Automatic driving control method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of automobile intelligent driving, and relates to an automatic driving control method and system based on artificial intelligence, and the method comprises the steps: carrying out the vehicle kinematics modeling, creating an action execution calculation model, and training; current environment data in the vehicle automatic driving engineering are collected in real time and input into the trained action execution calculation model; the action execution calculation model calculates an execution action and transmits the execution action to the control command generation module, and the execution action is converted into a corresponding control command for a vehicle execution system; the vehicle execution system performs action execution based on the control command, measures the actual pose of the vehicle, compares the actual pose with the theoretical pose, and calculates the geometric parameter error of the actual pose and the theoretical pose; and the real-time parameter adjustment amount in the control command is reversely calculated and transmitted to the control command generation module to generate a new control command, and the new control command is transmitted to the vehicle execution system. The safety risk of intelligent driving is reduced by reducing the error between the actual action and the theoretical action in the intelligent driving process.

Description

Technical Field

[0001] The present invention belongs to the field of automobile intelligent driving technology, and in particular relates to an automatic driving control method and system based on artificial intelligence. Background Art

[0002] Intelligent driving refers to the use of technologies such as artificial intelligence, visual computing, radar, monitoring devices and global positioning systems to enable vehicles to complete driving tasks autonomously and safely without active human operation. The core of intelligent driving lies in environmental perception, path planning and autonomous decision-making capabilities. These functions rely on multiple sensors such as lidar, cameras and radar to perceive the surrounding environment in real time and make decisions through complex algorithms.

[0003] For example, in a certain actual environment, based on the actual road conditions or when encountering an emergency during the autonomous driving process, the intelligent driving system will first determine the action that the vehicle should perform and the magnitude of the action based on the current actual road conditions or emergency situations after a series of internal algorithms. For example, for steering, it is necessary to calculate the direction and angle of rotation, as well as the rate of rotation, and then generate the corresponding control command, and issue the corresponding control command to the vehicle. Then the vehicle's action execution system executes the relevant action based on the control command of the intelligent driving system. However, due to various reasons, when the execution system is executing an action, there is an error between the actual action and the action that the intelligent driving system expects the execution system to execute. When the error between the actual action and the action that the intelligent driving system expects the execution system to execute exceeds the specified range, there will be safety risks in the process of intelligent driving.

[0004] Therefore, how to reduce the error between the actual actions in the intelligent driving system process and the actions that the intelligent driving system expects to execute, and reduce the safety risks of the intelligent driving process, is a technical problem that urgently needs to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide an automatic driving control method and system based on artificial intelligence, so as to reduce the error between the actual action in the intelligent driving system process and the action that the intelligent driving system expects to execute, and reduce the safety risk of the intelligent driving process.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, an automatic driving control method based on artificial intelligence is provided, comprising the following steps: S1: establishing a vehicle coordinate system, and modeling a vehicle kinematics model based on the vehicle coordinate system; S2: creating an action execution calculation model, acquiring actual environment data during driving based on big data, constructing training samples based on the actual environment data, and inputting the training samples into the action execution calculation model for model training; S3: collecting current environmental data in the vehicle automatic driving project in real time, performing data preprocessing on the collected environmental data, and inputting the preprocessed environmental data into the trained action execution calculation model; S4: the action execution calculation model calculates the execution action based on the current environment data, and transmits the calculation result of the execution action to the control command generation module, and the control command generation module converts the execution action into a corresponding control command for the vehicle execution system; S5: The vehicle execution system executes the corresponding action based on the control command, and the real-time posture measurement module measures the actual posture of the vehicle, compares the actual posture of the vehicle with the theoretical posture of the execution action in step S4, and calculates the geometric parameter error between the actual posture and the theoretical posture of the execution action; S6: reversely calculating the real-time parameter adjustment amount in the control command based on the geometric parameter error, and reversely transmitting the real-time parameter adjustment amount to the control command generation module, the control command generation module generates a new control command based on the real-time parameter adjustment amount, and transmits the new control command to the vehicle execution system; S7: Iteratively execute steps S3 to S6 to perform real-time adjustment and control during the automatic driving process of the vehicle.

[0007] Preferably, the specific process of modeling the vehicle kinematic model in step S1 is as follows: S11: In the established vehicle coordinate system, simplify the vehicle into a two-wheel model, simplify the two front wheels into one front wheel, and simplify the two rear wheels into one rear wheel; S12: Perform inverse kinematics solution on the simplified two-wheel model and create an inverse kinematics solution formula, as follows: cosθ1= a tan( y , x ); ; ; θ3=π- a cos[( a 2 + b 2 - c 2 ) / 2 ab ]; The front wheel and the rear wheel are simplified into two long strips, and θ1 is the line connecting the frontmost point of the front wheel and the rearmost point of the rear wheel and the established coordinate system. x The angle between the axes, θ2 is the angle between the straight line where the rear wheel is located and x The angle between the axes, θ3 is the angle between the straight line where the rear wheel is located and the line connecting the frontmost point of the front wheel and the rearmost point of the rear wheel, a is the distance between the intersection of the straight line where the front wheel is and the straight line where the rear wheel is and the last point of the rear wheel, b It is the distance between the intersection of the straight line where the front wheels are located and the straight line where the rear wheels are located and the frontmost point of the front wheel.

[0008] Preferably, the specific process of constructing the training sample based on the actual environment data in step S2 is as follows: S21: randomly selecting a designated small portion of data from the actual environment data for labeling, inputting the small portion of data into the action execution calculation model for prediction, and predicting the unlabeled actual environment data through the action execution calculation model; S22: Score the prediction results of the unlabeled actual environment data according to the specified rules, and label the actual environment data whose score value is greater than the specified threshold; S23: Add the actual environment data annotated in step S22 to the small part of the data annotated in step S21 to construct a training sample.

