Classification model based on fuzzy logic

Abstract

The invention relates to a method for classifying at least one object (O), in particular a computer-implemented method, to a method for training a fuzzy logic model (FM) using at least one membership function (f) for classifying at least one object (O), to a computer program for carrying out the method according to the invention, and to a method for training a fuzzy logic model (FM) using at least one membership function (f) for classifying at least one object (O). And a computer program product. The classification method comprises the following method steps:-providing input data (I (O)) containing at least one item of information about at least one object (O),-providing at least the input data (I (O)) as an input to a trained fuzzy logic model (FM), the fuzzy logic model (FM) is trained using at least one membership function (f) of at least one predefinable category (K) of the fuzzy logic model (FM), the fuzzy logic model (FM) being designed to determine a probability (PK (O)) of the at least one object (O) belonging to the at least one predefined category (K) on the basis of the input, and outputting a classification result (PK (O)).

Description

Technical Field

[0001] The present invention relates to a method, in particular a computer-implemented method, for classifying objects by means of a trained fuzzy logic model, a method, in particular a computer-implemented method, for training a fuzzy logic model, a computer program for executing the method according to the invention, and a computer program product on which the method is stored. Background Art

[0002] For object classification, particularly in the field of driver assistance systems, input data from sensors (e.g., radar sensors, ultrasonic sensors, lidar sensors, or cameras) is checked for the presence of various predefinable features and, if necessary, assigned to specific object classes. Many different methods are known in this regard, particularly in the field of machine learning. For example, object class assignment is performed using activation functions that can be used to define specific value ranges for detected features.

[0003] Regarding methods from machine learning, the following possibilities are also increasingly being discussed: considering background information about the system under investigation and simple logical statements, or introducing symbolic methods (such as those from the field of fuzzy logic). Fuzzy logic models are commonly used in control systems and are mainly used to formalize background knowledge and take uncertainty into account. However, the introduction of fuzzy logic models in machine learning methods raises various problems based on differentiability and associated with the gradient vanishing problem, such as "Analyzing Differentiable Fuzzy Logic Operators" (arXiv:2002.06100 (DOI: https: / / doi.org / 10.1016 / j.artint.2021.103602 )).

[0004] To overcome the obstacles associated with the vanishing gradient problem, the use of logic tensor networks (LTNs) has been proposed. In this case, membership functions are determined using a neural network with a sigmoid activation function. While this approach provides continuously differentiable membership functions, it disadvantageously comes with a significant computational effort and low interpretability.

[0005] Furthermore, US Pat. No. 5,761,326 A discloses a method for classifying and tracking objects. Each pixel in a camera image is analyzed for the intensity of horizontal and vertical edge elements, and a vector comprising an amplitude and an angle is calculated for each pixel. Fuzzy set theory is then applied to the vectors within a predefinable region to generate a unique vector representing the region. This vector is then interpreted by a neural network to perform object classification. Consequently, this method is time-consuming. Summary of the Invention

[0006] It would be desirable to have an object classification method that is suitable for embedded systems with limited computing power but still capable of classifying a large number of objects.

[0007] This object is achieved by a method as claimed in claim 1 , a computer program as claimed in claim 14 and a computer program product as claimed in claim 15 .

[0008] In terms of the method, the object of the invention is achieved by a method, in particular a computer-implemented method, for classifying at least one object, comprising the following method steps:

[0009] - providing input data comprising at least one item of information about the at least one object,

[0010] - providing at least the input data as input to a trained fuzzy logic model, the fuzzy logic model being trained using at least one membership function of at least one predefinable class of the fuzzy logic model, wherein the fuzzy logic model is designed to determine a probability that the at least one object belongs to the at least one predefined class based on the input, and

[0011] - Output classification results.

[0012] That is, according to the present invention, membership functions are determined for different classes and / or characteristic features within a class, and classification is performed directly using a fuzzy logic model. For example, the membership functions can be defined at least in part based on prior knowledge about the respectively selected class and / or feature. However, the membership functions can also be determined at least in part using machine learning methods. For classification, the object can then be assigned, for example, to the predefined class with the highest probability of belonging.

[0013] Advantageously, relatively low computing power is required for the process according to the invention. The method is therefore very suitable for embedded systems and can still classify a large number of different objects in each acquisition cycle for acquiring / detecting input data.

