Vehicle control torque determination method and device, electronic equipment and storage medium

Abstract

The application provides a vehicle control torque determination method and device, electronic equipment and storage medium. The method comprises the following steps: obtaining target state information of a target vehicle at a current time and a target image of a driving road surface of the target vehicle; inputting the target state information and the target image into a torque control model; processing the target state information through a first attention network to obtain vehicle state features; processing the target image through a second attention network to obtain road surface features; processing the vehicle state features and the road surface features through a feature splicing network to obtain spliced features; and processing the spliced features through a deep neural network to obtain a target control torque of the target vehicle. By using the above technical means, the problem that the vehicle torque output is insufficient and unstable due to the fact that the road conditions are not considered in vehicle torque control in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle control, and particularly relates to a vehicle control torque determination method and device, an electronic device and a storage medium. BACKGROUND

[0002] Existing vehicle torque control algorithms are mainly based on engine speed and vehicle speed for control. This control method only considers part of the state parameters of the vehicle and completely ignores road conditions. This control method may cause insufficient torque output and instability in some road conditions (such as climbing and rainy days), thereby affecting the stability and comfort of driving. SUMMARY

[0003] Therefore, the present application provides a vehicle control torque determination method and device, an electronic device and a storage medium to solve the problem of insufficient torque output and instability caused by the lack of consideration of road conditions in vehicle torque control.

[0004] In a first aspect, the present application provides a vehicle control torque determination method, comprising: obtaining target state information of a target vehicle at a current time and a target image of a driving road surface thereof; inputting the target state information and the target image into a torque control model; processing the target state information through a first attention network to obtain vehicle state features; processing the target image through a second attention network to obtain road surface features; processing the vehicle state features and the road surface features through a feature concatenation network to obtain concatenated features; and processing the concatenated features through a deep neural network to obtain a target control torque of the target vehicle.

[0005] In a second aspect, the present application provides a vehicle control torque determination device, comprising: an obtaining module configured to obtain target state information of a target vehicle at a current time and a target image of a driving road surface thereof; a first processing module configured to input the target state information and the target image into a torque control model; a second processing module configured to process the target state information through a first attention network to obtain vehicle state features; a third processing module configured to process the target image through a second attention network to obtain road surface features; a fourth processing module configured to process the vehicle state features and the road surface features through a feature concatenation network to obtain concatenated features; and a fifth processing module configured to process the concatenated features through a deep neural network to obtain a target control torque of the target vehicle.

[0006] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0007] In a fourth aspect, the present disclosure provides a computer-readable storage medium storing a computer program, the computer program being executed by a processor to implement the steps of any of the above methods.

[0008] Compared with the prior art, the present disclosure has at least the following beneficial effects: the first attention network is used to process the target state information to obtain the vehicle state feature; the second attention network is used to process the target image to obtain the road surface feature; the feature splicing network is used to process the vehicle state feature and the road surface feature to obtain the spliced feature; and the deep neural network is used to process the spliced feature to obtain the target control torque of the target vehicle. Therefore, the above technical means can solve the problem that the vehicle torque control does not consider the road condition in the prior art, resulting in insufficient and unstable vehicle torque output, thereby ensuring the vehicle power output and smooth driving, and improving the user comfort and satisfaction. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creative labor.

[0010] Figure 1 is a flow diagram of a vehicle control torque determination method provided by the present disclosure;

[0011] Figure 2 is a flow diagram of a torque control model training method provided by the present disclosure;

[0012] Figure 3 is a structure diagram of a torque control model provided by the present disclosure

[0013] Figure 4 is a structure diagram of a vehicle control torque determination device provided by the present disclosure;

[0014] Figure 5 is a structure diagram of an electronic device provided by the present disclosure. DETAILED DESCRIPTION

[0015] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present disclosure. However, it will be apparent to those skilled in the art that the present disclosure can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present disclosure with unnecessary detail.