[0009] Preferably, the specific process of preprocessing the collected environmental data in step S3 is as follows: S31: removing noise data from the collected environmental data, calculating the standard deviation of the environmental data after removing abnormal data, marking data points in the environmental data whose values ​​exceed the mean value plus or minus a specified number of standard deviations as abnormal values, and removing them; S32: performing feature processing on the environmental data of different forms with different sources from which the noise data and the abnormal data are removed, including extracting image features from the image data and sound features from the sound data, and converting the extracted image features and sound features into text data; S33: converting the text data into vector data, segmenting the text data to obtain individual words, constructing a vocabulary for the words obtained by segmenting the text data, assigning a unique index to each word, counting the frequency of the word, and constructing a text vector based on the word and the corresponding frequency; S34: performing normalization processing on the text vectors and scaling processing on the text vectors to scale the text vectors to the same specified length.

[0010] Preferably, the action execution calculation model in step S4 includes an input layer, a hidden layer and an output layer, the hidden layer includes a combination layer of a convolution layer and a pooling layer, an adaptive transformation layer and a stream projection layer, the adaptive transformation layer is connected to the stream projection layer and is located after the combination layer of the convolution layer and the pooling layer; The input layer is used to input the current environment data collected in real time; The convolution layer is used to extract spatial hierarchical features of image data from the input current environment data collected in real time; The pooling layer is used to reduce the spatial dimension of the spatial hierarchical features output by the previous convolutional layer; The adaptive transformation layer is used to perform pixel transformation processing on the image frame with reduced spatial hierarchical features of spatial dimension, and extract image data driving features; The flow projection layer is used to perform flow displacement on the image frames before and after the image frames undergoing pixel transformation processing, and to project the image frames undergoing pixel transformation processing to simulate the flow between the image frames undergoing pixel transformation processing.

[0011] Preferably, the adaptive transformation layer uses an interpolation kernel and optical flow to perform pixel transformation processing on image frames with reduced spatial hierarchical features of spatial dimensions in a manner that fuses bilinear deformation and kernel convolution in single-frame image processing, thereby extracting image data-driven features.

[0012] Preferably, when there are multiple flow vectors projected to the same position of the central image frame in the process of projecting the image frame processed by pixel transformation in the flow projection layer to simulate the flow between the image frames processed by pixel transformation, the multiple flow vectors at the position are averaged; when there are multiple hole positions where the flow vectors are not projected, the hole positions are filled inward from the outer edge of the hole to the designated image frame.

[0013] Preferably, the specific process of calculating the geometric parameter error between the actual posture and the calculation result of the execution action posture in step S5 is: S51: Synchronizing the actual posture to the coordinate system of the theoretical posture according to the actual posture; S52: Calculate the error value between the geometric parameters of the actual posture and the corresponding geometric parameters of the theoretical posture in the coordinate system.

[0014] Preferably, in step S6, the real-time parameter adjustment amount in the control command is reversely calculated based on the geometric parameter error, the continuous parameter adjustment amount data within the specified period is obtained, and a mapping relationship table between the environmental data and the parameter adjustment amount is constructed. In the intelligent driving process, after the real-time environmental data is collected, the mapping relationship table is directly retrieved to obtain the corresponding parameter adjustment amount, and then the parameter adjustment amount is reversely transmitted to the control command generation module. The control command generation module generates a new control command based on the actual parameter adjustment amount, and transmits the new control command to the vehicle execution system. By establishing a mapping relationship table, the parameter adjustment amount in the current environment can be quickly obtained, and the execution efficiency of the vehicle execution system can be improved while ensuring the accuracy of the execution action of the vehicle execution system.

[0015] Preferably, in step S6, the real-time parameter adjustment amount in the control command is reversely calculated based on the geometric parameter error, the continuous parameter adjustment amount data within the specified period is obtained, and the parameter adjustment data sample is obtained. A parameter adjustment prediction model is constructed, and the parameter adjustment data sample and the corresponding environmental data are input into the parameter adjustment prediction model to train the model. The parameter adjustment model is a neural network model, which automatically captures the relationship between the learning environment data and the parameter adjustment amount during the model training process. Then the real-time environmental data is collected and input into the parameter adjustment prediction model, and the parameter adjustment prediction model automatically outputs the corresponding parameter adjustment amount under the current environmental data. Then the parameter adjustment amount is transmitted in reverse to the control command generation module, and the control command generation module generates a new control command based on the actual parameter adjustment amount, and transmits the new control command to the vehicle execution system. By establishing a parameter adjustment prediction model of a neural network, the parameter adjustment amount under the current environment can be quickly predicted, and the execution efficiency of the vehicle execution system can be improved while ensuring the accuracy of the execution action of the vehicle execution system.

