Method and device for determining the accuracy of an occupancy grid to be evaluated for a driver assistance system

Abstract

Method for determining an accuracy (G) of an occupancy grid (BG) to be evaluated for a driver assistance system of a vehicle, comprising the following process steps: (a) Segmentation (S A ) the occupancy grid (BG) to be evaluated and at least one reference occupancy grid (R-BG) for generating binarized grids (b BG, b RBG); (b) cell-by-cell comparison (S B ) the binarized grids to generate an error map (FK) which indicates for each cell an error of the occupancy grid (BG) to be evaluated with respect to the reference occupancy grid (R-BG); (c) Carry out (S C ) a distance transformation on the error map (FK) to generate an error deviation map (FAK) which indicates a geometric deviation of the occupancy grid (BG) to be evaluated from the reference occupancy grid (R-BG); and (d) Evaluate (S D) of the generated error-deviation map (FAK) to determine the accuracy (G) of the occupancy grid (BG) to be evaluated.

Description

[0001] The invention relates to a method and a device for determining an accuracy of an occupancy grid to be evaluated for a driver assistance system of a vehicle.

[0002] An occupancy grid is a data structure for describing the environment of a vehicle. The environment of the vehicle is divided into cells, and for each cell, a probability is specified as to whether the respective cell is occupied or unoccupied. Each cell contains a probability value between 0 and 1, which indicates the probability that the respective cell is occupied. For example, the probability value can indicate whether the respective cell is occupied by a static obstacle in the vehicle's environment.

[0003] In many use cases, it is necessary or helpful to determine the accuracy of an occupancy grid. For example, it is necessary to evaluate how changes made to a data processing algorithm for calculating an occupancy grid affect its accuracy. Another use case is proving the accuracy of an occupancy grid for systems that use data stored in occupancy grids. Furthermore, different sensors that provide sensor data that is processed by data processing units into an occupancy grid can be compared with each other based on the generated occupancy grids.

[0004] In the field of robotics, methods have been used that enable different occupancy grids to be compared with one another. These conventional methods can be divided into methods that compare cell-by-cell and methods that compare extracted features. A disadvantage of conventional cell-by-cell comparison methods is that the determined key figures do not correlate sufficiently with the features extracted for applications. Although a comparison based on extracted features delivers relatively good results for applications based on the corresponding features, these results cannot be transferred to other applications or application groups. For example, for road edge estimation, features that describe lateral boundaries of the drivable space, i.e. boundaries extending in the direction of travel, are primarily extracted.While this allows for an assessment of lateral accuracy, it does not allow for the assessment of structures perpendicular to the direction of travel, such as those required for emergency braking or emergency evasive maneuvers. Therefore, an assessment of the motion grid (BG) is required for each individual application. This conventional approach is therefore complex, cumbersome, and time-consuming.

[0005] From WO 2013060323 A1 a method for a sensor system for environment detection is known, wherein a grid-based environment model is calculated.

[0006] From DE 102013217491 A1 a method for transmitting data from a grid-based environment model in a vehicle is known.

[0007] US 8315433 B2 shows a technology for determining obstacles around a vehicle using bird's eye view images.

[0008] DE 102009006214 A1 discloses a method for providing an environment representation of a vehicle.

[0009] It is therefore an object of the present invention to provide a device and a method for determining an accuracy of an occupancy grid to be evaluated, which provides meaningful and reliable results for different applications.

[0010] This object is achieved according to the invention by a method having the features specified in claim 1.

[0011] The invention therefore provides a method for determining the accuracy of an occupancy grid to be evaluated for a driver assistance system of a vehicle, comprising the following method steps: Segmenting the occupancy grid to be evaluated and at least one reference occupancy grid to generate binarized grids, Cell-by-cell comparison of the binarized grids to generate an error map, which indicates for each cell an error of the occupancy grid to be evaluated with respect to the reference grid, Performing a distance transformation on the error map to generate an error deviation map indicating a geometric deviation of the occupancy grid to be evaluated from the reference occupancy grid, and Evaluate the generated error-deviation map to determine the accuracy of the occupancy grid to be evaluated.

[0012] In one possible embodiment of the method according to the invention, the occupancy grid to be evaluated and the reference occupancy grid have cells that each contain probability values ​​that indicate the probability that the respective cell is occupied.