[0014] Within the scope of the present invention, a fuzzy logic model is understood to be a model based on fuzzy set theory and / or fuzzy logic. Fuzzy logic is based on fuzzy sets, the elements of which belong to the set to a certain extent. The assignment to the fuzzy set is performed based on one or more membership functions that can be combined to form a fuzzy function.

[0015] In one advantageous embodiment, the membership functions used are piecewise linear functions or piecewise constant functions, in particular step functions or trapezoidal functions. Compared to other types of functions, such membership functions advantageously reduce the computational effort. Furthermore, the classification results obtained using such functions are particularly easy and accurate to interpret, and both linear and nonlinear behavior can be considered. Different nonlinear object variables can be distinguished. Furthermore, object classification requires less computation time.

[0016] It is also advantageous if the membership function is described based on predefinable function parameters, which are parameters of the fuzzy logic model. If multiple membership functions are used, they can also be combined with one another. For example, based on multiple membership functions, the probability that at least one object belongs to a predefined class can be determined using the sum, maximum, minimum, or product of the different function parameters.

[0017] According to an advantageous embodiment of the method, at least two intervals are predefined for at least one function variable of the membership function, wherein a membership subfunction is determined for each interval, wherein the membership function is composed of membership subfunctions. For example, one or more intervals may be specified, within which the membership function takes one or more constant values, for example, associated with different probabilities. Preferably and particularly advantageously, each membership subfunction is a linear function or a constant function, in view of the computational effort.

[0018] In this respect, it is advantageous if each interval is characterized by at least one interval value, which is in particular a start value or an end value of the interval.

[0019] According to a further embodiment of the method, the input data are sensor data from a sensor, in particular a radar sensor or a camera. However, input data from other suitable sensors can also be used.

[0020] It is also advantageous if the at least one item of information about the object is a geometric variable related to the object, in particular a geometric dimension such as the length, width or area of ​​the object, or a position, a velocity, an acceleration or a radar cross section.

[0021] The object of the present invention is also achieved by a method, in particular a computer-implemented method, for training a fuzzy logic model using at least one membership function for classifying at least one object, comprising the following method steps:

[0022] - providing training data comprising at least one item of information about the at least one object,

[0023] - providing, for all training data, an indication relating to the category of the at least one object,

[0024] - generating outputs of the fuzzy logic model for all training data by processing the training data using the fuzzy logic model according to parameters of the fuzzy logic model,

[0025] - comparing the generated output to the indication using at least one error function, and

[0026] - Optimize parameters based on this comparison.

[0027] A fuzzy logic model is preferably used in the method of the present invention for classifying at least one object according to one of the previously described embodiments. The output of the fuzzy logic model is preferably a probability that the at least one object belongs to at least one category. In the case of multiple characterizing features, the classification result can also be determined by combining the individual probabilities (e.g., by summing or multiplying them). For assignment to one of the multiple categories, the category with the highest probability is output. This output, preferably determined using an argmax operator, is then the classification result.

[0028] The error function may be given, for example, by the number of incorrect classifications, or the difference between the determined probability or classification result and the indication. For example, the probability may be normalized to have a value between 0 and 1. Preferably, the error function is minimized during the training method. The training method may be initiated or executed several times.

[0029] Regarding the method for training the fuzzy logic model, it is advantageous to optimize the parameters of the fuzzy logic model based on a gradient method or based on a steepest descent method. In this case, a random initialization can be advantageously chosen.

[0030] It is also advantageous to perform regularization. In this regard, a regularization term can be added to the error function. Regularization ensures that the output of the fuzzy logic model (particularly the probability or classification result) is greater than zero and as small as possible. Thus, for example, for scenarios where little training data is available, lower values ​​or probabilities are typically output. Furthermore, regularization ensures that the starting and ending values ​​of the membership function intervals lie within the value range of the fuzzy logic model inputs. This eliminates the need to define unnecessary membership sub-functions, for example.

[0031] In a further embodiment of the training method, at least two intervals are predefined for at least one function variable of the membership function, wherein a membership subfunction is determined for each interval, wherein the membership function is composed of the membership subfunctions.

