Estimation device, estimation method, storage medium, vehicle control device, and vehicle

Abstract

An estimation device is provided with: a correction unit that inputs target data into a machine learning model that has been learned so as to output a logit that at least a portion of the data belongs to a category representing a certain category by using data as an input, acquires the logit, and corrects the logit by using the output of the machine learning model; calculating a correction value for correcting the logit according to the correction value; and an estimation unit that corrects the logit on the basis of the calculated correction value, and estimates, on the basis of the corrected logit, a category to which the at least part of the target data belongs.

Description

Technical Field

[0001] The present invention relates to an estimation device, an estimation method, a storage medium, a vehicle control device, and a vehicle. Background Art

[0002] In the learning of a machine learning model that takes an image as input and classifies the types of objects included in the image, a technique for preventing a decrease in the estimation accuracy of the model due to the imbalance between classes associated with the number of teaching data (samples) is conventionally known. For example, in Menon et al., “Long-tail learning via logit adjustment,” a technique is described in which the weights of the error functions used in learning (the weights of the error functions corresponding to each class) and the logit (the output value of the machine learning model) are corrected according to the number of samples corresponding to each class.

[0003] The technique described in Menon et al., “Long-tail learning via logit adjustment” uses the total number of samples in the teaching data and the number of samples in each class to correct the logit. However, if such a simple method is adopted, the logit may not be corrected with high accuracy, and as a result, the accuracy of the generated machine learning model becomes low. Summary of the Invention

[0004] The present invention has been made in view of such circumstances, and one of its objects is to provide an estimation device, an estimation method, a storage medium, a vehicle control device, and a vehicle that can improve the accuracy of a generated machine learning model by accurately correcting the logit output by the machine learning model.

[0005] The estimation device, the estimation method, the storage medium, the vehicle control device, and the vehicle according to the present invention adopt the following configuration.

[0006] (1): One aspect of the present invention relates to an estimation device, wherein the estimation device includes: a correction unit that inputs object data to a machine learning model that has been learned to output a logit in which at least a part of the data belongs to a class representing a certain type, obtains the logit, and calculates a correction value for correcting the logit using the output of the machine learning model; and an estimation unit that corrects the logit based on the calculated correction value and estimates the class to which at least a part of the object data belongs based on the corrected logit.

[0007] (2): Based on the solution in (1) above, the correction unit calculates the correction value using preset hyperparameters and the output of the machine learning model.

[0008] (3): Based on the solution in (2) above, the correction unit calculates the correction value by multiplying the hyperparameters by the prior probability of the samples of the category, and the prior probability is calculated according to the surrounding distribution defined by the machine learning model.

[0009] (4): Based on the solution in (1) above, the estimation unit corrects the logit by adding the correction value to the logit, and estimates the category to which at least a part of the object data belongs as the category with the maximum probability value obtained based on the corrected logit.

[0010] (5): Based on the solution in (1) above, the data is an image containing a plurality of pixels, and at least a part of the data is one or more pixel groups in the image.

[0011] (6): Based on the solution in (1) above, the data is sound, and at least a part of the data is an interval of the sound.

[0012] (7): Another solution of the present invention relates to a vehicle control device, wherein the vehicle control device includes: the estimation device according to the solution in (1) above; and a vehicle control unit that controls the running of the vehicle based on the estimation result estimated by the estimation device.

[0013] (8): Another solution of the present invention relates to a vehicle that includes the vehicle control device according to the solution in (7) above.

[0014] (9): Another solution of the present invention relates to an estimation method, wherein the estimation method causes a computer to perform the following processing: inputting object data into a machine learning model that has been learned to output a logit indicating that at least a part of the data belongs to a category representing a certain type to obtain the logit, and using the output of the machine learning model to calculate a correction value for correcting the logit; and correcting the logit based on the calculated correction value, and estimating the category to which at least a part of the object data belongs based on the corrected logit.

[0015] (10): Another aspect of the present invention relates to a storage medium storing a program, wherein the program causes a computer to perform the following processing: inputting object data into a machine learning model that has been learned in such a way that at least a part of the data output as an input belongs to the logit of a class representing a certain category to obtain the logit, and using the output of the machine learning model to calculate a correction value for correcting the logit; and correcting the logit based on the calculated correction value and estimating the category to which at least a part of the object data belongs based on the corrected logit.

