\sigma }_{t-n+1}^2+{\sigma }_{t-n}^2+2{\sigma }_{t-n+1}{\sigma }_{t-n}}{\max \left(2{\epsilon }^2,{\max \left(0.2\epsilon -{\sigma }_{t-n+1}-{\sigma }_{t-n}\right)}^2\right)}+\underset{j=0}{\overset{n-1}{.Math.}}\underset{i=0}{\overset{j-1}{.Math.}}\frac{{\sigma }_{t-i}^2+{\sigma }_{t-i-1}^2+2{\sigma }_{t-i}{\sigma }_{t-i-1}}{\left(1-\gamma \right)\max \left(2{\epsilon }^2,{\max \left(0.2\epsilon -{\sigma }_{t-i}-{\sigma }_{t-i-1}\right)}^2\right)}\left(\gamma -\frac{{\epsilon }^2-{\sigma }_{t-j-1}^2-{\sigma }_{t-j}^2}{{\epsilon }^2}\right)=F\left(\epsilon ,\gamma ,n,{\sigma }_i\right)$

5. The method as claimed in claim 1, wherein monitoring the results further comprises: establishing, for the estimated values of the Kalman filter, whether with a given accuracy probability they have an acceptable estimation error.

6. The method as claimed in claim 1, further comprising: determining a vehicle position using the different sensor information items of vehicle sensors.

7. The method as claimed in claim 1, wherein a computer program is configured to carry out the method.

8. The method as claimed in claim 7, wherein the computer program is stored on a non-transitory machine-readable storage medium.

9. The method as claimed in claim 8, wherein a controller for a motor vehicle includes the storage medium.

Description

[0026] FIG. 1: schematically shows an exemplary flowchart of the described method, and

[0027] FIG. 2: schematically shows a graphical illustration of an exemplary establishment of the accuracy probability.

[0028] FIG. 1 schematically shows an exemplary flowchart of the described method. The method is used to determine an object position by using different sensor information items. The sequence of steps a), b) and c), represented by the blocks 110, 120 and 130, is exemplary and may be carried out in this way at least once or several times in succession. Furthermore, the steps a), b) and c), in particular the steps b) and c), may also be carried out at least partially in parallel or simultaneously.

[0029] In block 110, according to step a) the sensor information items are read into a Kalman filter. In block 120, according to step b) fusion of the sensor information items is carried out using the Kalman filter, the Kalman filter delivering as a result estimated values x.sub.estimate for states x and information σ.sup.2, associated with these estimated values x.sub.estimate, concerning the accuracy of the estimations. In block 130, according to step c) the results of the Kalman filter are monitored by estimating an accuracy probability P(|X.sub.t|<ε) with which the estimation error X=x−x.sub.estimate lies within an error band ±F, the accuracy probability P(|X.sub.t|<ε) being established by employing a plurality of conditional probabilities, the conditions of which respectively relate to estimation errors X.sub.t−1, X.sub.t−2, . . . from at least one earlier time step.

[0030] The Kalman filter generally delivers not only an estimation of the states x but also their variances σ.sup.2. It therefore provides information relating to their robustness. The question is when the estimations of the Kalman filter lie within a particular error tolerance. This is of especial importance in the starting phase since the filter first needs to stabilize after initialization. One fundamental task in the use of Kalman filters is therefore to determine the end of this stabilization process.

[0031] This may particularly advantageously be determined using the method. Furthermore, with the presented method it is also possible to determine at any desired later instant whether with a (given) (minimum) probability or a (minimum) accuracy probability (symbol γ) the error lies in a (given) error band (symbol ±ε) and is therefore dependable. For example, a warning may be triggered if the expected errors become too great for the application in question because of inaccurate measurement values.

[0032] In particular, a lower bound γ is derived for the probability P(|x−x.sub.est|≤ε)≥γ. Here, ε is a (given) maximum acceptable error. One possible application in this context is as follows: if the difference between the estimated state x.sub.est and the real state x is with a sufficiently high probability less than the (given) error limit ε, the estimation is assumed to be reliable.

[0033] One particular difficulty in this context is to derive a lower bound γ for the probability P(|x−x.sub.est|≤ε) without having knowledge relating to the underlying probability distribution and correlation. While a solution on the basis of a normal distribution would be simple to calculate, the general case raises some problems. Here again, for this case a possible solution may particularly advantageously be given by using the Bienaymé-Tchebycheff inequality and/or the Berge inequality in order to establish bounds for the conditional probabilities. The Bienaymé-Tchebycheff inequality used and the Berge inequality used are given in ‘R. Savage: “Probability Inequalities of the Tchebycheff Type”. In: Journal of Research of the National Bureau of Standards—B. Mathematics and Mathematical Statistics 65B.3 (1961)’.

[0034] In this case, as in particular the only assumption, the assumption may be made that the variances from the Kalman filter are not affected by error. Furthermore, in particular model errors and/or measurement outliers are also not taken into account. The methodology applied in the method may be mathematically derived on the basis of this assumption, in particular without further heuristic assumptions. This further advantageously makes it possible to apply the methodology without further adaptations in a very wide variety of fields. In addition, for particularly safety-critical applications, the safeguard is provided that the derived bounds apply reliably and are not merely approximately correct. Of course, the latter is true only with the requirement of validity of the variances calculated by the Kalman filter.