[0016] Figure 1 is a flowchart of a vehicle control torque determination method provided by an embodiment of the present disclosure. Figure 1 The vehicle control torque determination method can be executed by a computer or a server, or a processor arranged on the computer or the server, or software on the computer or the server. The vehicle control torque determination method comprises:

[0017] S101, constructing a torque control model by using a first attention network, a second attention network, a feature splicing network and a deep neural network;

[0018] S102, obtaining target state information of a target vehicle at a current time and a target image of a driving road surface of the target vehicle;

[0019] S103, inputting the target state information and the target image into the torque control model: processing the target state information by the first attention network to obtain vehicle state features;

[0020] S104, processing the target image by the second attention network to obtain road surface features;

[0021] S105, processing the vehicle state features and the road surface features by the feature splicing network to obtain spliced features;

[0022] S106, processing the spliced features by the deep neural network to obtain a target control torque of the target vehicle.

[0023] The target state information includes vehicle speed, vehicle acceleration, steering wheel angle, steering wheel angular velocity, steering wheel angular acceleration and yaw rate. The yaw rate refers to the deflection of the vehicle around the vertical axis, and the size of the yaw rate represents the stability of the vehicle. The prior art only considers part of the vehicle state information, and the present application considers all the vehicle state information, thereby avoiding the problems of insufficient vehicle torque output and instability.

[0024] The target image relates to information of a road surface on which the vehicle travels, and the prior art only considers that the vehicle travels on a regular road (a road under normal weather conditions with straightness and flatness), without considering special road surfaces such as a potholed road surface, a speed bump, a snow-covered road surface, an ice-covered road surface, a rain-covered road surface, an uphill, and a downhill. The present application considers various special road surfaces, thereby ensuring that the vehicle can output torque sufficiently and stably under all road conditions.

[0025] According to the technical scheme provided by the embodiment of the present disclosure, the target state information of the target vehicle at the current moment and the target image of the road surface on which the target vehicle travels are obtained; the target state information and the target image are input into a torque control model; the target state information is processed through a first attention network to obtain vehicle state features; the target image is processed through a second attention network to obtain road surface features; the vehicle state features and the road surface features are processed through a feature splicing network to obtain spliced features; and the spliced features are processed through a deep neural network to obtain a target control torque of the target vehicle. By using the above technical means, the problem that the vehicle torque output is insufficient and unstable due to the fact that the vehicle torque control does not consider the road conditions in the prior art can be solved, thereby ensuring the power output and smooth driving of the vehicle and improving the comfort and satisfaction of the user.

[0026] According to the vehicle state information and the road surface image, the torque control model is determined according to the embodiment of the present disclosure. The torque control model has been trained, and the corresponding relationship between the vehicle state information and the road surface image and the control torque has been learned and saved. Specifically, the trained first attention network extracts the features of the target state information, the trained second attention network extracts the features of the target image, the feature splicing network splices the features of the target state information and the features of the target image, and the trained deep neural network realizes the mapping from the spliced features to the control torque.

[0027] Figure 2 is a flowchart of a training method of a torque control model provided by the embodiment of the present disclosure, as shown in Figure 2 , comprising:

[0028] S201, obtaining a training data set, wherein the training data set comprises a plurality of groups of training data, each group of training data being state information of a vehicle and an image of a road surface on which the vehicle travels, and the state information and the image in each group of training data being collected at the same time;

[0029] S202, inputting each group of training data into a torque control model to output a first control torque corresponding to each group of training data;

[0030] S203, calculating a first classification loss between the first control torque corresponding to each group of training data and a label through a cross-entropy loss function;

[0031] S204, updating the network parameters of the torque control model according to the first classification loss corresponding to each group of training data, to complete the training of the torque control model.

[0032] The training data set can be obtained by collecting a large number of vehicle groups of training data. The label corresponding to each group of training data is the annotated expected (ideal) control torque. Each group of training data is input into the torque control model, and the first control torque corresponding to each group of training data is output. The first control torque corresponding to each group of training data is predicted by the torque control model based on each group of training data. The cross-entropy loss function is used to calculate the loss (first classification loss) between the predicted value (first control torque) and the labeled value (label) corresponding to each group of training data.