[0016] According to the aspects (1) to (10), it is possible to improve the accuracy of the generated machine learning model by accurately correcting the logit output by the machine learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a diagram showing the structure of the estimation device according to the first embodiment.

[0018] Figure 2 It is a diagram showing an example of the structure of learning data.

[0019] Figure 3 It is a diagram for explaining an outline of machine learning for generating a machine learning model based on learning data.

[0020] Figure 4 It is a flowchart showing an example of the process executed by the estimation device according to the first embodiment.

[0021] Figure 5 It is a diagram showing the structure of the estimation device according to the second embodiment.

[0022] Figure 6 It is a flowchart showing an example of the process executed by the estimation device according to the second embodiment.

[0023] Figure 7 It is a diagram showing the structure of the present vehicle equipped with a vehicle control device having an estimation device. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] [First Embodiment]

[0025] Hereinafter, embodiments of the estimation device, estimation method, storage medium, vehicle control device, and vehicle of the present invention will be described with reference to the drawings.

[0026] [Structure]

[0027] Figure 1FIG. 0 is a diagram showing the configuration of the estimation device 100 according to the first embodiment. The estimation device 100 is an information processing device that uses a machine learning model trained to output logits for one or more pixel groups in an image belonging to a class representing the type of an object, given the image including a plurality of pixels as input, to estimate the type information of the one or more pixel groups. The estimation device 100 includes, for example, a correction unit 110, an estimation unit 120, and a storage unit 130. The correction unit 110 and the estimation unit 120 are each implemented, for example, by a hardware processor such as a CPU (Central Processing Unit) executing a program (software). In addition, some or all of these components may be implemented by hardware (including circuitry such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), GPU (Graphics Processing Unit), SOC (System On Chip)), or may be implemented by the cooperation of software and hardware. The program may be pre-stored in a storage device (a storage device having a non-transitory storage medium) such as an HDD or a flash memory of the estimation device 100, or may be stored in a removable storage medium such as a DVD or a CD-ROM, and installed in the HDD or flash memory of the estimation device 100 by being mounted on a drive device through the storage medium (non-transitory storage medium). The storage unit 130 stores, for example, learning data 130A and a machine learning model 130B. The storage unit 130 is implemented, for example, by a RAM, a flash memory, an SD card, or the like.

[0028] Figure 2 FIG. 4 is an example diagram showing the configuration of the learning data 130A. In the learning data 130A, for example, for one or more pixel groups of an image IM, a class representing the type of the object shown by the pixel group is corresponded as the correct answer class. Here, the "type of the object" is the type of an object related to the surrounding environment of a vehicle, and includes, for example, at least two of a four-wheeled vehicle, a two-wheeled vehicle, a pedestrian, a falling object, a traffic signal, a road sign, a guardrail, a median strip, and a curb. In addition, the "type of the object" may be locked to a moving object, or at least two of a four-wheeled vehicle, a two-wheeled vehicle, and a pedestrian may be set as the type of the object.

[0029] For example, in Figure 2In this case, the image IM includes a pixel group P1 representing a four-wheeled vehicle and a pixel group P2 representing a two-wheeled vehicle. At this time, for example, the pixel group P1 is associated with a vector in which only the component representing the category of the four-wheeled vehicle is set to 1 and the components of other categories are set to 0, whereby learning data 130A (hereinafter sometimes referred to as a sample) with the four-wheeled vehicle category as the correct answer category can be obtained. Additionally, for example, the pixel group P2 is associated with a vector in which only the component representing the category of the two-wheeled vehicle is set to 1 and the components of other categories are set to 0, whereby learning data 130A with the two-wheeled vehicle category as the correct answer category can be obtained. The storage unit 130 stores the correspondence between the above pixel group and the correct answer category as learning data 130A.

[0030] The learning data 130A is generated, for example, by the manager or operator of the estimation device 100 specifying each pixel of the image IM and the correct answer category in advance on their own terminal and stored in the storage unit 130. Alternatively, the estimation device 100 can download the learning data 130A stored in an external server to the storage unit 130 via a network at the timing of learning execution.

[0031] The machine learning model 130B is a machine learning model that takes an image containing multiple pixels as input and outputs the logit (accuracy) of one or more pixel groups in the image belonging to a category representing the type of object. Here, the logit is a value related to the probability that the pixel group of the image IM belongs to a category representing the type of object. The machine learning model 130B is, for example, a convolutional neural network (CNN: Convolution Neural Network), which extracts the feature amounts of the input pixels and outputs the logit of each pixel belonging to each category.

