FR3120695A1 - Method for automatic and on-board supervision of the accuracy of cartographic information by an autonomous vehicle - Google Patents

Abstract

A method of supervising (3) a digital navigation map (14) of an autonomous motor vehicle, comprising: A step of smoothing (322) the data of the trajectory of the vehicle during said route taken; A step of recalculating ( 323) observation residues associated with elements of the navigation map (13);A merging step (324, 325) of the residues associated with the same element of the digital navigation map ;An estimation step (326 ) the reliability of each residue obtained after said merging step; and a step of correcting (327) said digital navigation map based on the reliability estimate of each residual. Abstract Figure: Figure 1

Description

Method for automatic and on-board supervision of the accuracy of cartographic information by an autonomous vehicle

The invention relates to the field of guidance of autonomous motor vehicles and more particularly the automatic and on-board supervision of the accuracy of the cartographic information of the autonomous vehicle.

When a motor vehicle operates in an autonomous driving mode, it calculates its route by matching its geographical position with data from a digital navigation map.

The general geographical positioning of the vehicle and the elements of the map is based on positioning by Geolocation and Navigation by a Satellite System, better known by its acronym GNSS for English Global Navigation Satellite System .

However, since this system is not precise enough to guarantee good progress on a road, it is coupled with other means of positioning, in particular cameras or laser sensors, such as Lidars, making it possible to detect markings on the ground. , road signs, or other exteroceptive information representative of the environment of the motor vehicle.

Thus, the motor vehicle is positioned by matching the elements detected by the on-board means, cameras, Lidars, etc., with the GNSS positioning and the associated navigation map.

In particular, the system matches the road signs detected by the motor vehicle with the known road signs of the digital navigation map.

However, as the road network is changing rapidly, some signs present in the digital navigation map may have disappeared in reality; and vice versa, some signs that actually exist may not be indicated on the map.

The objective of the present application is therefore to automatically evaluate the quality of the digital navigation map in order to provide an accurate method of locating the motor vehicle, despite the potential information errors of the digital navigation map.

We know in particular the scientific publication " HD Map Change Detection with a Boosted Particle Filter ", D. Pannen, M. Liebner and W. Burgard, 2019 International Conference on Robotics and Automation (ICRA), Montreal, QC, Canada, 2019, pp . 2561-2567. This publication describes an error detection on a digital navigation map based on a particle filter. However, particle filtering as described in this publication is relatively complex to implement and involves a significant computational load.

Therefore, there is a need for a method of supervising a digital navigation map, robust and lighter from a computational point of view.

To this end, a method for supervising a digital navigation map of an autonomous motor vehicle is proposed, said motor vehicle comprising a location module suitable for acquiring, storing and associating with elements of the navigation map a plurality of data status of the trajectory of the motor vehicle and observation measurements of the environment of the motor vehicle during a journey, said observation measurements being associated with elements present on said navigation map so that the module calculates the observation residuals of the motor vehicle in relation to the data of the navigation map, and when the journey of the motor vehicle is finished, said localization module is then adapted to implement the steps of said supervision method comprising:

– A step of smoothing the data of the trajectory of the vehicle during said journey made;

– A step of recalculating the observation residues associated with elements of the navigation map after said smoothing of the trajectory data and the calculation of the covariance matrix associated with each residue;

– A step of merging residues associated with the same element of the digital navigation map;

– A step of estimating the reliability of each residue obtained after said fusion step; and

– A step of correcting said digital navigation map according to the reliability estimate of each residue.

The supervision method is thus robust, fast, efficient and does not need to be carried out in real time during the operation of the motor vehicle.

Advantageously, said smoothing step comprises the implementation of Kalman smoothing by retroactive propagation of the position corrections of the motor vehicle. Thus, it is possible to obtain high-performance smoothing that is inexpensive in terms of calculation time, allowing a significant improvement in the real position of the motor vehicle with respect to the navigation map.

Advantageously, said merging step comprises a first merging of the recalculated residuals associated with the same element of the digital navigation map by the covariance intersection method. Thus, we obtain a method for calculating the first fusion of residues that is robust and efficient.