[0033] Specifically, each group of training data is input into the torque control model, and the first control torque corresponding to each group of training data is output, including: processing the state information in each group of training data through the first attention network to obtain the vehicle state feature corresponding to each group of training data; processing the image in each group of training data through the second attention network to obtain the road surface feature corresponding to each group of training data; processing the vehicle state feature and the road surface feature corresponding to each group of training data through the feature splicing network to obtain the splicing feature corresponding to each group of training data; processing the splicing feature corresponding to each group of training data through the deep neural network to obtain the first control torque corresponding to each group of training data.

[0034] Specifically, the network parameters of the torque control model are updated according to the first classification loss corresponding to each group of training data, to complete the training of the torque control model, including: calculating the gradient corresponding to each group of training data through the back propagation algorithm according to the first classification loss corresponding to each group of training data; updating the network parameters of the torque control model through the gradient descent method based on the gradient corresponding to each group of training data, to complete the training of the torque control model.

[0035] Further, after obtaining the training data set, the method further comprises: inputting each group of training data into the torque control model, and outputting the splicing feature corresponding to each group of training data through the feature splicing network; calculating the class center feature of the class to which each group of training data belongs based on the splicing feature corresponding to each group of training data; calculating the supervision loss between the splicing feature corresponding to each group of training data and the class center feature through the mean square error loss function; updating the network parameters of the torque control model according to the first classification loss and the supervision loss corresponding to each group of training data, to complete the training of the torque control model.

[0036] The embodiments of the present disclosure actually supervise the spliced features with the class center features, and calculate the supervision loss at the feature level. Supervising the spliced features with the class center features means that the spliced features processed by the first attention network, the second attention network and the feature splicing network in the torque control model are close to the class center features. The method has the advantage of flexibly utilizing the relationship between the groups of training data, and optimizing the network parameters of the torque control model without using labels.

[0037] Further, based on the spliced features corresponding to the groups of training data, the class center features of the classes to which the groups of training data belong are calculated, including: clustering the groups of training data based on the spliced features corresponding to the groups of training data, so as to divide the groups of training data into multiple classes, wherein each class includes multiple groups of training data; and averaging the spliced features corresponding to the groups of training data included in each class, and taking the calculation result as the class center feature of each class.

[0038] The embodiments of the present disclosure find the relationship between the groups of training data by clustering, that is, divide the groups of training data into multiple classes. Further, the class center features of each class are calculated. Because a group of training data belongs to a class, the spliced feature corresponding to the group of training data should be close to the class center feature of the class.

[0039] Further, the supervision loss between the spliced features corresponding to the groups of training data and the class center features is calculated by the following mean square error loss function:

[0040]

[0041] Wherein, MSE() is a mean square error function (the mean square error loss function includes the mean square error function), W i is the class center feature of the class to which the i-th group of training data belongs, Z i is the spliced feature corresponding to the i-th group of training data, i is a natural number, i takes a value between 1 and N, and N is the number of groups of training data in the training data set.

[0042] According to the first classification loss and the supervision loss corresponding to each group of training data, the network parameters of the torque control model can be updated by weighted sum of the first classification loss and the supervision loss corresponding to each group of training data, and the network parameters of the torque control model are updated by using the result of the weighted sum.

[0043] Specifically, according to the result of the weighted sum of the first classification loss and the supervision loss corresponding to each group of training data, the gradients corresponding to each group of training data are calculated by the back propagation algorithm; based on the gradients corresponding to each group of training data, the network parameters of the torque control model are updated by the gradient descent method, so as to complete the training of the torque control model.

[0044] Further, after obtaining the training data set, the method further comprises: performing first stage training on the torque control model: freezing the network parameters of the deep neural network, training the first attention network and the second attention network using the training data set; performing second stage training on the torque control model: freezing the network parameters of the first attention network and the second attention network, training the deep neural network using the training data set; and performing third stage training on the torque control model: training the torque control model as a whole using the training data set.

[0045] After completing the first stage training, the network parameters of the deep neural network are unfrozen, and the second stage training is performed on the torque control model. After completing the second stage training, the network parameters of the first attention network and the second attention network are unfrozen, and the third stage training is performed on the torque control model. The third stage training is a whole training of the torque control model, and does not involve freezing of network parameters. Because the feature concatenation network is constructed based on feature concatenation processing, the network parameters of the feature concatenation network are not actually updated during training of the torque control model.