[0016] In a second aspect, an automatic driving control system based on artificial intelligence is provided, which is used to implement any one of the automatic driving control methods based on artificial intelligence, and is characterized in that it includes a vehicle modeling module, a model creation module, an action execution calculation model, a data acquisition module, a data preprocessing module, a control command generation module, a real-time posture measurement module, a posture comparison module and a parameter adjustment amount calculation module; The model creation module is connected to the action execution calculation model, the data acquisition module is connected to the data preprocessing module, the vehicle modeling module and the data preprocessing module are both connected to the action execution calculation model, the action execution calculation model is connected to the control command generation module, the real-time posture measurement module is connected to the posture comparison module, and the posture comparison module is connected to the parameter adjustment amount calculation module; The vehicle modeling module establishes a vehicle coordinate system and models a vehicle kinematic model based on the vehicle coordinate system; The model creation module is used to create an action execution calculation model; The action execution calculation model is used to calculate the execution action based on the current environment data, and transmit the calculation result of the execution action to the control command generation module; The data acquisition module is used to acquire the actual environment data during driving based on big data, and to acquire the current environment data in the vehicle automatic driving project in real time; The data preprocessing module is used to perform data preprocessing on the collected environmental data, and input the preprocessed environmental data into the trained action execution calculation model; The control command generation module is used to convert the execution action into a corresponding control command for the vehicle execution system; The real-time posture measurement module is used to measure the actual posture of the vehicle in real time; The posture comparison module is used to compare the actual posture of the vehicle with the theoretical posture of the execution action, and calculate the geometric parameter error between the actual posture and the theoretical posture of the execution action; The parameter adjustment amount calculation module is used to reversely calculate the real-time parameter adjustment amount in the control command based on the geometric parameter error, and reversely transmit the actual parameter adjustment amount to the control command generation module.

[0017] The beneficial effects of the present invention include: The automatic driving control method and system based on artificial intelligence provided by the present invention, through vehicle kinematic modeling, creates and trains an action execution calculation model; collects the current environmental data in the vehicle automatic driving project in real time and inputs it into the trained action execution calculation model; the action execution calculation model calculates the execution action and transmits it to the control command generation module, converting it into a corresponding control command for the vehicle execution system; the vehicle execution system performs the action based on the control command, measures the actual posture of the vehicle, and compares it with the theoretical posture, and calculates the geometric parameter error between the actual posture and the theoretical posture; reversely calculates the real-time parameter adjustment amount in the control command and transmits it to the control command generation module to generate a new control command and transmit it to the vehicle execution system. The safety risk of intelligent driving is reduced by reducing the error between the actual action and the theoretical action during intelligent driving.

[0018] First, by establishing the vehicle coordinate system, the vehicle is simplified into a two-wheel model, the two front wheels are simplified into one front wheel, and the two rear wheels are simplified into one rear wheel; the simplified two-wheel model is inverse kinematically solved, and the inverse kinematics solution formula is created. The vehicle's motion model is simplified and modeled, which improves the modeling efficiency and reduces the complexity of the subsequent error calculation between the actual posture and the theoretical posture, thereby improving the efficiency of the execution adjustment of the execution system.

[0019] Second, a small portion of data is randomly selected from the actual environment data for labeling, the small portion of data is input into the action execution calculation model for prediction, and the unlabeled actual environment data is predicted through the action execution calculation model; the prediction results of the unlabeled actual environment data are scored according to specified rules, and the actual environment data corresponding to the score value greater than the specified threshold is labeled; the labeled actual environment data is added to the labeled small portion of data to construct a training sample process, which only requires a small amount of labeled data, and is based on the automatic selection and prediction of unlabeled data during the training process, thereby improving the performance of the model while improving the efficiency of model training.

[0020] Third, by removing the noise data in the collected environmental data and identifying and eliminating the outliers, the influence of noise data and abnormal data on model training is reduced. By processing the features of environmental data in different forms from different sources, the extracted image features and sound features are converted into text data; the text data is converted into vector data, the text data is segmented to obtain individual words, a vocabulary is constructed for the words obtained by the text data segmentation, a unique index is assigned to each word, the frequency of the words is counted, and a text vector is constructed based on the words and the corresponding frequency, so that the computer can directly process it; the text vector is standardized to eliminate the differences between different text lengths and feature scales, and the text vector is scaled to the same specified length to facilitate model training.

[0021] Fourth, the action execution calculation model includes an input layer, a hidden layer and an output layer, wherein the hidden layer includes a combination layer of a convolution layer and a pooling layer, an adaptive transformation layer and a flow projection layer, wherein the convolution layer extracts the spatial hierarchical features of the image data from the input current environment data collected in real time; the pooling layer reduces the spatial dimension of the spatial hierarchical features output by the previous convolution layer; the adaptive transformation layer performs pixel transformation processing on the image frames with the spatial hierarchical features having reduced spatial dimensions, and extracts the driving features of the image data; the projection layer performs flow displacement on the image frames before and after the image frames subjected to pixel transformation processing, and projects the image frames subjected to pixel transformation processing to simulate the flow between the image frames subjected to pixel transformation processing, so that the accuracy of subsequent motion estimation and error compensation can be improved based on the driving features of the image data.

[0022] Fifth, when the flow projection layer projects the image frame processed with pixel transformation to simulate the flow between the image frames processed with pixel transformation, when there are multiple flow vectors projected to the same position of the central image frame, the multiple flow vectors at the position are averaged; when there are multiple hole positions where the flow vectors are not projected, the hole positions are filled in the designated image frame from the outer edge of the hole to the inside, so that the model has better robustness.

[0023] Sixth, in step S6, the real-time parameter adjustment amount in the control command is reversely calculated based on the geometric parameter error, the continuous parameter adjustment amount data within the specified period is obtained, and a mapping relationship table between the environmental data and the parameter adjustment amount is constructed. After the real-time environmental data is collected, the mapping relationship table is retrieved to obtain the corresponding parameter adjustment amount, and then the parameter adjustment amount is reversely transmitted to the control command generation module to generate a new control command, which is transmitted to the vehicle execution system. By establishing a mapping relationship table, the parameter adjustment amount in the current environment can be quickly obtained, and the execution efficiency of the vehicle execution system can be improved while ensuring the accuracy of the execution action of the vehicle execution system.