[0013] In one possible embodiment of the method according to the invention, the probability values ​​contained in the cells of the occupancy grid to be evaluated and the reference occupancy grid are compared with an adjustable threshold value for generating the binarized grids.

[0014] In a further possible embodiment of the method according to the invention, sensor data originating from at least one vehicle sensor mounted on the vehicle are evaluated in order to generate the occupancy grid to be evaluated.

[0015] In a further possible embodiment of the method according to the invention, the occupancy grid to be evaluated, generated from sensor data, is temporarily stored in a data memory.

[0016] In a further possible embodiment of the method according to the invention, the binarized grids are subtracted from one another to generate the error map.

[0017] In a further possible embodiment of the method according to the invention, the method steps are carried out iteratively with gradually increasing threshold values.

[0018] In a further possible embodiment of the method according to the invention, the set threshold values ​​lie in a selected probability range.

[0019] In a further possible embodiment of the method according to the invention, the values ​​contained in the cells of the error deviation maps are statistically evaluated to determine the accuracy of the occupancy grid to be evaluated.

[0020] In a further possible embodiment of the method according to the invention, vehicle sensors are recalibrated whose sensor data result in an occupancy grid whose determined accuracy is insufficient.

[0021] In a further possible embodiment of the method according to the invention, the vehicle sensors are recalibrated by readjusting their orientation, position or location on the vehicle body of the vehicle.

[0022] In a further possible embodiment of the method according to the invention, the reference occupancy grid is received via a communication interface of the vehicle and stored in a data memory.

[0023] In an alternative embodiment of the method according to the invention, the reference occupancy grid is read from an exchangeable data memory.

[0024] The invention further provides a calculation device for a driver assistance system of a vehicle having the features specified in claim 14.

[0025] The invention therefore provides a calculation device for a driver assistance system of a vehicle with: a segmentation unit which segments an occupancy grid to be evaluated and a reference occupancy grid to generate binarized grids; and a comparison unit which compares the binarized grids cell by cell to generate an error map which indicates for each cell an error of the occupancy grid to be evaluated with respect to the reference grid; a transformation unit that performs a distance transformation on the error map to generate an error deviation map that indicates a geometric deviation of the occupancy grid to be evaluated from the reference occupancy grid; and with an evaluation unit that evaluates the generated error deviation map to determine the accuracy of the occupancy grid to be evaluated.

[0026] In one possible embodiment of the calculation device according to the invention, a data processing unit is provided which evaluates sensor data originating from at least one vehicle sensor of the vehicle in order to generate the occupancy grid to be evaluated.

[0027] In one possible embodiment of the calculation device according to the invention, vehicle sensors whose sensor data result in an occupancy grid whose accuracy determined by the evaluation unit is insufficient can be recalibrated by a calibration unit of the vehicle.

[0028] In the following, possible embodiments of the method according to the invention and the device according to the invention for determining an accuracy of an occupancy grid to be evaluated are explained in more detail with reference to the attached figures.

[0029] They show: Fig. 1 is a block diagram illustrating an embodiment of a calculation device according to the invention for a driver assistance system of a vehicle; Fig. 2 is a flowchart illustrating an embodiment of the method according to the invention for determining an accuracy of an occupancy grid to be evaluated; Fig. 3 shows a further flowchart illustrating an embodiment of the method according to the invention for determining an accuracy of an occupancy grid to be evaluated; Fig. 4A, Fig. 4B Diagrams to explain the operation of a transformation unit used in the device according to the invention.