[0046] Whether each stage of training is completed can be determined by determining whether the accuracy of the corresponding network or model is greater than a preset value. For example, in the first stage training, whether the first stage training is completed is determined by determining whether the accuracy of the first attention network and the second attention network is greater than a preset value.

[0047] In some embodiments: the first stage training on the torque control model: freezing the network parameters of the deep neural network; inputting each group of training data into the torque control model, outputting the concatenation features corresponding to each group of training data through the feature concatenation network; calculating the class center features of the classes to which each group of training data belongs based on the concatenation features corresponding to each group of training data; calculating the supervision loss between the concatenation features corresponding to each group of training data and the class center features through the mean square error loss function; and updating the network parameters of the first attention network and the second attention network according to the supervision loss corresponding to each group of training data.

[0048] Because the concatenation features corresponding to each group of training data are obtained through the first attention network, the second attention network and the feature concatenation network in the torque control model, the supervision loss calculated based on the concatenation features corresponding to each group of training data can be used to update the network parameters of the first attention network and the second attention network.

[0049] In some embodiments, the network parameters of the first attention network, the second attention network and the feature concatenation network can also be updated according to the supervision loss corresponding to each group of training data. Updating the network parameters of the feature concatenation network means finding the proportion of concatenating the vehicle state features and the road surface features corresponding to each group of training data.

[0050] The second stage of training the torque control model is performed as follows: freezing the network parameters of the first attention network and the second attention network; inputting the class center features corresponding to each group of training data into the torque control model, and outputting the second control torque corresponding to each group of training data through the deep neural network; calculating the second classification loss between the second control torque corresponding to each group of training data and the label through the cross-entropy loss function; and updating the network parameters of the deep neural network according to the second classification loss corresponding to each group of training data.

[0051] The class center features corresponding to each group of training data are related to the spliced features, so the second control torque corresponding to each group of training data is related to the first control torque, and the second classification loss corresponding to each group of training data is related to the first classification loss. Therefore, updating the network parameters of the deep neural network according to the second classification loss corresponding to each group of training data can supplement the updating of the network parameters of the deep neural network according to the first classification loss corresponding to each group of training data.

[0052] The third stage of training the torque control model is performed as follows: inputting each group of training data into the torque control model, and outputting the control torque corresponding to each group of training data; calculating the first classification loss between the control torque corresponding to each group of training data and the label through the cross-entropy loss function; and updating the network parameters of the torque control model according to the first classification loss corresponding to each group of training data to complete the third stage of training.

[0053] In some embodiments, the first classification loss and the second classification loss corresponding to each group of training data are weighted and summed to obtain the third classification loss corresponding to each group of training data; and the network parameters of the torque control model are updated according to the third classification loss corresponding to each group of training data to complete the third stage of training.

[0054] Figure 3 is a structural schematic diagram of a torque control model provided by an embodiment of the present disclosure. As shown in Figure 3 , the torque control model sequentially comprises a first attention network, a second attention network, a feature splicing network, and a deep neural network.

[0055] Specifically, the first attention network and the second attention network are connected in parallel, and both the first attention network and the second attention network are connected in series to the feature splicing network, and the feature splicing network is connected in series to the deep neural network to obtain the torque control model.

[0056] That is, in the torque control model, the first attention network and the second attention network, the feature splicing network, and the deep neural network are sequentially connected in series, and the first attention network and the second attention network are connected in parallel.

[0057] The first attention network and the second attention network can be the same attention mechanism, except that the two are trained to process different types of information. The feature splicing network is constructed based on feature splicing processing, for splicing the features processed by the first attention network and the second attention network together. The first attention network, the second attention network and the deep neural network are existing networks, which will not be described again.

[0058] All the optional technical solutions described above can be combined to form optional embodiments of the present application, which will not be described again.

[0059] The following is an apparatus embodiment of the present application, which can be used to execute the method embodiments of the present application. For details not disclosed in the apparatus embodiments of the present application, please refer to the method embodiments of the present application.