[0024] Seventh, obtain continuous parameter adjustment data within a specified period and obtain parameter adjustment data samples. Build a parameter adjustment prediction model, input the parameter adjustment data samples and corresponding environmental data into the parameter adjustment prediction model to train the model. The parameter adjustment model is a neural network model, which automatically captures the relationship between learning environment data and parameter adjustment amounts during model training. The parameter adjustment prediction model predicts the corresponding parameter adjustment amount based on real-time environmental data. By establishing a neural network parameter adjustment prediction model, the parameter adjustment amount in the current environment can be quickly predicted, while ensuring the accuracy of the vehicle execution system's execution actions, while improving the execution efficiency of the vehicle execution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the automatic driving control method based on artificial intelligence of the present invention.

[0026] Figure 2 It is a schematic diagram of the process of constructing training samples based on actual environment data according to the present invention.

[0027] Figure 3 Schematic diagram of the architecture of the artificial intelligence-based automatic driving control system of the present invention.

[0028] Figure 4 It is a simplified schematic diagram of the two-wheel model structure of the present invention. DETAILED DESCRIPTION

[0029] The following is combined with Figure 1~Figure 3 The present invention is further described in detail: Example 1 See attached Figure 1 As shown, an automatic driving control method based on artificial intelligence includes the following steps: S1: establishing a vehicle coordinate system, and modeling a vehicle kinematics model based on the vehicle coordinate system; S2: creating an action execution calculation model, acquiring actual environment data during driving based on big data, constructing training samples based on the actual environment data, and inputting the training samples into the action execution calculation model for model training; S3: collecting current environmental data in the vehicle automatic driving project in real time, performing data preprocessing on the collected environmental data, and inputting the preprocessed environmental data into the trained action execution calculation model; S4: the action execution calculation model calculates the execution action based on the current environment data, and transmits the calculation result of the execution action to the control command generation module, and the control command generation module converts the execution action into a corresponding control command for the vehicle execution system; S5: The vehicle execution system executes the corresponding action based on the control command, and the real-time posture measurement module measures the actual posture of the vehicle, compares the actual posture of the vehicle with the theoretical posture of the execution action in step S4, and calculates the geometric parameter error between the actual posture and the theoretical posture of the execution action; S6: reversely calculating the real-time parameter adjustment amount in the control command based on the geometric parameter error, and reversely transmitting the real-time parameter adjustment amount to the control command generation module, the control command generation module generates a new control command based on the real-time parameter adjustment amount, and transmits the new control command to the vehicle execution system; S7: Iteratively execute steps S3 to S6 to perform real-time adjustment and control during the automatic driving process of the vehicle.

[0030] Due to the actual road conditions or unexpected situations encountered during the autonomous driving process, the intelligent driving system will first determine the action that the vehicle should perform and the magnitude of the action based on the current actual road conditions or unexpected situations after a series of internal algorithms, generate corresponding control commands, and issue corresponding control commands to the vehicle. Then the vehicle's action execution system executes the relevant actions based on the control commands of the intelligent driving system. When the execution system is executing an action, there is an error between the actual action and the action that the intelligent driving system expects the execution system to execute. When the error between the actual action and the action that the intelligent driving system expects the execution system to execute exceeds the specified range, there will be safety risks in the process of intelligent driving.

[0031] Therefore, the present invention creates and trains an action execution calculation model through vehicle kinematic modeling; collects the current environmental data in the vehicle automatic driving project in real time and inputs it into the trained action execution calculation model; the action execution calculation model calculates the execution action and transmits it to the control command generation module, converting it into a corresponding control command for the vehicle execution system; the vehicle execution system performs the action based on the control command, measures the actual posture of the vehicle, and compares it with the theoretical posture, and calculates the geometric parameter error between the actual posture and the theoretical posture; reversely calculates the real-time parameter adjustment amount in the control command and transmits it to the control command generation module to generate a new control command and transmits it to the vehicle execution system. By reducing the error between the actual action in the intelligent driving system and the action that the intelligent driving system expects the execution system to perform, the safety risk of the intelligent driving process is reduced, and the safety and reliability of intelligent driving are improved.

[0032] In this embodiment, see Figure 4 The specific process of modeling the vehicle kinematic model in step S1 is as follows: S11: In the established vehicle coordinate system, simplify the vehicle into a two-wheel model, the two front wheels are simplified into one front wheel, and the two rear wheels are simplified into one rear wheel. The reason why the vehicle can be simplified into a two-wheel model is that the left and right tires of the vehicle have the same steering angle and speed at any time, ignore the change in vehicle speed, ignore the transfer of front and rear axle loads, and assume that the body and suspension system are rigid systems. The two-wheel model makes it simple and efficient to describe the movement of the vehicle on a two-dimensional plane. This makes the model more intuitive and easy to calculate when describing the movement of the vehicle, and can also provide sufficient accuracy to describe and control the movement of the vehicle.

[0033] S12: Perform inverse kinematics solution on the simplified two-wheel model and create an inverse kinematics solution formula, as follows: cosθ1= a tan( y , x ); ; ; θ3=π- a cos[( a 2 + b 2 - c 2 ) / 2 ab ]; The front wheel and the rear wheel are simplified into two long strips, and θ1 is the line connecting the frontmost point of the front wheel and the rearmost point of the rear wheel and the established coordinate system. xThe angle between the axes, θ2 is the angle between the straight line where the rear wheel is located and x The angle between the axes, θ3 is the angle between the straight line where the rear wheel is located and the line connecting the frontmost point of the front wheel and the rearmost point of the rear wheel, a is the distance between the intersection of the straight line where the front wheel is and the straight line where the rear wheel is and the last point of the rear wheel, b It is the distance between the intersection of the straight line where the front wheels are located and the straight line where the rear wheels are located and the frontmost point of the front wheel.