[0032] [Overview of Machine Learning]

[0033] Figure 3 This is a diagram for explaining the overview of machine learning for generating the machine learning model 130B based on the learning data 130A. In Figure 3 , C represents the total number of categories to be classified, and Z0~Z C-1 represents the logit output by the machine learning model 130B. As Figure 3 shown, by substituting each output logit into the Softmax function (normalized exponential function), it is transformed into probability values P0~P C-1 normalized to values between 0 and 1, and an error function L such as the Cross Entropy error (cross-entropy error) is used to calculate the error between the transformed probability values and the correct answer vector, and a method such as the error backpropagation method is used to learn the machine learning model 130B in such a way that the calculated error is minimized.

[0034] Note that, in this embodiment, as an example, the multi-valued class classification of the pixels included in the image is described. Therefore, the Softmax function is applied in Figure 3 . However, the present invention is not limited to such a structure. In the case of binary classification of the pixels included in the image, the Sigmoid function may be applied instead of Figure 3 the Softmax function. Moreover, in this case, the Cross Entropy error L can also be calculated as the Binary Cross Entropy error L instead of the Softmax Cross Entropy error.

[0035] The machine learning model 130B obtained by learning as described above does not consider the imbalance between classes related to the number of samples for the learning data 130A prepared for each class. As a result, the machine learning model 130B learns in such a way that in the learning process, the higher the number of samples in a class, the higher the logit, that is, the higher the probability value is output. Sometimes, the estimation accuracy of the model decreases.

[0036] [Correction by the correction unit 110]

[0037] Against this background, the correction unit 110 adds correction terms Δ0 to Δ C-1 to the logits Z0 to Z C-1 output by the machine learning model 130B after learning is completed, and calculates the corrected logits Z’0 to Z’ C-1 . The estimation unit 120 substitutes the corrected logits Z’0 to Z’ C-1 for the logits Z0 to Z C-1 in the Softmax function at the estimation stage, and thus calculates the corrected probability values P’0 to P’ C-1 . The estimation unit 120 estimates the class corresponding to the highest probability value among the calculated probability values P’0 to P’ C-1 as the object included in the input image.

[0038] More specifically, the correction unit 110 calculates the corrected logit Z’ according to the following formula (1) i (i is an arbitrary integer from 0 to C - 1).

[0039] Z’ i = Z i - α log P(y i ) ··· (1)

[0040] In formula (1), α represents a hyperparameter set in advance, and P(y irepresents the prior probability of the occurrence of samples of class i in the entire sample of the learning data 130A. The hyperparameter α is adjusted (tuned) using validation data different from the learning data 130A when calculating the corrected logitZ'. When P(x, y) is set as the joint probability, P(y) is calculated according to the following equation (2). i When calculating the corrected logitZ', the hyperparameter α is adjusted (tuned) using validation data different from the learning data 130A. When P(x, y) is set as the joint probability, P(y) is calculated according to the following equation (2).

[0041] ···(2)

[0042] In Equation (2), P(y|x) represents the conditional probability of the occurrence of samples of class y on the premise of the occurrence of samples of class x (where x represents any class other than class y in the learning data 130A), and can be calculated using the output of the learned machine learning model 130B (more specifically, the marginal distribution defined by the probability values P0 to P output via the learned machine learning model 130B). C-1 defined marginal distribution).

[0043] Next, with reference to Figure 4 the process flow executed by the estimation device 100 according to the first embodiment will be described. Figure 4 is a flowchart showing an example of the process flow executed by the estimation device 100 according to the first embodiment.

[0044] First, the correction unit 110 obtains logitZ0 to Z by inputting the image to be estimated into the learned machine learning model 130B (step S100). Next, the correction unit 110 adds a correction term Δ C-1 reflecting the output of the machine learning model 130B to the obtained logitZ0 to Z C-1 =-αlogP(y i ) to calculate the corrected logitZ'0 to Z' i (step S102). Next, the estimation unit 120 substitutes the corrected logitZ'0 to Z' C-1 into the Softmax function, obtains and outputs the corrected probability values P'0 to P' C-1 (step S104). Thus, the process of this flowchart ends. C-1 (step S104). Thus, the process of this flowchart ends.