[0060] Figure 4 is a schematic diagram of a vehicle control torque determination device provided by an embodiment of the present disclosure. As shown in Figure 4 , the vehicle control torque determination device comprises:

[0061] The construction module 401 constructs a torque control model by using a first attention network, a second attention network, a feature splicing network and a deep neural network.

[0062] The acquisition module 402 is configured to acquire target state information of a target vehicle at a current time and a target image of a driving road surface thereof.

[0063] The first processing module 403 is configured to input the target state information and the target image into the torque control model: the target state information is processed by the first attention network to obtain vehicle state features.

[0064] The second processing module 404 is configured to process the target image by the second attention network to obtain road surface features.

[0065] The third processing module 405 is configured to process the vehicle state features and the road surface features by the feature splicing network to obtain spliced features.

[0066] The fourth processing module 406 is configured to process the spliced features by the deep neural network to obtain a target control torque of the target vehicle.

[0067] The target state information includes vehicle speed, vehicle acceleration, steering wheel angle, steering wheel angular velocity, steering wheel angular acceleration and yaw rate, etc. The yaw rate refers to the deflection of the vehicle around the vertical axis, and the size of the yaw rate represents the stability of the vehicle. The prior art only considers part of the vehicle state information, and the present application considers all the vehicle state information, thereby avoiding the problems of insufficient vehicle torque output and instability.

[0068] The target image relates to information of a driving road surface of the vehicle, and the prior art only considers that the vehicle drives on a regular road (a road under normal weather with straightness and flatness), without considering special road surfaces such as a potholed road surface, a speed bump, a snow-covered road surface, an ice-covered road surface, a rain-covered road surface, an uphill and a downhill, and the like. The present application considers various special road surfaces, thereby ensuring that the vehicle can output sufficient and smooth torque under all road conditions.

[0069] In some embodiments, the first attention network and the second attention network are connected in parallel, and the first attention network and the second attention network are both connected in series to the feature concatenation network, and the feature concatenation network is connected in series to the deep neural network, to obtain the torque control model.

[0070] According to the technical scheme provided by the embodiments of the present disclosure, the target state information of the target vehicle at the current time and the target image of the driving road surface thereof are obtained; the target state information and the target image are input into a torque control model; the target state information is processed by a first attention network to obtain vehicle state features; the target image is processed by a second attention network to obtain road surface features; the vehicle state features and the road surface features are processed by a feature concatenation network to obtain concatenated features; and the concatenated features are processed by a deep neural network to obtain a target control torque of the target vehicle. By using the above technical means, the problem that the vehicle torque control does not consider the road condition in the prior art, resulting in insufficient and unstable vehicle torque output, can be solved, thereby ensuring vehicle power output and smooth driving, and improving user comfort and satisfaction.

[0071] In some embodiments, the first processing module 403 is further configured to obtain a training data set, wherein the training data set includes a plurality of groups of training data, each group of training data being state information of a vehicle and an image of a driving road surface of the vehicle, and the state information and the image in each group of training data being collected at the same time; inputting each group of training data into the torque control model to output a first control torque corresponding to each group of training data; calculating a first classification loss between the first control torque corresponding to each group of training data and a label by using a cross-entropy loss function; and updating network parameters of the torque control model according to the first classification loss corresponding to each group of training data, to complete training of the torque control model.

[0072] In some embodiments, the first processing module 403 is further configured to input each group of training data into the torque control model, output the spliced features corresponding to each group of training data through the feature splicing network, calculate the class center features of the class to which each group of training data belongs based on the spliced features corresponding to each group of training data, calculate the supervision loss between the spliced features corresponding to each group of training data and the class center features through the mean square error loss function, and update the network parameters of the torque control model according to the first classification loss and the supervision loss corresponding to each group of training data to complete the training of the torque control model.

[0073] In some embodiments, the first processing module 403 is further configured to cluster the multiple groups of training data based on the spliced features corresponding to each group of training data to divide the multiple groups of training data into multiple classes, wherein each class includes multiple groups of training data, and average the spliced features corresponding to each group of training data included in each class to take the calculation result as the class center features of each class.