Advantageously, said merging step comprises a second merging of the residue obtained by the first merging with, if it exists, a residue already recorded during a previous implementation of the method for the same element of the digital navigation map, of so that the remaining residue includes all the information of the previous residues. Thus, we make sure to obtain a relevant residual to correct the digital navigation map.

Advantageously, the step of estimating the reliability of each residue comprises the implementation of a test of the , as a function of a sensitivity value predetermined. This statistical test thus ensures a robust and efficient detection of the detected observation errors.

Advantageously, when said test of fails, an exclusion variable indicating that said map is erroneous for said element associated with said residual is assigned to the element of the navigation map associated with said residual having failed. Thus, it is possible in a very efficient and simple manner to identify the erroneous elements on said navigation map.

Advantageously, said step of correcting said digital navigation map comprises correcting the residuals associated with an exclusion variable. Thus, it is relatively efficient to identify the elements to be corrected.

The invention also relates to a module for locating an autonomous motor vehicle comprising means for acquiring, storing and associating with elements of the navigation map a plurality of state data of the trajectory of the vehicle motor vehicle and observation measurements of the environment of the motor vehicle during a trip, said observation measurements being associated with elements present on said navigation map such that the module is suitable for calculating the residuals of observation of the motor vehicle in relation to the data of the navigation map, and when the journey of the motor vehicle is finished, said location module comprises means for implementing the following supervision steps:

– A step of smoothing the data of the trajectory of the vehicle during said journey made;

– A step of recalculating the observation residues associated with elements of the navigation map after said smoothing of the trajectory data and the calculation of the covariance matrix associated with each residue;

– A step of merging residues associated with the same element of the digital navigation map;

– A step of estimating the reliability of each residue obtained after said fusion step; and

– A step of correcting said digital navigation map according to the reliability estimate of each residue.

The invention also relates to an autonomous motor vehicle comprising a location module as described previously.

Other features and advantages of the invention will become apparent on reading the description given below of a particular embodiment of the invention, given by way of indication but not limitation, with reference to the appended drawings in which:

is a schematic view of a location module according to one embodiment of the invention;

is a schematic representation of the observation model of a traffic sign by an autonomous vehicle according to the invention;

is a flowchart of the method for supervising a navigation map according to the embodiment of the invention.

With reference to the , an autonomous motor vehicle comprises a location module 1, which will also be referred to more simply as module 1.

Module 1 communicates with a plurality of other modules including:

– a CAN 10 Bus (Controller Area Network) adapted to supply digitized signals from proprioceptive sensors;

– A GNSS receiver 11 adapted to provide the geographical position of the motor vehicle;

– An intelligent camera 12 suitable for detecting and identifying the lines demarcating the traffic lanes;

– A module 13 for detecting elements of the road network, such as road signs or markings on the ground, here using data in particular from a Lidar sensor 15 or on-board cameras;

– A module 14 for providing a digital navigation map.

This location module 1 then implements the following general steps:

– a first step called Real time filtering 16, in French real time filtering 16, during which the state of the vehicle is estimated at high frequency using the inertial data provided by the series sensors of the vehicle via the CAN bus 10 as well as the measurements from the GNSS receiver 11.

The location module 1 thus provides an up-to-date status of the vehicle including its estimated position, its longitudinal speed as well as its yaw rate.

These estimated states are available at high frequency but may be marred by errors, due to the limitations of real-time filtering.

A second smoothing step 17, in English Smoothing , then consists in reprocessing the present state according to the future states in order to improve its precision and reduce its uncertainty.

- The level of quality of the smoothed trajectory then makes it possible to implement an association step 18 matching the elements perceived by the exteroceptive sensors of the vehicle (road signs, markings on the ground) with the elements referenced in the navigation map digital.

This association step 18 provides new information on the position of the vehicle relative to its environment and makes it possible to refine the estimation of its pose by the real-time filtering function 16.

However, as the road network is constantly changing, the elements referenced in the digital navigation map may be inaccurate or erroneous.

- The study of the associations calculated in the previous step then makes it possible to evaluate their reliability at the end of each journey. To this end, a supervision method 3 of the digital navigation map, in English Map reliability , is implemented.

