Vehicle driving control method, device, electronic equipment and storage medium

Abstract

The vehicle driving control method, device, electronic device and storage medium proposed in the embodiments of the present application include: first, obtaining image information and navigation information of the vehicle at the current moment during driving; second, performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory; then, performing multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision, and obtaining a preliminary control decision based on the combination of the first control decision and the second control decision; next, inputting the image information into a collision correction model for data processing to obtain a safety control decision, and performing integration processing based on the safety control decision and the preliminary control decision to obtain a target control decision; finally, driving control of the vehicle is performed based on the target control decision, which can improve the control reliability of the scheduled route and the safety of operation control.

Description

Technical Field

[0001] The present application relates to the field of vehicle control technology, and in particular to a vehicle driving control method, device, electronic device and storage medium. Background Art

[0002] As more and more commercial robots or embodied intelligent systems begin to be deployed in practical applications, the most typical embodied intelligent robot is the self-driving car. Some cities have begun to deploy self-driving vehicles on public roads, which shows that autonomous driving has made significant progress.

[0003] In the related art, intelligent automatic driving for vehicles is usually carried out by installing sensors and processors on the vehicle, and then processing the real-time data of the vehicle's surroundings collected by the sensors and the preset operation path to generate appropriate control decisions to control the operation of the vehicle. However, this operation control decision relies on the input of fixed data format and fixed control strategy. However, the actual road conditions are usually more complicated, and there are situations where pedestrians or vehicles suddenly appear in the blind spot of the sensor's line of sight. The originally generated control decision is difficult to adapt to complex road conditions, changes in traffic rules or different driving behaviors due to its fixed control thinking limitations, and it is impossible to avoid all sudden risks, resulting in low safety of vehicle control based on the control decision. Summary of the invention

[0004] The embodiments of the present application provide a vehicle driving control method, device, electronic device and storage medium, which can improve the safety of vehicle operation control.

[0005] To achieve the above objectives, a first aspect of an embodiment of the present application provides a vehicle driving control method, the method comprising:

[0006] Obtain the image information and navigation information of the vehicle at the current moment during driving;

[0007] Performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory;

[0008] Performing a multi-step control process based on the navigation information and the predicted trajectory to obtain a second control decision, and obtaining a preliminary control decision based on a combination of the first control decision and the second control decision;

[0009] Inputting the image information into a collision correction model for data processing to obtain a safety control decision, and integrating the safety control decision and the preliminary control decision to obtain a target control decision;

[0010] The vehicle is driven and controlled based on the target control decision.

[0011] In some embodiments, the image information includes a current image frame, and performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory, includes:

[0012] Extracting features from the current image frame to obtain path prediction features;

[0013] Inputting the path prediction feature and the navigation information into a path prediction network to obtain the predicted trajectory, wherein the predicted trajectory includes predicted path points at multiple consecutive moments;

[0014] A path vector between every two adjacent predicted path points is calculated, and the first control decision is obtained based on the path vector and PID control.

[0015] In some embodiments, performing multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision includes:

[0016] Performing attention processing on the path prediction feature and the navigation information to obtain an initial fusion feature;

[0017] Based on the initial fusion feature and the plurality of predicted path points, generating prediction decision features at a plurality of consecutive moments;

[0018] Based on the plurality of prediction decision features, a prediction action at each continuous moment is generated, and the second control decision is obtained based on all the prediction actions.

[0019] In some embodiments, generating a plurality of prediction decision features at consecutive moments based on the initial fusion features and the plurality of predicted path points comprises:

[0020] At each continuous moment, based on the probability distribution of the initial fusion feature and the superposition processing of the corresponding predicted path point, an updated fusion feature is obtained;

[0021] Performing splicing processing based on the updated fusion feature and the navigation information to obtain a spliced ​​fusion feature;

[0022] Attention processing is performed on the spliced ​​fusion features to obtain the prediction decision features corresponding to the continuous moments.

[0023] In some embodiments, obtaining a preliminary control decision based on the combination of the first control decision and the second control decision includes:

[0024] Obtaining a first control factor, and obtaining a second control factor based on one minus the first control factor, the first control factor being less than one-half;

[0025] When the image information indicates that the vehicle is in a straight-moving state, the preliminary control decision is obtained based on the product of the first control factor and the second control decision, plus the product of the second control factor and the first control decision;

[0026] When the image information represents that the vehicle is in a turning state, the preliminary control decision is obtained based on the product of the first control factor and the first control decision plus the product of the second control factor and the second control decision.