[0074] In some embodiments, the first processing module 403 is further configured to calculate the supervision loss between the spliced features corresponding to each group of training data and the class center features through the following mean square error loss function:

[0075]

[0076] wherein MSE() is a mean square error function (the mean square error loss function includes the mean square error function), W i is the class center features of the class to which the i-th group of training data belongs, Z i is the spliced features corresponding to the i-th group of training data, i is a natural number, i takes a value between 1 and N, and N is the number of groups of training data in the training data set.

[0077] In some embodiments, the first processing module 403 is further configured to freeze the network parameters of the deep neural network, train the first attention network and the second attention network using the training data set, perform second-stage training on the torque control model by freezing the network parameters of the first attention network and the second attention network, and train the deep neural network using the training data set, and perform third-stage training on the torque control model by training the torque control model as a whole using the training data set.

[0078] After the first-stage training is completed, the network parameters of the deep neural network are unfrozen, and the second-stage training is performed on the torque control model. After the second-stage training is completed, the network parameters of the first attention network and the second attention network are unfrozen, and the third-stage training is performed on the torque control model. The third-stage training is the training of the torque control model as a whole, and does not involve freezing of network parameters. Because the feature splicing network is constructed based on feature splicing processing, the network parameters of the feature splicing network are not updated during the training of the torque control model.

[0079] In some embodiments, the first processing module 403 is further configured to perform first stage training on the torque control model: freeze the network parameters of the deep neural network; input each set of training data into the torque control model, output the spliced features corresponding to each set of training data through the feature splicing network; calculate the class center features of the class to which each set of training data belongs based on the spliced features corresponding to each set of training data; calculate the supervision loss between the spliced features corresponding to each set of training data and the class center features through the mean square deviation loss function; update the network parameters of the first attention network and the second attention network according to the supervision loss corresponding to each set of training data. Perform second stage training on the torque control model: freeze the network parameters of the first attention network and the second attention network; input the class center features corresponding to each set of training data into the torque control model, output the second control torque corresponding to each set of training data through the deep neural network; calculate the second classification loss between the second control torque corresponding to each set of training data and the label through the cross-entropy loss function; update the network parameters of the deep neural network according to the second classification loss corresponding to each set of training data. Perform third stage training on the torque control model: input each set of training data into the torque control model, output the control torque corresponding to each set of training data; calculate the first classification loss between the control torque corresponding to each set of training data and the label through the cross-entropy loss function; update the network parameters of the torque control model according to the first classification loss corresponding to each set of training data, to complete the third stage training.

[0080] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

[0081] Figure 5 is a schematic diagram of an electronic device 5 provided by the embodiments of the present disclosure. As shown in the figure, the electronic device 5 of this embodiment includes a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. The processor 501 implements the steps in each of the above method embodiments when executing the computer program 503. Alternatively, the processor 501 implements the functions of each module / unit in the above device embodiments when executing the computer program 503. Figure 5

[0082] The electronic device 5 can include but is not limited to the processor 501 and the memory 502. Those skilled in the art can understand that the electronic device 5 can include more or fewer components than those shown in the figure, or different components. Figure 5 The electronic device 5 is only an example and does not constitute a limitation on the electronic device 5, which can include more or fewer components than those shown in the figure, or different components.

[0083] ​The processor 501 can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0084] The memory 502 can be an internal storage unit of the electronic device 5, for example, a hard disk or a memory of the electronic device 5. The memory 502 can also be an external storage device of the electronic device 5, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 5. The memory 502 can also include both the internal storage unit and the external storage device of the electronic device 5. The memory 502 is used to store computer programs and other programs and data required by the electronic device.

[0085] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0086] The integrated modules / units, if implemented in the form of software functional units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can be executed by a processor to implement the steps of the above-mentioned various method embodiments. The computer program can include computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electric carrier signals and telecommunication signals.