[0045] According to the first embodiment described above, different from the technique described in Non-Patent Document 1 in which the correction term is calculated only by addition, subtraction, multiplication, and division of the total number of samples of the teaching data and the number of samples of each class, the correction term Δ in Equation (1) i =-αlogP(y i ) the prior probability P(y i)Calculated using the output of the learned machine learning model 130B. That is, according to the first embodiment, the logit is corrected with higher precision, and the accuracy of the generated machine learning model can be improved.

[0046] [Second Embodiment]

[0047] In the first embodiment, by adding a correction term calculated using the output of the learned machine learning model 130B to the logit output by the machine learning model 130B, a decrease in the estimation accuracy of the model caused by the imbalance between classes related to the number of samples is prevented. As a second embodiment, it may also be that during the learning process of the machine learning model 130B, the error function L corrected based on the output of the machine learning model 130B obtained by preliminary learning is used for re-learning the machine learning model 130B to prevent a decrease in the estimation accuracy of the model.

[0048] Figure 5This is a diagram showing the structure of the estimation device 200 according to the second embodiment. Similar to the first embodiment, the estimation device 200 is an information processing device that uses a machine learning model trained to output the logit of one or more pixel groups in an image belonging to a category representing the type of an object, given the image including a plurality of pixels as input, to estimate the type information of the one or more pixel groups. The estimation device 200 includes, for example, a learning unit 202, a correction unit 210, an estimation unit 220, and a storage unit 230. The learning unit 202, the correction unit 210, and the estimation unit 220 are each implemented, for example, by a hardware processor such as a CPU (Central Processing Unit) executing a program (software). In addition, some or all of these components may also be implemented by hardware (including a circuitry part) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), a GPU (Graphics Processing Unit), or an SOC (System On Chip), or may be implemented by the cooperation of software and hardware. The program may be pre-stored in a storage device (a storage device having a non-transitory storage medium) such as an HDD or a flash memory of the estimation device 100, or may be stored in a removable storage medium such as a DVD or a CD-ROM, and installed in the HDD or the flash memory of the estimation device 200 by being mounted on a driving device through the storage medium (non-transitory storage medium). The storage unit 230 stores, for example, learning data 230A, a first machine learning model 230B, and a second machine learning model 230C. The storage unit 230 is implemented, for example, by a RAM, a flash memory, an SD card, or the like. The structures of the correction unit 210 and the estimation unit 220 according to the second embodiment are the same as those of the correction unit 110 and the estimation unit 120 according to the first embodiment, respectively, and thus the description thereof is omitted. The learning unit 202 is an example of the "learning device" in the technical solution. An image including a plurality of pixels is an example of the "data" in the technical solution, and one or more pixel groups in the image are an example of the "at least a part of the data" in the technical solution.

[0049] The learning unit 202 first communicates with the Figure 3The described machine learning similarly learns a machine learning model such as a convolutional neural network based on the learning data 130A, thereby generating a first machine learning model 230B. The generated first machine learning model 230B is the same as the learned machine learning model 130B in the first embodiment. Next, the learning unit 202 uses the output of the learned first machine learning model 230B (more specifically, the probability values P0 to P output via the learned first machine learning model 230B), similar to the first embodiment, to calculate the prior probability P(y C-1 of the occurrence of the samples of class i in the whole of the samples of the learning data 130A, as defined by the peripheral distribution). i ).

[0050] Next, the correction unit 210 uses the calculated prior probability P(y i ) to correct the error function representing the general Cross Entropy error according to the following equation (3), thereby redefining and calculating the error function L. In equation (3), k represents all integers from 0 to C-1 inclusive.

[0051] ···(3)

[0052] Alternatively, the correction unit 210 can redefine and calculate the error function L according to the following equation (4), which is obtained by adding an entropy regularization term for increasing the entropy of the first machine learning model 230B to equation (3). By adding an entropy regularization term to the error function L and performing re-learning of the first machine learning model 230B while maintaining a high entropy, it is possible to prevent the generation of a model that is biased towards a specific class (i.e., the class with a large amount of learning data 130A).

[0053] ···(4)

[0054] In equation (4), λ is a hyperparameter set in advance, and b i represents a bias term with an initial value of logP(y i ), and both are adjusted during the re-learning process. The hyperparameter λ determines the influence degree of the entropy regularization term ε(z). The bias term b i is used to prevent the mismatch between the empirical distribution defined by the learning data 130A and the prior distribution estimated by the model from expanding due to the re-learning of the first machine learning model 230B when the entropy regularization term is added. The entropy regularization term ε(z) is calculated, for example, according to the following equation (5).