[0032] A further embodiment includes each interval being characterized by at least one interval value, wherein if the value of the nth interval value (particularly within a training cycle) is greater than the value of the n-1th interval value, the values ​​of the nth interval value and the n-1th interval value are swapped before evaluating the parameters. In contrast, the function values ​​of the corresponding membership subfunctions are not swapped. If the interval value is the starting value and / or ending value of the interval, for example, if the starting value is greater than the ending value within a training cycle, the starting and ending values ​​are swapped. In this way, the error function is mirrored at the boundaries of the search space. Thus, interruptions to the training method due to invalid parameter values ​​can be prevented.

[0033] The method for classifying at least one object according to the present invention is preferably used for object classification, in particular within driver assistance systems. The method according to the present invention is particularly preferably used for classifying objects using radar sensors or cameras. It is also advantageous to use the method for classifying object types (e.g., different types of vehicles, such as trucks or buses).

[0034] The present invention also relates to a driver assistance system that is designed to implement the method according to the invention for classifying at least one object. In the case of a driver assistance system, the object is, for example, another road user, in particular a vehicle or pedestrian, or a road boundary or other object within the corresponding detection range of the sensor used.

[0035] The object of the invention is also achieved by a computer program having instructions which, when executed by a computer, cause the computer to carry out the classification method according to the invention according to one of the described embodiments or the method according to the invention for training a fuzzy logic model.

[0036] The objects of the invention are also achieved by a computer program product on which the computer program according to the invention is stored.

[0037] It should be mentioned that what relates to the method according to the invention for classifying at least one object also applies mutatis mutandis to the training method according to the invention, the computer program according to the invention and the computer program product according to the invention, and vice versa. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention and its advantageous embodiments will be explained in more detail with reference to the following drawings. In the drawings:

[0039] Figure 1 An example of a piecewise constant membership function composed of multiple membership sub-functions is shown;

[0040] Figure 2Another example of a membership function is shown;

[0041] Figure 3 A flow chart showing a design of a method for classifying objects according to the present invention; and

[0042] Figure 4 A flow chart showing a design solution of the training method according to the present invention is shown.

[0043] In the figures, identical elements are provided with the same reference numerals. DETAILED DESCRIPTION

[0044] exist Figure 1 The membership function f is plotted in the form of a piecewise linear function or a piecewise constant function (here in the form of a trapezoidal function). The trapezoidal function f can be based on the a to t c The three membership subfunctions f a to f c To define, in each interval, the membership function f is a constant. In the constant interval t a to t c In other designs, the interval t a to t c They can also be adjacent, or the interval can be defined as a range where the membership function f is not a constant.

[0045] Three intervals t a to t c It can also be characterized based on the four interval values ​​t1-t4 of the function variable t, each of which is an interval t a to t c The example shown here is: t a : [-∞; t1], t b :[t2; t3], t c : [t4; ∞]. For each interval t a to t c , and also define the probability p a to p c .

[0046] A different number of intervals and / or a different number of interval values ​​can be used to characterize other membership functions f. Figure 2 In the figure, a graph with three constant ranges t is depicted. a to t c Another example of the membership function f in the form of a step function.

[0047] exist Figure 3A flow chart of an advantageous embodiment of the method according to the invention for classifying objects is shown in FIG. In a first method step, input data I(O) are provided, which contain at least one item of information about at least one object O to be classified. The input data I(O) are preferably measurement data from one or more sensors S (e.g., radar sensors or cameras). The input data I(O) serve as input to a trained fuzzy logic model FM, which determines the probability P of belonging to at least one predefinable class K. K (O) and output the corresponding classification results.

[0048] The fuzzy logic model FM is based on at least one membership function f for at least one predefinable class K. The membership function f is preferably based on predefinable function parameters (eg t1 to t4, p a to p c ) to describe, these predefined function parameters are also the parameters of the fuzzy logic model FM.

[0049] In the case where there are multiple categories K, determine the probability P for each category K K (O). In this case, the probability P K (O) The highest category can be output as a classification result, for example.

[0050] Finally, in Figure 4 In the embodiment of the invention, an advantageous embodiment of the training method for training a fuzzy logic model FM using at least one membership function f for classifying at least one object O is described. K The training dataset T D (O) Start, according to the parameters t1 to t4, p a to p c To generate the output PK(O) of the fuzzy logic model. The output PK(O) is respectively related to the indicator I when the error function Error is used. K Compare and optimize the parameters t1 to t4, p based on the comparison a to p c .