[0087] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of determining a vehicle control torque, characterized by, include: A torque control model is constructed using a first attention network, a second attention network, a feature splicing network, and a deep neural network, and the torque control model is trained using a training dataset. The training dataset includes multiple sets of training data, each set of training data consisting of a vehicle state information and an image of the road surface on which the vehicle is traveling. The state information and the image in each set of training data are collected at the same time, and the label corresponding to each set of training data is the labeled ideal control torque. Acquire the target status information of the target vehicle at the current moment and the target image of the road surface on which it is traveling; The target state information and the target image are input into the torque control model: The target state information is processed by the first attention network to obtain vehicle state features; The target image is processed by the second attention network to obtain road surface features; The vehicle state features and road surface features are processed by the feature stitching network to obtain stitched features; The target control torque of the target vehicle is obtained by processing the spliced ​​features through the deep neural network.

2. The method of claim 1, wherein, The method further includes: The first attention network and the second attention network are connected in parallel, and both the first attention network and the second attention network are connected in series to the feature splicing network. The feature splicing network is then connected in series to the deep neural network to obtain the torque control model.

3. The method of claim 1, wherein, The step of training the torque control model using a training dataset includes: Input each set of training data into the torque control model and output the first control torque corresponding to each set of training data. The first classification loss between the first control torque and the label corresponding to each group of training data is calculated using the cross-entropy loss function. The network parameters of the torque control model are updated based on the first classification loss corresponding to each group of training data to complete the training of the torque control model.

4. The method of claim 3, wherein, After obtaining the training dataset, the method further includes: Each set of training data is input into the torque control model, and the spliced ​​features corresponding to each set of training data are output through the feature splicing network. Based on the splicing features corresponding to each group of training data, calculate the class center features of the class to which each group of training data belongs; The supervision loss between the spliced ​​features and class center features corresponding to each group of training data is calculated using the mean squared error loss function. The network parameters of the torque control model are updated based on the first classification loss and the supervised loss corresponding to each group of training data to complete the training of the torque control model.

5. The method of claim 4, wherein, Based on the concatenated features corresponding to each set of training data, the class center features of the class to which each set of training data belongs are calculated, including: Based on the splicing features corresponding to each group of training data, multiple groups of training data are clustered to divide them into multiple classes, where each class includes multiple groups of training data. The average of the concatenated features corresponding to each group of training data included in each class is calculated, and the result is used as the class center feature of each class.

6. The method of claim 4, wherein, After obtaining the training dataset, the method further includes: The torque control model is trained in the first stage: the network parameters of the deep neural network are frozen, and the first attention network and the second attention network are trained using the training dataset; The torque control model is trained in the second stage: the network parameters of the first attention network and the second attention network are frozen, and the deep neural network is trained using the training dataset; The torque control model is trained in the third stage: the torque control model is trained as a whole using the training dataset.

7. The method according to claim 6, characterized in that, The method further includes: The torque control model is trained in the first stage: the network parameters of the first attention network and the second attention network are updated according to the supervised loss corresponding to each group of training data; The torque control model is trained in the third stage: the network parameters of the torque control model are updated according to the first classification loss corresponding to each group of training data; The torque control model is then trained in the second stage: The class center features corresponding to each group of training data are input into the torque control model, and the second control torque corresponding to each group of training data is output through the deep neural network. The second classification loss between the second control torque and the label corresponding to each group of training data is calculated using the cross-entropy loss function. The network parameters of the deep neural network are updated based on the second classification loss corresponding to each group of training data.

8. A device for determining vehicle control torque, characterized in that, include: The construction module is configured to build a torque control model using a first attention network, a second attention network, a feature concatenation network, and a deep neural network, and to train the torque control model using a training dataset. The training dataset includes multiple sets of training data, each set of training data consisting of a vehicle state information and an image of the road surface on which the vehicle is traveling. The state information and the image in each set of training data are collected at the same time, and the label corresponding to each set of training data is the labeled ideal control torque. The acquisition module is configured to acquire the target status information of the target vehicle at the current moment and the target image of the road surface on which it is traveling. The first processing module is configured to input the target state information and the target image into the torque control model: process the target state information through a first attention network to obtain vehicle state features; The second processing module is configured to process the target image through a second attention network to obtain road surface features; The third processing module is configured to process the vehicle state features and the road surface features through a feature stitching network to obtain stitched features; The fourth processing module is configured to process the spliced ​​features through a deep neural network to obtain the target control torque of the target vehicle.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images