This supervision method 3, with reference to the , is implemented after the end of the running of said motor vehicle.

However, beforehand, during driving, the method 3 implements a step 321 of storing in memory all the states received from the motor vehicle and the associations made during driving since the start of the autonomous navigation module.

The term “all the states received” means all of the information acquired by the location module 1 explained above.

Indeed, to analyze the quality of the elements referenced in the map, the measurements and the estimated states are used. These data are recorded during the trajectory to be processed only after the end of the rolling of the motor vehicle.

Module 1 receives observations z _j ^{( k )} (where k is the instant of measurement and i is the index of the measurement at this instant) and their associated map element m _i from the perception module. These data are stored in a buffer memory, also called in English buffer , so as to obtain at the end of the trajectory, an observation buffer :

where Z ^{( k )} Contains the observations obtained at time k:

In parallel, the estimations of the localization module 1 are also recorded.

The localization module 1 thus provides an estimate of the pose (position in the plane and orientation) of the vehicle, of its longitudinal speed and of rotation. These data are estimated by the localization module 1 at time k using at most the observations from time 0 to k.

These states x _k|k and their associated covariance matrices P _k|k are stored in memory so as to obtain the state buffer at the end of the trajectory:

Once the rolling is finished, the steps of the estimation method 3 are then implemented.

First of all, a smoothing step 322 is carried out on the entire trajectory.

In this embodiment, smoothing is implemented from the end of the trip until the start.

Since the states recorded in the buffer could only be estimated by using the observations available in real time, these estimates can be improved once taxiing and the trajectory are finished, to propagate the information obtained thanks to the new measurements.

This is achieved by Kalman smoothing, which consists of applying a corrective factor to each state.

This factor depends on the following estimates and is applied from the last state to the first.

Thus, when observations allow a better estimate of the state x _k+1 compared to the state x _k . The corrective term makes it possible to propagate this improvement to the state x _k . So the whole state buffer can be improved by applying the functions:

with

For the implementation of the smoothing step 322, the necessary observations only concern the elements of the navigation map, the other observations then not needing to be kept.

Once the smoothing has been carried out, a step of recalculation 323 of new observation residues as well as the covariance matrix of each observation residue is then implemented.

By residual, we mean the difference between an observation and the predicted value resulting from a statistical regression.

Indeed, to assess the quality of the elements of the navigation map, residual observations are used.

An observation error is modeled by a Gaussian distribution centered on the element of the map and a covariance matrix depending on the a priori precision of the map and the precision of the detection.

If the element of the map is well referenced, then we know that y ^T S ^-1 y

(where y is the measurement residual and S its associated covariance matrix) follows a law of .

By setting a risk α, remove the elements from the map if it does not verify the test of the will cause the deletion of 1 element out of 1/α.

If the map contains an error, the deletion rate will be much higher. Indeed if an element is well referenced, as the residues are used, the uncertainty of the residue will decrease and the residue will tend towards 0.

On the other hand, if the element is badly referenced, the uncertainty will always decrease but the residue will tend towards a non-zero constant. So in case of card error the test fails much more often.

The calculated residuals must be comparable. Whatever the point of view, the residuals must represent the same quantities.

Where for localization using a [ ^S x _{i S} y _i ] ^T landmark observed in the vehicle frame, the observation model with reference to the applies:

With observation:

Such a model is however not suitable for the evaluation of map elements because the residuals depend on the point of view (since they represent quantities expressed in the vehicle frame.

Thus, to obtain a usable residue, we choose a model of the form:

With observation:

This model produces observation residuals expressed in the local coordinate system on the navigation map. Thus, regardless of the point of view of the element of the map, the residuals represent the same magnitude. They can therefore be compared and merged from one trajectory to another.

To calculate the residual and its associated covariance matrix, two cases must be distinguished.

– If the observation was used to estimate the state, the error of the state estimate is correlated with the observation error. The residue is therefore calculated with:

And its covariance matrix is:

where R is the covariance matrix of the measurement noise and H is the Jacobian of the observation model with respect to x _k

– If the observation was not used to estimate the state, its error is not correlated with the state error and the covariance matrix of the residual is calculated with:

Once the residuals have been recalculated 323, it is possible that for the same element of the map there are several calculated residuals and the associated covariance matrices.