[0051] For parameters t1 to t4, p a to p c The optimization of can be performed, for example, by a gradient method or based on the steepest descent method. In addition, a regularization R can be optionally performed. In addition, if the nth interval value t in the optimization cycle n The value of is greater than the n-1th interval value t n-1 , then in the optimization parameters t1 to t4, p a to pc Before exchanging the nth interval value t n and the ntth interval value t n-1 This does not affect the probability p a to p c This means that if in the training cycle, for example, the second interval t b If the starting value t2 is greater than its ending value t3, the starting value t2 and the ending value t3 are swapped.

Claims

1. A method, in particular a computer-implemented method, for classifying at least one object (O), comprising the following method steps: - providing input data I(O), said input data comprising at least one item of information about the at least one object (O), - providing at least the input data (I(O)) as input to a trained fuzzy logic model (FM), the fuzzy logic model being trained using at least one membership function (f) of the fuzzy logic model (FM) for at least one predefinable class (K), wherein The fuzzy logic model (FM) is designed to determine a probability (P) that the at least one object (O) belongs to at least one predefined category (K) based on the input. K (O)), and - Output classification results (P K (O)).

2. The method according to claim 1, in, The membership functions (f) used are piecewise linear functions or piecewise constant functions, in particular step functions or trapezoidal functions.

3. The method according to claim 1 or 2, in, The membership function (f) is a function that can be predefined with the help of function parameters (t1-t4; p a -p c ) to describe these function parameters (t1-t4; p a -p c ) are the parameters of the fuzzy logic model (FM).

4. The method according to claim 1, in, At least two intervals (t a -t c ), where for each interval (t a -t c ), determine the membership function (f a -f c ), the membership function (f) is composed of these membership sub-functions (f a -f c )composition.

5. The method according to claim 4, in, Each interval (t a -t c ) is characterized by at least one interval value (t1-t4), the interval value (t1-t4) being in particular the interval (t a -t c )'s starting or ending value.

6. The method according to claim 1, in, Input data (I(O)) are sensor data from a sensor (S), in particular a radar sensor or a camera.

7. The method according to claim 1, in, The at least one item of information about the object (O) is a geometrical variable, a position, a velocity, an acceleration or a radar cross section of the object, wherein the geometrical variable is in particular a geometrical size.

8. A method, in particular a computer-implemented method, for training a fuzzy logic model (FM) using at least one membership function (f) for classifying at least one object (O), the method comprising the following method steps: - Provide training data (D T (O)), the training data comprises at least one item of information about the at least one object (O), - For all training data (D T (O)) provides an indication (I) related to a category (K) of the at least one object (O) K ), - By using the fuzzy logic model (FM) according to the parameters (t1-t4; p a -p c ) Processing training data (D T (O)), for all training data (D T (O)) generates the output (P) of the fuzzy logic model (FM) K (O)), - The generated output (P) is converted to K (O)) and the instructions (I K ) for comparison, and - Based on this comparison, optimize these parameters (t1-t4; p a -p c ).

9. The method according to claim 8, in, These parameters (t1-t4; p a -p c ) is optimized based on the gradient method or the steepest descent method.

10. The method according to at least one of claims 8 and 9, in, Perform regularization (R).

11. The method according to any one of claims 8 to 10, in, At least two intervals (t a -t c ), where for each interval (t a -t c ) Determine the membership function (f a -f c ), the membership function (f) is composed of these membership sub-functions (f a -f c )composition.

12. The method according to claim 11, in, Each interval (t a -t c ) is characterized by at least one interval value (t1-t4), the interval value (t1-t4) being in particular the interval (t a -t c ) of the starting or ending value, in the nth interval value (t n ) is greater than the value of the n-1th interval (t n-1 ) values, in optimizing these parameters (t1-t4; p a -p c ) before exchanging the nth interval value (t n ) and the n-1th interval value (t n-1 ) value. 13 . Use of the method according to claim 1 for object classification, in particular within a driver assistance system.

14. A computer program having instructions which, when executed by a computer, cause the computer to carry out the method according to one of claims 1 to 7 or the method according to one of claims 8 to 12.

15. A computer program product having stored thereon the computer program according to claim 14.

Images