[0030] How to Fig. 1, in the illustrated embodiment, a calculation device 1 for a driver assistance system of a vehicle according to the invention contains several units. In the illustrated embodiment, the calculation device 1 has a segmentation unit 1A, which segments an occupancy grid BG to be evaluated and at least one reference occupancy grid R-BG to generate binarized grids bGB, bRBG. The occupancy grid BG to be evaluated and the reference occupancy grid R-BG have cells Z, each containing a probability value P, which indicates the probability that the respective cell Z is occupied. The occupancy grid BG to be evaluated and the reference occupancy grid R-BG can each comprise a plurality of cells, for example, several thousand cells. The occupancy grid BG is in the Fig. 1 is temporarily stored in a data memory 2. The reference occupancy grid R-BG is also stored in a data memory 3 in the embodiment shown. Fig. In the embodiment shown in Figure 1, the occupancy grid is generated by a data processing unit 4, which evaluates sensor data supplied by a sensor 5. The sensor 5 can be a vehicle sensor of a vehicle, for example a radar sensor or a camera, which detects the surroundings of the vehicle. The reference occupancy grid R-BG can be stored on an exchangeable memory card. In one possible embodiment, the reference occupancy grid R-BG stored in the data memory 3 is received from outside via a communication interface 6 of the vehicle and stored in the data memory 3. The reference occupancy grid R-BG stored in the data memory 3 serves as a reference and, in one possible embodiment, is regarded as a valid, error-free occupancy grid.The occupancy grid BG and the reference grid R-BG contain a plurality of cells, each of which stores a probability value that indicates the probability that the respective cell is occupied. The probability values ​​lie between 0 and 1. The segmentation unit 1A of the calculation device 1 segments the reference occupancy grid R-BG, which is located in the data memory 3, and generates a first binarized grid b RBG therefrom by comparing it with a threshold value SW. Furthermore, the segmentation unit 1A also segments the motion grid BG to be evaluated, which is located in the data memory 2, and generates a second binarized grid b BG by comparing it with the adjustable threshold value SW. The two binarized grids orBinary grids contain only two logical values: a logically high value of 1 if the probability value P of the cell is above the threshold value SW, and a logically low value of 0 if the probability value P of the occupancy grid BG is below the threshold value SW. Segmentation is thus performed by threshold comparison. The threshold value SW is preferably increased in steps. The two binarized grids b RBG, b BG are transmitted from the segmentation unit 1A to the comparison unit 1B of the calculation device 1, as shown in FIG. Fig. 1. In one possible embodiment, the comparison unit 1B contains a comparator that compares the two binarized grids cell by cell to generate an error map FK, which indicates an error of the occupancy grid BG to be evaluated with respect to the reference grid for each cell. In one possible embodiment, the two binarized grids are subtracted from each other cell by cell to generate the error map FK. The generated error map FK is fed to a transformation unit 1C of the calculation device 1, as shown in Fig. 1. The transformation unit 1C performs a distance transformation DT on the error map FK to generate an error deviation map FAK. The generated error deviation map FAK indicates a geometric deviation of the occupancy grid BG to be evaluated from the reference occupancy grid R-BG. The geometric accuracy is determined from the error map FK using distance transformation. Fig. 4A, Fig. 4B shows an example of a distance transformation of an error map FK ( Fig. 4A) to generate an error deviation map FAK ( Fig. 4B). The error map FK, which is supplied by the comparison unit 1B, contains a logical zero for error-free cells, where the two binarized grids b RBG, b RG match, and a logical 1 for possibly faulty cells, where the two binarized grids do not match. Fig. 4A, Fig. In the example shown in Figure 4B, a faulty cell in the center of a region consisting of faulty cells has a displacement of three cells to the edge. The cells surrounding this cell have a displacement of two cells, and the outer edge of the region consisting of faulty cells of the error map FK has a displacement of only one cell. These displacements are shown in the Fig. 4B. The transformation unit 1C thus provides an error-deviation map FAK, as shown for example in Fig. 4B. This error-deviation map indicates a geometric deviation of the occupancy grid BG to be evaluated from the reference occupancy grid R-BG.

[0031] The generated error-deviation map FAK is transmitted from the transformation unit 1C to an evaluation unit 1D of the calculation device 1. The evaluation unit 1D evaluates the generated error-deviation map FAK to determine the accuracy G of the occupancy grid BG to be evaluated. The error-deviation map FAK is preferably statistically evaluated by the evaluation unit 1D. For example, the evaluation unit 1D can calculate a mean deviation by summing the values ​​of the error-deviation map FAK and dividing them by the total number of corresponding cells. A mean deviation can thus be calculated for each occupancy grid BG of a sequence and for each threshold. The evaluation unit 1D can thus perform a mean calculation or, alternatively, calculate a median as a location parameter as well as a variance or interquartile range for the dispersion of the distribution.Using the cell size, the obtained key figure can be additionally converted into an error with the unit length by the evaluation unit 1D. For example, a key figure of 0.25 with a cell size of 8 cm corresponds to an average error of 2 cm.