[0055] ···(5)

[0056] After the correction unit 210 corrects the error function L, the learning unit 202 uses the redefined error function L to relearn the first machine learning model 230B, thereby generating the second machine learning model 230C. The relearning at this time can be the learning of the entire first machine learning model 230B or the learning of a part thereof. Here, the learning of a part means the fine-tuning of the classification layer of the first machine learning model 230B. The classification layer refers to the last several layers (for example, 2 layers or 3 layers) counted from the output side in the first machine learning model 230B.

[0057] In the estimation stage, the estimation unit 220 inputs the image to be estimated into the second machine learning model 230C, thereby obtaining logitZ’’0~Z’’ C-1 and substitutes the obtained logitZ’’0~Z’’ C-1 into the Softmax function, and obtains and outputs the probability values P’’0~P’’ C-1 . At this time, the correction unit 210 can also use the obtained logitZ’’ i and the second machine learning model 230C to calculate the correction term Δ described in the first embodiment i , and by adding the correction term Δ i to logitZ’’ i to correct logitZ’’ i . Thus, according to the present invention, by combining the correction based on the correction term described in the first embodiment and the relearning described in the second embodiment, the estimation accuracy of the model can be further improved.

[0058] Next, the process executed by the estimation device 200 according to the second embodiment will be described with reference to Figure 6 . Figure 6 FIG. is a flowchart showing an example of the process executed by the estimation device 200 according to the second embodiment.

[0059] First, the learning unit 202 learns the first machine learning model 230B based on the learning data 230A (step S200). Next, the correction unit 210 corrects the error function using the correction term reflecting the output of the first machine learning model 230B (step S202). Next, the learning unit 202 uses the corrected error function to relearn the first machine learning model 230B, and generates the second machine learning model 230C (step S204).

[0060] Next, the estimation unit 220 obtains logitZ’’0~Z’’ by inputting the image to be estimated into the second machine learning model 230C C-1(Step S206). Next, the correction unit 110 corrects the obtained logitZ’’0 to Z’’ C-1 by adding a correction term Δ that reflects the output of the first machine learning model 230B i to calculate the corrected logitZ’’’0 to Z’’’ C-1 (Step S208). Next, the estimation unit 120 substitutes the corrected logitZ’’’0 to Z’’’ C-1 into the Softmax function, and obtains and outputs the corrected probability values P’’0 to P’’ C-1 (Step S210). Thus, the processing of this flowchart ends.

[0061] According to the second embodiment described above, the error function L representing the Cross Entropy error shown in Equation (3) is calculated using the output of the learned first machine learning model 230B. That is, according to the second embodiment, the error function L for learning the machine learning model is corrected with high precision, and the accuracy of the generated machine learning model can be improved.

[0062] It should be noted that in the above embodiment, the case where the machine learning model is configured and learned in such a way that it takes an image as input and outputs the logit of the type of the object shown by the pixel group of the image has been described. However, the present invention is not limited to such a structure, and the machine learning model can also be applied to the case where it is configured and learned in such a way that it takes other types of data (for example, sound) as input and outputs the logit of the type of the data. For example, the machine learning model may be configured and learned in such a way that it takes sound as input and outputs the logit of the type shown by the section of the sound. Even if the data involved is of a type other than an image, as long as at least the machine learning model is configured to output logit, the logit correction process performed by the correction unit 110 (210) described above can also be applied. Sound is another example of "data" in the technical solution, and the section of sound is another example of "at least a part of the data" in the technical solution.

[0063] [Vehicle]

[0064] The above-mentioned estimation device 100 (200) may also be mounted on a vehicle control device and used for controlling the present vehicle M. Figure 7 FIG. is a diagram showing the structure of the present vehicle M equipped with a vehicle control device 300 having an estimation device.

[0065] The present vehicle M includes, for example, a camera 10, an object recognition device 12, a vehicle sensor 14, a driving operation member 20, a steering wheel 22, a driving force output device 30, a braking device 32, a steering device 34, and a vehicle control device 300.