[0034] Example 2 On the basis of Example 1, the specific process of constructing the training sample based on the actual environment data in step S2 is as follows: S21: randomly selecting a designated small portion of data from the actual environment data for labeling, inputting the small portion of data into the action execution calculation model for prediction, and predicting the unlabeled actual environment data through the action execution calculation model; S22: Score the prediction results of the unlabeled actual environment data according to the specified rules, and label the actual environment data whose score value is greater than the specified threshold; S23: Add the actual environment data annotated in step S22 to the small part of the data annotated in step S21 to construct a training sample.

[0035] A small portion of data is randomly selected from the actual environment data for labeling, and the small portion of data is input into the action execution calculation model for prediction, and the unlabeled actual environment data is predicted through the action execution calculation model; the prediction results of the unlabeled actual environment data are scored according to specified rules, and the actual environment data corresponding to the score value greater than the specified threshold is labeled; the labeled actual environment data is added to the labeled small portion of data to construct a training sample process. Only a small amount of labeled data is needed. Based on the automatic selection and prediction of unlabeled data during the training process, the model performance is improved while the model training efficiency is improved.

[0036] In this embodiment, the specific process of performing data preprocessing on the collected environmental data in step S3 is as follows: S31: removing noise data from the collected environmental data, calculating the standard deviation of the environmental data after removing abnormal data, marking data points in the environmental data whose values ​​exceed the mean value plus or minus a specified number of standard deviations as abnormal values, and removing them; S32: performing feature processing on the environmental data of different forms with different sources from which the noise data and the abnormal data are removed, including extracting image features from the image data and sound features from the sound data, and converting the extracted image features and sound features into text data; S33: converting the text data into vector data, segmenting the text data to obtain individual words, constructing a vocabulary for the words obtained by segmenting the text data, assigning a unique index to each word, counting the frequency of the word, and constructing a text vector based on the word and the corresponding frequency; S34: performing normalization processing on the text vectors and scaling processing on the text vectors to scale the text vectors to the same specified length.

[0037] The above process reduces the impact of noise data and abnormal data on model training by removing noise data from the collected environmental data and identifying and eliminating abnormal values. By processing the features of environmental data in different forms from different sources, the extracted image features and sound features are converted into text data; the text data is converted into vector data, the text data is segmented to obtain individual words, a vocabulary is constructed for the words obtained by segmenting the text data, a unique index is assigned to each word, the frequency of the words is counted, and a text vector is constructed based on the words and the corresponding frequency, so that the computer can directly process it; the text vector is standardized to eliminate the differences between different text lengths and feature scales, and the text vector is scaled to the same specified length to facilitate model training.

[0038] Example 3 Based on Example 1 or Example 2, the action execution calculation model described in step S4 includes an input layer, a hidden layer and an output layer, the hidden layer includes a combination layer of a convolutional layer and a pooling layer, an adaptive transformation layer and a stream projection layer, the adaptive transformation layer is connected to the stream projection layer, and is located after the combination layer of the convolutional layer and the pooling layer.

[0039] The input layer is used to input the current environment data collected in real time; the convolution layer is used to extract the spatial hierarchical features of the image data from the input current environment data collected in real time; the pooling layer is used to reduce the spatial dimension of the spatial hierarchical features output by the previous convolution layer; the adaptive transformation layer is used to perform pixel transformation processing on the image frames with the spatial hierarchical features with reduced spatial dimensions, and extract the image data driving features; the flow projection layer is used to perform flow displacement on the image frames before and after the image frames undergoing pixel transformation processing, and project the image frames undergoing pixel transformation processing to simulate the flow between the image frames undergoing pixel transformation processing.

[0040] In this embodiment, the adaptive transformation layer uses the interpolation kernel and optical flow to perform pixel transformation processing on the image frame with reduced spatial hierarchical features of spatial dimensions by fusing bilinear deformation and kernel convolution in single-frame image processing, and extracts image data driving features. When the flow projection layer projects the image frame processed by pixel transformation to simulate the flow between the image frames processed by pixel transformation, when there are multiple flow vectors projected to the same position of the central image frame, the multiple flow vectors at the position are averaged; when there are multiple hole positions where the flow vectors are not projected, the hole positions are filled in the designated image frame by the outer edge of the hole.

[0041] In this embodiment, the action execution calculation model includes an input layer, a hidden layer and an output layer, and the hidden layer includes a combination layer of a convolution layer and a pooling layer, an adaptive transformation layer and a flow projection layer. The convolution layer extracts the spatial hierarchical features of the image data from the current environment data collected in real time; the pooling layer reduces the spatial dimension of the spatial hierarchical features output by the previous convolution layer; the adaptive transformation layer performs pixel transformation processing on the image frames with the spatial hierarchical features with reduced spatial dimensions, and extracts the driving features of the image data; the projection layer performs flow displacement on the image frames before and after the image frames subjected to pixel transformation processing, and projects the image frames subjected to pixel transformation processing to simulate the flow between the image frames subjected to pixel transformation processing, so that the accuracy of subsequent motion estimation and error compensation can be improved based on the driving features of the image data.

[0042] Example 4 On the basis of Example 1 or Example 2 or Example 3, the specific process of calculating the geometric parameter error between the actual posture and the calculation result of the execution action posture in step S5 is: S51: Synchronizing the actual posture to the coordinate system of the theoretical posture according to the actual posture; S52: Calculate the error value between the geometric parameters of the actual posture and the corresponding geometric parameters of the theoretical posture in the coordinate system.