[0032] The evaluation unit 1D of the calculation device 1 provides, for example, a key figure that reflects the accuracy G of the occupancy grid BG to be evaluated, which is temporarily stored in the data memory 2. In one possible embodiment of the calculation device 1 according to the invention, the calculation accuracy G is fed to a calibration unit 8, which recalibrates or readjusts the vehicle sensor 5 depending on the determined accuracy G of its occupancy grid BG. In this way, vehicle sensors 5 are recalibrated whose sensor data result in an occupancy grid BG whose determined accuracy G is not yet sufficient. In one possible embodiment, the vehicle sensor 5 is recalibrated by readjusting its orientation, position, or location on a vehicle body.The readjustment is preferably carried out until the occupancy grid BG based on the sensor data from vehicle sensor 5 matches a reference occupancy grid R-BG to such an extent that the calculated accuracy G of the occupancy grid BG to be evaluated lies above an accuracy threshold value G-SW. Alternatively, the recalibration can be performed by recalculating a transformation matrix between a coordinate system of the sensor and the coordinate system of the vehicle. The transformation matrix is ​​changed iteratively until the error falls below a threshold value. In the case of the . Fig. In the embodiment shown in Figure 1, the calculated accuracy value G, which is output by the evaluation unit 1D, can be compared with an accuracy threshold value G-SW using a comparator 9. If the calculated accuracy value G exceeds the accuracy threshold value G-SW, the readjustment or recalibration by the calibration unit 8 is terminated.

[0033] Fig. 2 shows a flowchart illustrating an embodiment of the method according to the invention for determining an accuracy of an occupancy grid BG to be evaluated for a driver assistance system of a vehicle.

[0034] In a first step S AThe occupancy grid BG to be evaluated is segmented by at least one reference occupancy grid R-BG to generate binarized grids. Segmentation is preferably performed by threshold comparison, taking into account a plurality of different, gradually increasing threshold values ​​SW. In the first step S Athe reference occupancy grid R-BG and the occupancy grid BG to be evaluated are compared with a threshold value SW, resulting in two binarized grids. The two binarized grids depend on the respective threshold value SW. Since the binarized grids b BG, b RBG depend on the selected threshold value SW, a single threshold or threshold value SW can always lead to a randomly "good" or "bad" result, so that a statement that can be transferred to other threshold values ​​is not possible on this basis alone. In order to be able to make a systematic statement about the accuracy G of the occupancy grid BG to be evaluated rather than a random one, the threshold value formation and comparison are carried out for a large number of threshold values, for example in a range from 0 to 100% in 1% steps.In one possible embodiment, instead of covering the entire possible probability range between 0 and 100%, a restriction to a reasonable selected probability range is also applied. For example, the thresholds are set in a selected probability range of 50 to 90%, provided that the thresholds of the applications under investigation are within this range.

[0035] After segmenting the occupancy grid BG to be evaluated and at least one reference occupancy grid R-BG, the generated binarized grids are binarized in step S B Compared cell by cell to generate an error map FK. The binarized grids represent binary grids, whereby one can distinguish between a binary free map, i.e., each cell is free or not free, and an occupied map, i.e., each cell is occupied or not occupied. The segmentation in step S Aand the further steps are preferably carried out for each set threshold value SW, ie the method steps are preferably carried out iteratively at gradually increasing threshold values.

[0036] In step S B The binarized grids are compared with each other, and a binary error map FK is generated. In this error map FK, each cell is marked with either an error, e.g., logic 1, or no error, e.g., logic 0. By adding all cells of the error map FK, the total number of all faulty or deviating cells is obtained, which can also be related to the total number of free or occupied cells. However, this calculated key figure does not yet take into account how the faulty cells affect the geometry of the structure represented by the occupancy grid BG or occupancy grid. Therefore, in the method according to the invention, in the next step SC a distance transformation DT is applied, which is exemplified in the Fig. 4A, Fig. 4B. The Fig. The input image shown in Figure 4A represents an error map FK, where logical 0 stands for error-free cells and logical 1 for faulty cells. After the distance transformation DT has been performed, an error deviation map FAK results, as shown in Fig. 4B. After the distance transformation DT in step S CThis results in an error-deviation map FAK, which describes the geometric deviation of an occupancy grid BG to be evaluated from a reference occupancy grid R-BG resulting from the threshold comparison. To obtain an average deviation in cells, in one possible embodiment, the values ​​of the error-deviation map FAK are summed in step S4 and divided by the total number of corresponding cells. For each occupancy grid BG and for each threshold SW, an average deviation can thus be calculated, which can be statistically summarized, for example, by averaging or by calculating a median. The evaluation of the generated error-deviation map to determine the accuracy of the occupancy grid BG to be evaluated takes place in step S D .