[0027] In some embodiments, the step of constructing the collision correction model includes:

[0028] Acquire a plurality of sample image information, and generate a state space based on the sample image information, each of the sample image information including a plurality of continuous image frames;

[0029] generating an action space based on control parameters of the vehicle;

[0030] Obtaining an anti-collision objective function, and generating a reward model based on the anti-collision objective function;

[0031] Obtaining initial action network parameters of the action network and initial evaluation network parameters of the evaluation network;

[0032] An initial collision correction model is generated based on the state space, the action space, the reward model, the initial action network parameters and the initial evaluation network parameters, and the initial collision correction model is trained for multiple rounds, and the collision correction model is obtained based on the trained initial collision correction model.

[0033] In some embodiments, obtaining the anti-collision objective function includes:

[0034] generating a centralization factor function based on the degree of deviation of the driving center of the vehicle during road driving;

[0035] generating an angle factor function based on a driving angle of the vehicle during road driving;

[0036] generating a target factor function based on a travel distance between the vehicle and a target travel position during road travel;

[0037] The centering factor function, the angle factor function and the target factor function are cumulatively multiplied to generate the anti-collision target function.

[0038] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a vehicle driving control device, the device comprising:

[0039] An information acquisition module is used to acquire image information and navigation information of the vehicle at the current moment during driving;

[0040] A first control decision generating module, configured to perform trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtain a first control decision based on the predicted trajectory;

[0041] a second control decision generating module, configured to perform a multi-step control process based on the navigation information and the predicted trajectory to obtain a second control decision, and to obtain a preliminary control decision based on a combination of the first control decision and the second control decision;

[0042] a target control decision generating module, configured to input the image information into a collision correction model for data processing to obtain a safety control decision, and to integrate the safety control decision and the preliminary control decision to obtain a target control decision;

[0043] A vehicle control module is used to control the driving of the vehicle based on the target control decision.

[0044] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the vehicle driving control method described in the first aspect when executing the computer program.

[0045] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a storage medium, which is a computer-readable storage medium, and the storage medium stores a computer program. When the computer program is executed by a processor, it implements the vehicle driving control method described in the first aspect above.

[0046] The vehicle driving control method, device, electronic device and storage medium proposed in the embodiments of the present application include: first, obtaining image information and navigation information of the vehicle at the current moment during driving; second, performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory; then, performing multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision, and obtaining a preliminary control decision based on the combination of the first control decision and the second control decision; next, inputting the image information into a collision correction model for data processing to obtain a safety control decision, and performing integration processing based on the safety control decision and the preliminary control decision to obtain a target control decision; finally, performing driving control on the vehicle based on the target control decision. The embodiment of the present application utilizes the predicted trajectory obtained by performing trajectory prediction based on image information around the vehicle to generate a corresponding first control decision, and further utilizes the predicted trajectory for guidance to generate a predicted multi-step control decision that matches the control navigation information, thereby generating a preliminary control decision that accurately corresponds to the original control information; in addition, a collision correction model that can be trained and generated in advance based on reinforcement learning for complex road environments is utilized to generate a safety control decision to prevent the vehicle from being hit, so as to utilize the high learning adaptability of reinforcement learning to avoid safety accidents as much as possible when encountering sudden risks, thereby improving the control reliability of the scheduled route and the safety of operation control when the vehicle is controlled using the preliminary control decision and the safety control decision.

[0047] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a vehicle driving control method provided in one embodiment of the present application.

[0049] Figure 2 yes Figure 1 Flow chart of step 102 in FIG.

[0050] Figure 3 yes Figure 1 Flow chart of step 103 in FIG.

[0051] Figure 4 yes Figure 3 Flow chart of step 302 in FIG.

[0052] Figure 5 yes Figure 1 Another flow chart of step 103 in FIG.

[0053] Figure 6It is a flowchart of generating a preliminary control decision provided by another embodiment of the present application.

[0054] Figure 7 This is a flow chart for generating a collision correction model provided in yet another embodiment of the present application.