[0066] The camera 10 is, for example, a digital camera that uses solid-state imaging elements such as a CCD (Charge Coupled Device) or a CMOS (Complementary Metal Oxide Semiconductor). The camera 10 is installed at an arbitrary position of the vehicle M. When shooting forward, the camera 10 is installed on the upper part of the windshield, the back of the interior rearview mirror, etc. The camera 10, for example, periodically and repeatedly shoots the periphery of the vehicle M. The camera 10 may also be a stereo camera. The camera 10 sends the captured image to the object recognition device 12.

[0067] The object recognition device 12 performs image processing on the image received from the camera 10 to detect pedestrians, other vehicles, road structures (road markings, walls, etc.) reflected in the image, and sends the detection result to the vehicle control device 300. The detection result in this case, as described with reference to Figure 2 is to recognize the recognized object as a pixel group.

[0068] The vehicle sensor 14 includes a vehicle speed sensor that detects the speed of the vehicle M, an acceleration sensor that detects acceleration, a yaw rate sensor that detects the angular velocity about the vertical axis, an azimuth sensor that detects the orientation of the vehicle M, and the like.

[0069] The driving operation member 20 includes, for example, an accelerator pedal, a brake pedal, a shift lever, and other operation members in addition to the steering wheel 22. A sensor that detects the operation amount or the presence or absence of an operation is installed on the driving operation member 20, and the detection result is output to a part or all of the vehicle control device 300, the driving force output device 30, the braking device 32, and the steering device 34. The operation member does not necessarily have to be ring-shaped, and may be in the form of an irregular steering device, a joystick, a button, or the like.

[0070] The driving force output device 30 outputs the driving force (torque) for the vehicle M to travel to the drive wheels. The driving force output device 30 includes, for example, a combination of an internal combustion engine, an electric motor, and a transmission, and an ECU (Electronic Control Unit) that controls them. The ECU controls the above structure according to the information input from the vehicle control device 300 or the information input from the driving operation member 20.

[0071] The braking device 32 includes, for example, a brake caliper, a hydraulic cylinder that transmits hydraulic pressure to the brake caliper, an electric motor that generates hydraulic pressure in the hydraulic cylinder, and a brake ECU. The brake ECU controls the electric motor in accordance with information input from the vehicle control device 300 or information input from the driving operation member 20, so that a braking torque corresponding to the braking operation is output to each wheel. The braking device 32 may include a mechanism that transmits the hydraulic pressure generated by the operation of the brake pedal included in the driving operation member 20 to the hydraulic cylinder via the master cylinder as a backup. It should be noted that the braking device 32 is not limited to the structure described above, and may also be an electronically controlled hydraulic braking device that controls an actuator in accordance with information input from the vehicle control device 300 and transmits the hydraulic pressure of the master cylinder to the hydraulic cylinder.

[0072] The steering device 34 includes, for example, a steering ECU and an electric motor. The electric motor, for example, applies a force to a rack-pinion mechanism to change the orientation of the steering wheel. The steering ECU drives the electric motor in accordance with information input from the vehicle control device 300 or information input from the driving operation member 20 to change the orientation of the steering wheel.

[0073] The vehicle control device 300 includes, for example, an estimation device 100 (200) and a vehicle control unit 310. The vehicle control unit 310 is implemented, for example, by a hardware processor such as a CPU (Central Processing Unit) executing a program (software). In addition, some or all of these components may be implemented by hardware (including a circuitry unit) such as an LSI (Large Scale Integration), an ASIC (Application-Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit), or may be implemented by the cooperation of software and hardware. The program may be pre-stored in a storage device (a storage device having a non-transitory storage medium) such as an HDD or a flash memory of the vehicle control device 300, or may be stored in a removable storage medium such as a DVD or a CD-ROM, and installed in the HDD or flash memory of the vehicle control device 300 by being mounted on a driving device through the storage medium (non-transitory storage medium). The storage unit 320 is implemented, for example, by a ROM (Read Only Memory), a flash memory, an SD card, a RAM (Random Access Memory), a register, or the like. The storage unit 320 stores the image including the recognition result of the object received from the object recognition device 12 as image data 320A.

[0074] The estimation device 100 (200) uses the estimation unit 120 (220) to estimate the type of the object belonging to the pixel group included in the image data 320A. More specifically, the estimation device 100 (200) estimates the type with the maximum value among the probability values output for the pixel group included in the image data 320A as the type of the object belonging to the pixel group.