The recalculated residues 324 are then merged.

This step of merging the recalculated residuals 324 is carried out by the covariance intersection method.

However, the invention is not limited to this method alone.

In particular, the computational needs can be reduced even further by approximating the fusion with a method known as fast covariance intersection, better known in English as “ Fast Covariance Intersection ” which offers an analytical approximated solution for the parameters ω .

The method can also be approximated further by selecting the worst residual instead of merging them all. The worst residual can be selected by choosing the residual with the largest Mahalanobis distance.

However, returning to the covariance intersection method implemented for this embodiment, for each map element, theLresiduesthere ₁ atthere _L (and associated covariance matricesS ₁ atS _L ) with :

where the parameters ω are positive and such that . These parameters being estimated by optimization to minimize Tr( S ), the trace of the matrix S.

However, it is possible that for a given element of the map, during a previous implementation of the method, a previous residue, also called saved residue, has already been calculated.

In this case, the merging 325 of the saved residue is then carried out with the residue resulting from the merging of step 324.

The residue saved in memory is then merged with the new residues so as to obtain a single residue containing all the information.

This merger is carried out as follows:

Then we proceed to a step 326 of evaluation of the reliability of the element according to the residue obtained after the fusion steps 324, 325.

Indeed, the residuals now contain all the information obtained during this passage in the vicinity of the element of the map and potential previous passages.

These residuals can now be used to assess the reliability of map elements.

As explained above, if the map has no error, each residual follows a Gaussian distribution centered at 0 and with a covariance matrix S .

In the event of a map error, this distribution will not be centered at 0, it will be centered around the map bias, in other words the error.

So to detect card errors, a test of the is carried out on the residue with:

dim(y) being the dimension of the matrix of each residue.

If the test fails, two cases are possible:

– If there is no error in the map, the test will fail in α Percent of cases. α must be chosen low enough to limit these occurrences.

– If there is a card error, the test will fail in a much higher proportion. In both cases, an exclusion variable, for example a "do not use" variable or in English don't use is entered in the map to allow the localization system not to use these landmarks during the next passes.

Such a variable can evolve with each repetition of the method, reproducing itself after different paths or passages in front of an element of the map. When the test fails despite the absence of a card error, it is possible that on the next pass, the residual decreases enough to pass the test.

Thus, each new observation makes it possible to develop knowledge about the reliability of the map element.

In an alternative implementation of the invention, the test of the making it possible to determine the creation of the exclusion variable is not carried out directly by the method according to the invention, but by the systems seeking to use the element of the card in question. This can be useful as different systems may have different risk tolerances. Also a system can tolerate variable risk depending on the situation. The test of can be realized by each system because the residuals and covariance matrices are stored in the map.

Finally, a correction step 327 of the map is implemented.

When a card error is detected, the don't use variable is populated.

This allows the tracking system not to use erroneous elements. Losing a landmark for localization is not ideal as it can lead to errors that are too large for the autonomous driving application.

It is then preferable to correct this error to allow the localization system to continue to use the known landmarks.

To correct the error, the observation model is used. We know the relationship:

We can then find the correction Δ m to be applied to the references of the landmark to obtain a zero residual.

So for the example of a panel observation, we have the relationship:

We therefore deduce that to obtain a zero residual, the new coordinates of the landmark must be:

In the case where the residual has the same dimension as the landmark reference, there is a unique solution Δ m that can be found. If the residue is of dimension lower than the dimension of the elements describing the landmark, there is no uniqueness of the correction to be made.

For example, a measurement of distance to a marking has a residual of dimension 1 while the marking referenced by a straight line is parameterized by two terms.

To solve the system, an additional constraint must be added.

In this embodiment, the addition of a parallelism constraint makes it possible to find a single correction.

The method according to the invention thus makes it possible to improve the reliability of the location by means of the detection of erroneous map elements.