[0035] A Kalman filter estimates the states of a process on the basis of the preceding states (prediction) and measured quantities (measurement update). It also calculates the associated variance for each state. The calculation of the estimated variances is correspondingly based on the variance of the preceding state and the variances of the measurement values which are used to calculate the state. These variances are denoted below by σ.sup.2. Here, t refers to the t.sup.th time step of the Kalman filter. The variances deliver information relating to the reliability of the estimation of the states. The relationship between variance and expected error is, however, dependent on the underlying distribution function. If this is unknown, the conversion between variance and error can only be estimated, for example by applying the Chebyshev inequality. If the errors of the states are furthermore correlated, even the Chebyshev inequality can no longer be applied. For this case, a particular advantageous (specific) estimation is derived here:

[0036] To this end, the Bienaymé-Tchebycheff inequality and the Berge inequality, as are given for example in ‘Journal of Research of the National Bureau of Standards—B. Mathematics and Mathematical Statistics 65B.3 (1961)’, are preferably used. They are used to derive bounds for a plurality of conditional probabilities, in particular P(|X.sub.t|>ε| |X.sub.t−1|>ε), P(|X.sub.t|>ε| |X.sub.t−1|<ε), P(|X.sub.t|<ε| |X.sub.t−1|>ε) and P(|X.sub.t|<ε| |X.sub.t−1|<ε).

[0037] These probabilities are assigned to the edges of a trellis. This is represented by way of example in FIG. 2. In this case, the lower edge of the trellis represents the states in which the error limit e is complied with. The upper edge, on the other hand, shows the states in which the error limit e is not complied with but is exceeded.

[0038] An upper bound for P(|X.sub.t|>ε) at the instant t is now derived along the edges of the trellis. The final result uses the variances σ.sup.2 of the Kalman filter at the instant i:

[00002]$P\left(.Math.e_t.Math.>\epsilon \right)\le \underset{i=0}{\overset{n-2}{.Math.}}\frac{{\sigma }_{t-i}^2+{\sigma }_{t-i-1}^2+2{\sigma }_{t-i}{\sigma }_{t-i-1}}{\left(1-\gamma \right)\max \left(2{\epsilon }^2,{\max \left(0.2\epsilon -{\sigma }_{t-i}-{\sigma }_{t-i-1}\right)}^2\right)}.Math.\frac{{\sigma }_{t-n+1}^2+{\sigma }_{t-n}^2+2{\sigma }_{t-n+1}{\sigma }_{t-n}}{\max \left(2{\epsilon }^2,{\max \left(0.2\epsilon -{\sigma }_{t-n+1}-{\sigma }_{t-n}\right)}^2\right)}+\underset{j=0}{\overset{n-1}{.Math.}}\underset{i=0}{\overset{j-1}{.Math.}}\frac{{\sigma }_{t-i}^2+{\sigma }_{t-i-1}^2+2{\sigma }_{t-i}{\sigma }_{t-i-1}}{\left(1-\gamma \right)\max \left(2{\epsilon }^2,{\max \left(0.2\epsilon -{\sigma }_{t-i}-{\sigma }_{t-i-1}\right)}^2\right)}\left(\gamma -\frac{{\epsilon }^2-{\sigma }_{t-j-1}^2-{\sigma }_{t-j}^2}{{\epsilon }^2}\right)=F\left(\epsilon ,\gamma ,n,{\sigma }_i\right)$

[0039] Therefore,

P(|e.sub.t|<ε)≥1−F(ε,γ,n,σ.sub.i)

[0040] These bounds make it possible to evaluate whether the error e.sub.t, or X.sub.t, lies with a probability of at least 1−F(ε,γ,n,σ.sub.i) within an error band ±ε. For each estimated value X.sub.estimate of the Kalman filter, it is therefore possible to indicate whether with a given (minimum, or minimum accuracy) probability γ it has an acceptable (estimated) error e.sub.t, or X.sub.t. In this case, γ describes the (minimum, or minimum accuracy) probability with which the error should lie within the error band ∓ε. In other words, γ describes a predetermined target parameter for the accuracy probability P(|X.sub.t|<ε).

[0041] This means, in other words, in particular that ε and/or γ may be predetermined, so that a bound for P(|X.sub.t|<ε) can be established. The bound actually established in this way for the probability P(|X.sub.t|<ε) may be compared with the (fixed or predetermined) minimum probability γ. On the basis of this assessment, further measures may then be taken for the corresponding application, for example release for autonomous driving, after stabilization of the filter.

[0042] Preferably, the method is carried out in order to determine a vehicle position by using different sensor information items of vehicle sensors. In this case, the described method may possibly be used to decide whether the estimated position may be trusted and the vehicle may drive autonomously.

[0043] The method advantageously permits an improvement of the position accuracy. In particular, the methodology presented here permits a valid estimation even in the event of correlated errors.