[0075] The vehicle control unit 310 controls at least one of the driving force output device 30, the braking device 32, and the steering device 34 based on the estimation result estimated by the estimation device 100 (200), thereby controlling the travel of the own vehicle M. More specifically, for example, when the type of the object estimated by the estimation device 100 (200) belongs to any of a four-wheeled vehicle, a two-wheeled vehicle, a pedestrian, and a falling object, the vehicle control unit 310 controls the travel of the own vehicle M so as to avoid the object. Further, for example, when the type of the object estimated by the estimation device 100 (200) belongs to the type of a traffic signal or a road sign, the vehicle control unit 310 controls the travel of the own vehicle M according to the instruction information indicated by the traffic signal or the road sign (for example, when it is recognized that the instruction information indicated by the traffic signal is stop (red light), the own vehicle M is decelerated or stopped). Further, for example, when the type of the object estimated by the estimation device 100 (200) belongs to the type of a guardrail, a median strip, or a curb, the vehicle control unit 310 controls the own vehicle M to travel along the object.

[0076] As another aspect, it may be that the vehicle control unit 310 does not actually control the own vehicle M, but causes the type of the object estimated by the estimation device 100 (200) to be displayed on a display mounted on the own vehicle M. In other words, the vehicle control unit 310 may also function as a driving support unit.

[0077] The embodiment described above can be expressed as follows.

[0078] An estimation device configured to include: [[ID=--]]

[0079] A storage medium / memory that stores computer-readable instructions; and

[0080] A processor connected to the storage medium,

[0081] The processor performs the following processing by executing the computer-readable instructions: (theprocessor executing the computer-readable instructions to:)

[0082] Inputting target data into a machine learning model that has been learned so as to take data as input and output a logit indicating that at least a portion of the data belongs to a category representing a certain class to obtain the logit, and calculating a correction value for correcting the logit using the output of the machine learning model; and

[0083] The logit is corrected based on the calculated correction value, and the category to which the at least part of the target data belongs is estimated based on the corrected logit.

[0084] While specific embodiments of the present invention have been described above, the present invention is not limited to these embodiments at all, and various modifications and substitutions can be made without departing from the spirit of the present invention.

Claims

1. A presumption device, wherein the presumption device includes: a correction unit that inputs object data into a machine learning model that has been learned to output, as input, at least a part of the data as a logit representing a category of a certain type, obtains the logit, and uses the output of the machine learning model to calculate a correction value for correcting the logit; and a presumption unit that corrects the logit based on the calculated correction value and presumes the category to which at least a part of the object data belongs based on the corrected logit.

2. The presumption device according to claim 1, wherein the correction unit calculates the correction value using a preset hyperparameter and the output of the machine learning model.

3. The presumption device according to claim 2, wherein the correction unit calculates the correction value by multiplying the hyperparameter by the prior probability of a sample in which the category occurs, and the prior probability is calculated based on a marginal distribution defined by the machine learning model.

4. The presumption device according to claim 1, wherein the presumption unit corrects the logit by adding the correction value to the logit and presumes the category for which the probability value based on the corrected logit is maximized as the category to which at least a part of the object data belongs.

5. The presumption device according to claim 1, wherein the data is an image including a plurality of pixels, and at least a part of the data is one or more pixel groups in the image.

6. The presumption device according to claim 1, wherein the data is sound, and at least a part of the data is an interval of the sound.

7. A vehicle control device, wherein the vehicle control device includes: the presumption device according to claim 1; and a vehicle control unit that controls the running of the vehicle based on a presumption result presumed by the presumption device.

8. A vehicle, wherein the vehicle includes the vehicle control device according to claim 7.

9. A presumption method, wherein the presumption method causes a computer to perform the following processing: input object data into a machine learning model that has been learned to output, as input, at least a part of the data as a logit representing a category of a certain type, obtain the logit, and use the output of the machine learning model to calculate a correction value for correcting the logit; and correct the logit based on the calculated correction value and presume the category to which at least a part of the object data belongs based on the corrected logit.

10. A storage medium that stores a program, wherein the program causes a computer to perform the following processing: input object data into a machine learning model that has been learned to output, as input, at least a part of the data as a logit representing a category of a certain type, obtain the logit, and use the output of the machine learning model to calculate a correction value for correcting the logit; and The logit is corrected based on the calculated correction value, and the category to which the at least part of the target data belongs is estimated based on the corrected logit.

Images