[0043] By projecting the image frames processed with pixel transformation in the flow projection layer to simulate the flow between image frames processed with pixel transformation, when there are multiple flow vectors projected to the same position of the central image frame, the multiple flow vectors at the position are averaged; when there are multiple hole positions where the flow vectors are not projected, the hole positions are filled in the designated image frame with the outer edge of the hole facing inward, so that the model has better robustness.

[0044] An automatic driving control system based on artificial intelligence, used to implement any one of the automatic driving control methods based on artificial intelligence, characterized in that it includes a vehicle modeling module, a model creation module, an action execution calculation model, a data acquisition module, a data preprocessing module, a control command generation module, a real-time posture measurement module, a posture comparison module and a parameter adjustment amount calculation module.

[0045] The model creation module is connected to the action execution calculation model, the data acquisition module is connected to the data preprocessing module, the vehicle modeling module and the data preprocessing module are both connected to the action execution calculation model, the action execution calculation model is connected to the control command generation module, the real-time posture measurement module is connected to the posture comparison module, and the posture comparison module is connected to the parameter adjustment amount calculation module. The vehicle modeling module establishes a vehicle coordinate system and models the vehicle kinematic model based on the vehicle coordinate system. The model creation module is used to create an action execution calculation model. The action execution calculation model is used to calculate the execution action based on the current environmental data, and transmit the calculation result of the execution action to the control command generation module. The data acquisition module is used to obtain the actual environmental data in the driving process based on big data, and obtain the current environmental data in the vehicle automatic driving project for real-time collection.

[0046] The data preprocessing module is used to perform data preprocessing on the collected environmental data, and input the preprocessed environmental data into the trained action execution calculation model. The control command generation module is used to convert the execution action into the corresponding control command for the vehicle execution system. The real-time posture measurement module is used to measure the actual posture of the vehicle in real time. The posture comparison module is used to compare the actual posture of the vehicle with the theoretical posture of the execution action, and calculate the geometric parameter error between the actual posture and the theoretical posture of the execution action. The parameter adjustment amount calculation module is used to reversely calculate the real-time parameter adjustment amount in the control command based on the geometric parameter error, and reversely transmit the actual parameter adjustment amount to the control command generation module.

[0047] Example 5 On the basis of Example 1 or Example 2 or Example 3 or Example 4, in step S6, the real-time parameter adjustment amount in the control command is reversely calculated based on the geometric parameter error, the continuous parameter adjustment amount data within the specified period is obtained, and a mapping relationship table between the environmental data and the parameter adjustment amount is constructed. In the process of intelligent driving, after the real-time environmental data is collected, the mapping relationship table is directly retrieved to obtain the corresponding parameter adjustment amount, and then the parameter adjustment amount is transmitted reversely to the control command generation module, and the control command generation module generates a new control command based on the actual parameter adjustment amount, and transmits the new control command to the vehicle execution system. By establishing a mapping relationship table, the parameter adjustment amount in the current environment can be quickly obtained, and the execution efficiency of the vehicle execution system can be improved while ensuring the accuracy of the execution action of the vehicle execution system.

[0048] In another embodiment, in step S6, the real-time parameter adjustment amount in the control command is reversely calculated based on the geometric parameter error, and the continuous parameter adjustment amount data within the specified period is obtained, and the parameter adjustment data sample is obtained. A parameter adjustment prediction model is constructed, and the parameter adjustment data sample and the corresponding environmental data are input into the parameter adjustment prediction model to train the model. The parameter adjustment model is a neural network model, which automatically captures the relationship between the learning environment data and the parameter adjustment amount during the model training process. Then the real-time environmental data is collected and input into the parameter adjustment prediction model, and the parameter adjustment prediction model automatically outputs the corresponding parameter adjustment amount under the current environmental data. Then the parameter adjustment amount is transmitted in reverse to the control command generation module, and the control command generation module generates a new control command based on the actual parameter adjustment amount, and transmits the new control command to the vehicle execution system. By establishing a parameter adjustment prediction model of a neural network, the parameter adjustment amount under the current environment can be quickly predicted, and the execution efficiency of the vehicle execution system can be improved while ensuring the accuracy of the execution action of the vehicle execution system.

[0049] In summary, the automatic driving control method and system based on artificial intelligence provided by the present invention simplifies the vehicle into a two-wheel model under the established vehicle coordinate system, wherein the two front wheels are simplified into one front wheel, and the two rear wheels are simplified into one rear wheel; the simplified two-wheel model is subjected to inverse kinematics solving, and the inverse kinematics solving formula is created, and the vehicle's motion model is simplified and modeled, thereby improving the modeling efficiency, reducing the subsequent error calculation complexity between the actual posture and the theoretical posture, and thus improving the efficiency of the execution adjustment of the execution system. By randomly selecting a specified small portion of data from the actual environment data for annotation, the small portion of data is input into the action execution calculation model for prediction, and the unlabeled actual environment data is predicted by the action execution calculation model; the prediction results of the unlabeled actual environment data are scored according to the specified rules, and the actual environment data corresponding to the score value greater than the specified threshold is marked; the marked actual environment data is added to the marked small portion of data to construct a training sample process, which only requires a small amount of labeled data, and based on the automatic selection and prediction of unlabeled data during the training process, the performance of the model is improved while improving the model training efficiency.