In addition, the invention has several advantages:

– the method integrates into a standard fusion architecture without requiring any internal modification to the localization system. The only additional constraint is to save the states and observations processed by the localization system, which is inexpensive. This makes this method easily integrated into existing systems.

– the method does not require any real-time calculation but exploits the real-time calculations already carried out by the localization system. The rest of the calculations are done a posteriori. Thus this strategy does not weigh down the real-time system and is applicable to systems where the resources available in real time are limited.

– the method is suitable for exploiting several trajectories in order to detect errors. This ability allows this method to detect errors that could not have been observed during a single trajectory.

– reducing the number of map elements also reduces the number of possible observations and may therefore result in poorer localization. It is therefore preferable to correct the detected errors. The proposed correction thus makes it possible to keep the element rather than no longer using it for localization. This allows the proposed method to lead to better localization.

Claims (9)

Method of supervising (3) a digital navigation map (14) of an autonomous motor vehicle, said motor vehicle comprising a localization module (1) suitable for acquiring, storing (321) and associating with elements of the map navigation device (14) a plurality of status data (10, 11) of the trajectory of the motor vehicle and observation measurements (10, 12, 13, 15) of the environment of the motor vehicle during a route, said observation measurements (10,12,13,15) being associated with elements present on said navigation map (14) so that the module calculates the observation residuals of the motor vehicle with respect to the data of the navigation card (14), and when the journey of the motor vehicle is finished, said location module (1) is then adapted to implement the steps of said supervision method (3) comprising:

• A step of smoothing (322) the data of the trajectory of the vehicle during said route taken;
• A step of recalculating (323) the observation residues associated with elements of the navigation map (13) after said smoothing of the data of the trajectory and the calculation of the covariance matrix associated with each residue;
• A merging step (324, 325) of the residues associated with the same element of the digital navigation map;
• A step of estimating (326) the reliability of each residue obtained after said merging step; and
• A step of correcting (327) said digital navigation map according to the reliability estimate of each residue.

Method (3) according to claim 1, characterized in that said smoothing step (322) comprises the implementation of Kalman smoothing by retroactive propagation of the motor vehicle position corrections. Method (3) according to Claim 1 or 2, characterized in that the said merging step (324, 325) comprises a first merging (324) of the recalculated residues associated with the same element of the digital navigation map by the method of covariance intersection. Method (3) according to claim 3, characterized in that said merging step (324, 325) comprises a second merging (325) of the residue obtained by the first merging (324) with, if it exists, a residue already recorded during a previous implementation of the method (3) for the same element of the digital navigation map, so that the remaining residue comprises all the information of the previous residues. Method (3) according to any one of Claims 1 to 4, characterized in that the step of estimating (326) the reliability of each residue comprises the implementation of a test of the , as a function of a sensitivity value predetermined. Method (3) according to claim 5, characterized in that when said test of the fails, the element of the navigation map associated with said residue having failed, residue is assigned an exclusion variable indicating that said map is erroneous for said element associated with said residue. A method (3) according to claim 6, characterized in that said step of correcting (327) said digital navigation map includes correcting residuals associated with an exclusion variable. Locating module (1) of an autonomous motor vehicle comprising means for acquiring, storing and associating with elements of the navigation map (14) a plurality of status data (10, 11) of the trajectory of the motor vehicle and observation measurements (10, 12, 13, 15) of the environment of the motor vehicle during a journey, said observation measurements (10,12,13,15) being associated with elements present on said navigation map (14) so that the module is suitable for calculating the observation residuals of the motor vehicle with respect to the data of the navigation map (14), and when the path of the motor vehicle is completed, said location module (1) comprises means for implementing the following supervision steps (3):

• A step of smoothing (322) the data of the trajectory of the vehicle during said route taken;
• A step of recalculating (323) the observation residues associated with elements of the navigation map (13) after said smoothing of the data of the trajectory and the calculation of the covariance matrix associated with each residue;
• A merging step (324, 325) of the residues associated with the same element of the digital navigation map;
• A step of estimating (326) the reliability of each residue obtained after said merging step; and
• A step of correcting (327) said digital navigation map according to the reliability estimate of each residue.

Autonomous motor vehicle comprising a location module (1) according to claim 8.