[0055] Figure 8 yes Figure 7 Flowchart of step 703 in FIG.

[0056] Fig. 9 It is a schematic diagram of the training process of a deep reinforcement learning model provided in one embodiment of the present application.

[0057] Fig.10 It is a flowchart schematically illustrating a vehicle driving control method provided in one embodiment of the present application.

[0058] Fig.11 It is a structural schematic diagram of a vehicle driving control device provided in one embodiment of the present application.

[0059] Fig.12 It is a schematic diagram of the hardware structure of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0061] It should be noted that although the functional modules are divided in the device schematic and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0063] As more and more commercial robots or embodied intelligent systems begin to be deployed in practical applications, the most typical embodied intelligent robot is the self-driving car. Some cities have begun to deploy self-driving vehicles on public roads, which shows that autonomous driving has made significant progress. Despite the recent significant progress in the field of autonomous driving, current methods still have difficulties when encountering long-tail unforeseen events and challenging urban scenarios, which may cause serious accidents. At present, research on autonomous driving cars based on large language models has emerged, which aims to solve problems such as explainability, common sense reasoning, and interaction with humans.

[0064] Pedestrians or vehicles suddenly appearing from blind spots when vision is obstructed refers to a complex traffic phenomenon that poses an emergency avoidance challenge to the driver. Complex scenarios in autonomous driving are extremely challenging because they require the system to make fast and precise decisions in a very short period of time to deal with the sudden appearance of pedestrians or vehicles. The importance of decision-making in this scenario lies in its direct relationship to the safety of road users, and it also tests the perception ability, prediction accuracy and reaction speed of the autonomous driving system. Effective decision-making can not only prevent potential collisions, but also enhance the public's trust in autonomous driving technology, thereby promoting its widespread application in the real world. Therefore, developing autonomous driving systems that can accurately identify and properly handle complex situations is crucial to achieving safe and reliable autonomous driving.

[0065] In the related art, intelligent automatic driving for vehicles is usually carried out by installing sensors and processors on the vehicle, and then processing the real-time data of the vehicle's surroundings collected by the sensors and the preset operation path to generate appropriate control decisions to control the operation of the vehicle. However, this operation control decision relies on the input of fixed data format and fixed control strategy. However, the actual road conditions are usually more complicated, and there are situations where pedestrians or vehicles suddenly appear in the blind spot of the sensor's line of sight. The originally generated control decision is difficult to adapt to complex road conditions, changes in traffic rules or different driving behaviors due to its fixed control thinking limitations, and it is impossible to avoid all sudden risks, resulting in low safety of vehicle control based on the control decision.

[0066] In order to improve the safety of vehicle operation control, the embodiment of the present application uses the predicted trajectory obtained by trajectory prediction based on the image information around the vehicle to generate a corresponding first control decision, and further uses the predicted trajectory for guidance to generate a predicted multi-step control decision that matches the control navigation information, thereby generating a preliminary control decision that accurately corresponds to the original control information; in addition, a collision correction model that can be trained and generated in advance based on reinforcement learning for complex road environments is used to generate a safety control decision to prevent the vehicle from being hit, so as to utilize the high learning adaptability of reinforcement learning to avoid safety accidents as much as possible when encountering sudden risks, so that when the vehicle is controlled using the preliminary control decision and the safety control decision, the control reliability of the scheduled route and the safety of operation control can be improved.

[0067] The vehicle driving control method, device, electronic device and storage medium provided by the present application will be further described below. The vehicle driving control method provided in the embodiments of the present application can be applied to a processor on a vehicle or an intelligent terminal that communicates and interacts with the vehicle.

[0068] The vehicle driving control method in the embodiment of the present application will be described in detail below. Figure 1 , is an optional flow chart of the vehicle driving control method provided in an embodiment of the present application, Figure 1 The method may include but is not limited to steps 101 to 105. It can also be understood that this embodiment Figure 1 The order of step 101 to step 105 is not specifically limited, and the order of steps can be adjusted or some steps can be reduced or added according to actual needs.

[0069] Step 101: Acquire the image information and navigation information of the vehicle at the current moment during its driving process.

[0070] The following is a detailed description of step 101.

[0071] In some embodiments, when the vehicle is driving, the image information at the current moment is collected through sensors installed on the vehicle (such as traditional cameras, etc.), and the navigation information required in advance according to the user settings is obtained.