[0050] By removing the noise data in the collected environmental data and identifying and eliminating the outliers, the influence of noise data and abnormal data on model training is reduced. By processing the features of environmental data in different forms from different sources, the extracted image features and sound features are converted into text data; the text data is converted into vector data, the text data is segmented to obtain individual words, a vocabulary is constructed for the words obtained by the text data segmentation, a unique index is assigned to each word, the frequency of the words is counted, and a text vector is constructed based on the words and the corresponding frequency, so that the computer can directly process it; the text vector is standardized to eliminate the differences between different text lengths and feature scales, and the text vector is scaled to the same specified length to facilitate model training.

[0051] The action execution calculation model includes an input layer, a hidden layer and an output layer, wherein the hidden layer includes a combination layer of a convolution layer and a pooling layer, an adaptive transformation layer and a flow projection layer, wherein the convolution layer extracts the spatial hierarchical features of the image data from the input current environment data collected in real time; the pooling layer reduces the spatial dimension of the spatial hierarchical features output by the previous convolution layer; the adaptive transformation layer performs pixel transformation processing on the image frames with the spatial hierarchical features with reduced spatial dimensions, and extracts the driving features of the image data; the projection layer performs flow displacement on the image frames before and after the image frames subjected to pixel transformation processing, and projects the image frames subjected to pixel transformation processing to simulate the flow between the image frames subjected to pixel transformation processing, so that the accuracy of subsequent motion estimation and error compensation can be improved based on the driving features of the image data. By projecting the image frames processed with pixel transformation in the flow projection layer to simulate the flow between image frames processed with pixel transformation, when there are multiple flow vectors projected to the same position of the central image frame, the multiple flow vectors at the position are averaged; when there are multiple hole positions where the flow vectors are not projected, the hole positions are filled in the designated image frame with the outer edge of the hole facing inward, so that the model has better robustness.

[0052] In step S6, the real-time parameter adjustment amount in the control command is calculated inversely based on the geometric parameter error, the continuous parameter adjustment amount data within the specified period is obtained, and a mapping relationship table between the environmental data and the parameter adjustment amount is constructed. After the real-time environmental data is collected, the mapping relationship table is retrieved to obtain the corresponding parameter adjustment amount, and then the parameter adjustment amount is transmitted inversely to the control command generation module to generate a new control command, which is transmitted to the vehicle execution system. By establishing a mapping relationship table, the parameter adjustment amount in the current environment can be quickly obtained, while ensuring the accuracy of the vehicle execution system's execution action, the execution efficiency of the vehicle execution system is improved at the same time. By obtaining the continuous parameter adjustment amount data within the specified period, a parameter adjustment data sample is obtained. A parameter adjustment prediction model is constructed, and the parameter adjustment data sample and the corresponding environmental data are input into the parameter adjustment prediction model to train the model. The parameter adjustment model is a neural network model, which automatically captures the relationship between the learning environment data and the parameter adjustment amount during the model training process. The parameter adjustment prediction model predicts the corresponding parameter adjustment amount based on the real-time environmental data. By establishing a neural network parameter adjustment prediction model, the parameter adjustment amount in the current environment can be quickly predicted, while ensuring the accuracy of the vehicle execution system's execution action, the execution efficiency of the vehicle execution system is improved at the same time.

Claims

1. An automatic driving control method based on artificial intelligence, characterized in that: The following steps are involved: S1: establishing a vehicle coordinate system, and modeling a vehicle kinematics model based on the vehicle coordinate system; S2: creating an action execution calculation model, acquiring actual environment data during driving based on big data, constructing training samples based on the actual environment data, and inputting the training samples into the action execution calculation model for model training; S3: collecting current environmental data in the vehicle automatic driving project in real time, performing data preprocessing on the collected environmental data, and inputting the preprocessed environmental data into the trained action execution calculation model; S4: the action execution calculation model calculates the execution action based on the current environment data, and transmits the calculation result of the execution action to the control command generation module, and the control command generation module converts the execution action into a corresponding control command for the vehicle execution system; S5: The vehicle execution system executes the corresponding action based on the control command, and the real-time posture measurement module measures the actual posture of the vehicle, compares the actual posture of the vehicle with the theoretical posture of the execution action in step S4, and calculates the geometric parameter error between the actual posture and the theoretical posture of the execution action; S6: reversely calculating the real-time parameter adjustment amount in the control command based on the geometric parameter error, and reversely transmitting the real-time parameter adjustment amount to the control command generation module, the control command generation module generates a new control command based on the real-time parameter adjustment amount, and transmits the new control command to the vehicle execution system; S7: Iteratively execute steps S3 to S6 to perform real-time adjustment and control during the automatic driving process of the vehicle.

2. The automatic driving control method based on artificial intelligence according to claim 1, characterized in that: The specific process of modeling the vehicle kinematics model in step S1 is as follows: S11: In the established vehicle coordinate system, simplify the vehicle into a two-wheel model, simplify the two front wheels into one front wheel, and simplify the two rear wheels into one rear wheel; S12: Perform inverse kinematics solution on the simplified two-wheel model and create an inverse kinematics solution formula, as follows: cosθ1= a tan( y , x ); ； ； θ3=π- a cos[( a 2 + b 2 - c 2 ) / 2 ab ]; The front wheel and the rear wheel are simplified into two long strips, and θ1 is the line connecting the frontmost point of the front wheel and the rearmost point of the rear wheel and the established coordinate system. x The angle between the axes, θ2 is the angle between the straight line where the rear wheel is located and x The angle between the axes, θ3 is the angle between the straight line where the rear wheel is located and the line connecting the frontmost point of the front wheel and the rearmost point of the rear wheel, a is the distance between the intersection of the straight line where the front wheel is and the straight line where the rear wheel is and the last point of the rear wheel, b It is the distance between the intersection of the straight line where the front wheels are located and the straight line where the rear wheels are located and the frontmost point of the front wheel.