[0037] Fig. Figure 3 shows a flowchart illustrating another embodiment of the method according to the invention for determining an accuracy G of an occupancy grid BG to be evaluated. After a start step S1, the occupancy grid BG is read out, for example, from a data memory in a step S2. In one possible embodiment, the occupancy grid BG to be evaluated and one or more reference occupancy grids R-BG are read out.

[0038] In step S3, the threshold value SW is initialized to an initial value (SW = SW0). Subsequently, the occupancy grids BG, RBG are segmented by comparing the set threshold value SW with the threshold value to obtain corresponding binary grids b BG , b RBG .

[0039] In a further step S5, the binary grids b BG, b RBG are subtracted from each other to generate an error map FK.

[0040] The generated error map FK is subjected to a distance transformation DT (FK) in step S6 to generate an error deviation map FAK.

[0041] In a further step S7, a mean value M of the error deviation map FAK for the currently set threshold value SW i calculated (M SWi (FAK)).

[0042] This mean value M SWi the error deviation map FAK is temporarily stored in step S8.

[0043] In a further step S9 it is checked whether the threshold value M SWi already a maximum threshold SW MAX If this is not the case, the currently set threshold value SW i in step 10 gradually increased to the next threshold value (SW i+1 :=SW i +ΔSW). The process then returns to the segmentation step S4. If the maximum threshold value SW MAX reached in step S9, the Fig. 3 illustrated embodiment in step S11 the maximum of all intermediately stored mean values ​​M SWi is determined, i.e., the largest error. This value is preferably compared with an error threshold in step S12. If the largest error determined in step S11 is below the error threshold, the occupancy grid BG to be examined is recognized as sufficiently accurate in step S13. Conversely, if it is determined in step S12 that the largest error occurred is above the error threshold, the occupancy grid BG to be examined is assessed as insufficiently accurate, and if necessary, a recalibration of the sensor whose sensor data forms the basis for the occupancy grid BG to be examined is performed.

[0044] The method and device according to the invention yield a key figure for an occupancy grid BG to be examined, which takes into account the influence of the threshold value generation or threshold value comparison and establishes a direct relationship to the geometric accuracy of the physical structure represented by the occupancy grid BG. By considering a plurality of threshold values ​​SW for segmenting the occupancy grid BG and by creating and converting an error map FK into distances to error-free areas, the method according to the invention thus yields a key figure that is independent of the specific threshold or threshold value and correlates with the geometric accuracy of the structure represented by the occupancy grid BG.

[0045] The method according to the invention can preferably be used in an authenticity system, in particular a driver assistance system. In one possible embodiment, the accuracy G of an occupancy grid BG to be evaluated is determined during a design phase of the real-time system. In an alternative embodiment, the method for determining an accuracy G of an occupancy grid BG to be evaluated is carried out during normal operation of the respective system. In this case, the accuracy G of the occupancy grid BG to be evaluated can be determined in the background by the calculation device 1. In one possible embodiment, the calculation device 1 generates a warning message if the accuracy G of the occupancy grid BG to be evaluated is too low and cannot be sufficiently increased even by readjusting the associated sensor.

[0046] In another possible embodiment, several occupancy grids BG, which reflect the same surrounding area of ​​the vehicle, can be evaluated with regard to their accuracy G using the method according to the invention, wherein the remaining occupancy grids BG represent a reference occupancy grid R-BG. An occupancy grid BG that deviates too significantly from the other occupancy grids or whose accuracy is too low compared to the other occupancy grids can be rejected in a possible embodiment variant or disregarded in the further data evaluation. In this embodiment variant, the integrity of the occupancy grid BG to be evaluated can thus be verified.