[0072] The navigation information includes the vehicle's speed, navigation target point and high-level control instructions at each moment. Image information includes single-frame images and multi-frame images. Here, multi-frame refers to the current image frame corresponding to the current moment and the historical image frame of the nearest adjacent past moment, which can generally be 3 frames. In the navigation information, the speed is the current speed of the vehicle, the navigation target point is the location that the vehicle hopes to reach at the next moment, and the high-level control command is the high-level planning guidance of the intelligent body, such as turning left, turning right, keeping lane, changing lanes right, etc.

[0073] Step 102: performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory.

[0074] The following is a detailed description of step 102.

[0075] In some embodiments, in order to first meet the user's control needs, after obtaining the image information and navigation information, trajectory prediction processing will be performed based on the image information at the current moment to obtain the corresponding predicted trajectory after the vehicle runs at the current moment according to the current speed and driving direction, so as to subsequently generate the first control decision of the vehicle based on the predicted trajectory.

[0076] The following will further describe how to perform trajectory prediction based on image information.

[0077] Reference Figure 2 , performing trajectory prediction processing based on image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory, including the following steps 201 to 203.

[0078] Step 201: extract features from the current image frame to obtain path prediction features.

[0079] Step 202: Input the path prediction features and navigation information into the path prediction network to obtain a predicted trajectory.

[0080] Step 203: Calculate the path vector between every two adjacent predicted path points, and obtain a first control decision based on the path vector and PID control.

[0081] Steps 201 to 203 are described in detail below.

[0082] In some embodiments, unlike control prediction which directly predicts control actions, after obtaining the current image frame and navigation information (including speed v, navigation target point g, and high-level control instructions x) at the current moment, a CNN-based image encoder (such as ResNet) is used to extract features of the current image frame to obtain a path prediction feature F.

[0083] At the same time, the data in the navigation information (including speed v, navigation target point g, and high-level control instruction x) are spliced ​​to form the measurement input m, and then the MLP-based measurement encoder takes the measurement input m as input to output the measurement feature .

[0084] Next, the path prediction feature F is averaged and pooled with the measurement feature obtained from the navigation information. Concatenate to get prediction features , and then use the predicted feature In the input path prediction network GRU, the predicted path points of multiple (i.e., K) consecutive moments in the future are obtained one by one in an autoregressive manner, and these predicted path points form a predicted trajectory.

[0085] Then, two PID controllers are used to generate longitudinal and lateral control decisions respectively. That is, the path vectors between every two adjacent predicted path points in the predicted trajectory are first calculated, and these path vectors are used to represent the expected speed value; then these path vectors are sent to the two PID controllers to generate longitudinal and lateral control strategies respectively, so as to combine and obtain the first control decision .

[0086] Through the above steps 201 to 203, the path prediction features obtained by feature extraction from the current image frame are used to characterize the current image information feature data during the vehicle operation process, and then combined with the navigation information to perform accurate path prediction to obtain a predicted path that matches the current operation of the vehicle and also matches the user's needs, thereby combining with the PID controller to generate a first control decision containing longitudinal and lateral control data that matches the predicted path, thereby improving the reliability of the vehicle driving control method.

[0087] Step 103: Perform multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision, and obtain a preliminary control decision based on the combination of the first control decision and the second control decision.

[0088] The following is a detailed description of step 103.

[0089] In some embodiments, after obtaining a predicted path that matches the current operation of the vehicle and the user's needs, in order to further improve the safety and reliability of the vehicle in turning control, a multi-step control process will be performed based on the navigation information and the predicted trajectory to obtain a second control decision that is safer and more reliable for the vehicle when performing turning control, so that a preliminary control decision that meets both the user's control operation needs and the safety of turning control can be obtained based on the subsequent combination of the first control decision and the second control decision.

[0090] The following will further describe how to perform multi-step control processing based on navigation information and predicted trajectory.

[0091] Reference Figure 3 , performing multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision, including the following steps 301 to 303.

[0092] Step 301: Perform attention processing on the path prediction features and navigation information to obtain initial fusion features.

[0093] Step 302: Generate prediction decision features at multiple consecutive moments based on the initial fusion features and multiple predicted path points.

[0094] Steps 301 to 302 are described in detail below.