3. The automatic driving control method based on artificial intelligence according to claim 1, characterized in that: The specific process of constructing training samples based on the actual environment data in step S2 is as follows: S21: randomly selecting a designated small portion of data from the actual environment data for labeling, inputting the small portion of data into the action execution calculation model for prediction, and predicting the unlabeled actual environment data through the action execution calculation model; S22: Score the prediction results of the unlabeled actual environment data according to the specified rules, and label the actual environment data whose score value is greater than the specified threshold; S23: Add the actual environment data annotated in step S22 to the small part of the data annotated in step S21 to construct a training sample.

4. The automatic driving control method based on artificial intelligence according to claim 3, characterized in that: The specific process of data preprocessing of the collected environmental data in step S3 is as follows: S31: removing noise data from the collected environmental data, calculating the standard deviation of the environmental data after removing abnormal data, marking data points in the environmental data whose values ​​exceed the mean value plus or minus a specified number of standard deviations as abnormal values, and removing them; S32: performing feature processing on the environmental data of different forms with different sources from which the noise data and the abnormal data are removed, including extracting image features from the image data and sound features from the sound data, and converting the extracted image features and sound features into text data; S33: converting the text data into vector data, segmenting the text data to obtain individual words, constructing a vocabulary for the words obtained by segmenting the text data, assigning a unique index to each word, counting the frequency of the word, and constructing a text vector based on the word and the corresponding frequency; S34: performing normalization processing on the text vectors and scaling processing on the text vectors to scale the text vectors to the same specified length.

5. The automatic driving control method based on artificial intelligence according to claim 3, characterized in that: The action execution calculation model in step S4 includes an input layer, a hidden layer and an output layer, wherein the hidden layer includes a combination layer of a convolution layer and a pooling layer, an adaptive transformation layer and a stream projection layer, wherein the adaptive transformation layer is connected to the stream projection layer and is located after the combination layer of the convolution layer and the pooling layer; The input layer is used to input the current environment data collected in real time; The convolution layer is used to extract spatial hierarchical features of image data from the input current environment data collected in real time; The pooling layer is used to reduce the spatial dimension of the spatial hierarchical features output by the previous convolutional layer; The adaptive transformation layer is used to perform pixel transformation processing on the image frame with reduced spatial hierarchical features of spatial dimension, and extract image data driving features; The flow projection layer is used to perform flow displacement on the image frames before and after the image frames undergoing pixel transformation processing, and to project the image frames undergoing pixel transformation processing to simulate the flow between the image frames undergoing pixel transformation processing.

6. The automatic driving control method based on artificial intelligence according to claim 5, characterized in that: The adaptive transformation layer uses the interpolation kernel and optical flow to perform pixel transformation processing on the image frame with reduced spatial hierarchical features of spatial dimensions in a way of fusing bilinear deformation and kernel convolution in single-frame image processing, and extracts image data driving features.

7. The automatic driving control method based on artificial intelligence according to claim 5, characterized in that: When, in the process of projecting the pixel transformed image frame in the flow projection layer to simulate the flow between the pixel transformed image frames, there are multiple flow vectors projected to the same position of the central image frame, then the multiple flow vectors at the position are averaged; When there are multiple hole positions where flow vectors are not projected, the hole positions are filled inward from the outer edge of the hole.

8. The automatic driving control method based on artificial intelligence according to claim 1, characterized in that: The specific process of calculating the geometric parameter error between the actual posture and the calculated result of the executed action posture in step S5 is: S51: Synchronizing the actual posture to the coordinate system of the theoretical posture according to the actual posture; S52: Calculate the error value between the geometric parameters of the actual posture and the corresponding geometric parameters of the theoretical posture in the coordinate system.

9. An artificial intelligence-based automatic driving control system, used to implement an artificial intelligence-based automatic driving control method according to any one of claims 1 to 8, characterized in that: It includes a vehicle modeling module, a model creation module, an action execution calculation model, a data acquisition module, a data preprocessing module, a control command generation module, a real-time posture measurement module, a posture comparison module and a parameter adjustment amount calculation module; The model creation module is connected to the action execution calculation model, the data acquisition module is connected to the data preprocessing module, the vehicle modeling module and the data preprocessing module are both connected to the action execution calculation model, the action execution calculation model is connected to the control command generation module, the real-time posture measurement module is connected to the posture comparison module, and the posture comparison module is connected to the parameter adjustment amount calculation module; The vehicle modeling module establishes a vehicle coordinate system and models a vehicle kinematic model based on the vehicle coordinate system; The model creation module is used to create an action execution calculation model; The action execution calculation model is used to calculate the execution action based on the current environment data, and transmit the calculation result of the execution action to the control command generation module; The data acquisition module is used to acquire the actual environment data during driving based on big data, and to acquire the current environment data in the vehicle automatic driving project in real time; The data preprocessing module is used to perform data preprocessing on the collected environmental data, and input the preprocessed environmental data into the trained action execution calculation model; The control command generation module is used to convert the execution action into a corresponding control command for the vehicle execution system; The real-time posture measurement module is used to measure the actual posture of the vehicle in real time; The posture comparison module is used to compare the actual posture of the vehicle with the theoretical posture of the execution action, and calculate the geometric parameter error between the actual posture and the theoretical posture of the execution action; The parameter adjustment amount calculation module is used to reversely calculate the real-time parameter adjustment amount in the control command based on the geometric parameter error, and reversely transmit the actual parameter adjustment amount to the control command generation module.

Images