Claims

[1] Method for determining an accuracy (G) of an occupancy grid (BG) to be evaluated for a driver assistance system of a vehicle, comprising the following process steps: (a) Segmentation (S A ) the occupancy grid (BG) to be evaluated and at least one reference occupancy grid (R-BG) for generating binarized grids (b BG, b RBG); (b) cell-by-cell comparison (S B ) the binarized grids to generate an error map (FK) which indicates for each cell an error of the occupancy grid (BG) to be evaluated with respect to the reference occupancy grid (R-BG); (c) Carry out (S C ) a distance transformation on the error map (FK) to generate an error deviation map (FAK) which indicates a geometric deviation of the occupancy grid (BG) to be evaluated from the reference occupancy grid (R-BG); and (d) Evaluate (S D) of the generated error-deviation map (FAK) to determine the accuracy (G) of the occupancy grid (BG) to be evaluated. [2] Method according to claim 1, wherein the occupancy grid (BG) to be evaluated and the reference occupancy grid (R-BG) have cells each containing a probability value indicating the probability that the respective cell is occupied. [3] Method according to claim 2, wherein the probability values ​​contained in the cells of the occupancy grid (BG) to be evaluated and of the reference occupancy grid (R-BG) are compared with an adjustable threshold value for generating the binarized grids. [4] Method according to one of the preceding claims 1 to 3, wherein sensor data originating from at least one vehicle sensor (5) mounted on the vehicle are evaluated in order to generate the occupancy grid (BG) to be evaluated. [5] Method according to claim 4, wherein the occupancy grid (BG) to be evaluated generated from sensor data is temporarily stored in a data memory (2). [6] Method according to one of the preceding claims 1 to 5, wherein the binarized grids are subtracted to generate the error map (FK). [7] Method according to claim 3, wherein the method steps are carried out iteratively with gradually increasing threshold values. [8] Method according to claim 7, wherein the set threshold values ​​lie in a selected probability range. [9] Method according to one of the preceding claims 1 to 8, wherein the values ​​contained in the cells of the error deviation maps (FAK) are statistically evaluated to determine the accuracy (G) of the occupancy grid (BG) to be evaluated. [10] Method according to one of the preceding claims 4 to 9, wherein a vehicle sensor (5) is recalibrated, the sensor data of which result in an occupancy grid (BG) whose determined accuracy (G) is insufficient. [11] Method according to claim 10, wherein the vehicle sensor (5) is recalibrated by readjusting its orientation, position or attitude on the vehicle body or by changing a transformation matrix for transforming from a sensor coordinate system to a vehicle coordinate system. [12] Method according to one of the preceding claims 1 to 11, wherein the reference occupancy grid (R-BG) is received via a communication interface of the vehicle and is temporarily stored in a data memory (3). [13] Method according to one of the preceding claims 1 to 11, wherein the reference occupancy grid (R-BG) is read from an exchangeable data memory. [14] Driver assistance system for a vehicle with: (a) a segmentation unit (1A) which segments an occupancy grid (BG) to be evaluated and a reference occupancy grid (R-BG) to generate binarized grids (b BG, b RBG); (b) a comparison unit (1B) which compares the binarized grids cell by cell to generate an error map (FAK) which indicates for each cell an error of the occupancy grid (BG) to be evaluated with respect to the reference grid (R-BG); (c) a transformation unit (1C) which performs a distance transformation on the error map (FK) to generate an error deviation map (FAK) which indicates a geometric deviation of the occupancy grid (BG) to be evaluated from the reference occupancy grid (R-BG); and (d) an evaluation unit (1D) which evaluates the generated error deviation map (FAK) to determine the accuracy (G) of the occupancy grid (BG) to be evaluated. [15] Driver assistance system according to claim 14, wherein a data processing unit (4) is provided which evaluates sensor data originating from at least one vehicle sensor (5) of the vehicle in order to generate the occupancy grid (BG) to be evaluated. [16] Driver assistance system according to claim 15, wherein vehicle sensors (5) whose sensor data result in an occupancy grid (BG) whose accuracy (G) determined by the evaluation unit (1D) is insufficient can be recalibrated by a calibration unit (8) of the vehicle.

Images