[0095] In some embodiments, after obtaining the path prediction feature F, the hidden state corresponding to the path prediction feature F is Hidden state corresponding to navigation information Perform attention processing to obtain the initial fusion features As shown in the following formula (1).

[0096] (1)

[0097] Among them, MLP is attention processing. It is understandable that hidden state is a core concept in deep learning and sequence modeling (such as recurrent neural network RNN, long short-term memory network LSTM, gated recurrent unit GRU, etc.). It represents the state information that is stored internally and cannot be directly observed when the model processes sequence data. The hidden state can be understood as the model's "memory" or "summary" of the current input and historical input.

[0098] Next, based on the initial fusion feature And multiple predicted path points of the predicted path can further generate prediction decision features at multiple consecutive moments , as described below.

[0099] Reference Figure 4 , based on the initial fusion features and multiple predicted path points, generating prediction decision features at multiple consecutive moments, including the following steps 401 to 403.

[0100] Step 401: At each continuous moment, based on the probability distribution of the initial fusion feature and the superposition processing of the corresponding predicted path points, an updated fusion feature is obtained.

[0101] Step 402: Perform stitching processing based on the updated fusion feature and the navigation information to obtain a stitching fusion feature.

[0102] Step 403: Perform attention processing on the concatenated fusion features to obtain prediction decision features corresponding to consecutive moments.

[0103] Steps 401 to 403 are described in detail below.

[0104] In some embodiments, first, at each predicted continuous moment, based on the initial fusion feature The probability distribution of The superposition processing of the corresponding predicted path points is used to obtain the updated fusion feature , and then update the hidden state corresponding to the fusion feature and navigation information Perform splicing processing to obtain splicing fusion features Finally, the concatenated fusion features are processed with attention to obtain the prediction decision features corresponding to the continuous moments As shown in the following formula (2).

[0105] (2)

[0106] Among them, the prediction decision feature Used to contain a combined representation of both static and dynamic information about the environment and the vehicle.

[0107] Step 303: Based on multiple prediction decision features, generate a prediction action for each continuous moment, and obtain a second control decision based on all the prediction actions.

[0108] Step 303 is described in detail below.

[0109] In some embodiments, next, the prediction decision feature corresponding to each consecutive moment is is fed into a policy head shared across all time steps to predict the corresponding control actions Note that for the initial step, only the measured features are used to compute the initial attention map, and the aggregated image features are combined with the measured features to form the initial feature vector To ensure the characteristics Indeed, it describes the state of that step and contains important information for controlling the prediction. A feature loss is added at each step so that Close to user controlled features.

[0110] The above-mentioned strategy head can be a prediction strategy model obtained after data training such as reinforcement learning and deep learning.

[0111] Based on this, in each prediction decision feature Each corresponds to a control action , thereby generating predicted actions at multiple consecutive moments based on the predicted decision features at multiple consecutive moments , and then based on these predicted actions to get the second control decision .

[0112] Through the above steps 301 to 303, and steps 401 to 403, the path prediction features and navigation information are processed with attention, so as to fuse the static information and the dynamic information, and further utilize the first obtained predicted path as guidance information to capture the dependency between consecutive time steps, and utilize the future position information obtained from the predicted path to optimize the control prediction, so that multiple actions in the future can be predicted more accurately, so that the output second control decision can be planned in advance to avoid short-sighted decisions, further improving the control safety of the vehicle in the face of more complex road environments during turning, while also ensuring the control reliability of following the user's original control strategy corresponding to the navigation information.

[0113] In some embodiments, after obtaining the first control decision and the second control decision Afterwards, the first control decision and the second control decision need to be integrated to obtain a suitable preliminary control decision, so as to improve the operational control reliability of the vehicle. The following will further describe how to generate a suitable preliminary control decision.

[0114] Reference Figure 5 , obtaining a preliminary control decision based on the combination of the first control decision and the second control decision, including the following steps 501 to 503.

[0115] Step 501: Obtain a first control factor, and obtain a second control factor based on one minus the first control factor.

[0116] Step 502: When the image information indicates that the vehicle is in a straight-moving state, a preliminary control decision is obtained based on the product of the first control factor and the second control decision, plus the product of the second control factor and the first control decision.

[0117] Step 503: When the image information indicates that the vehicle is in a turning state, a preliminary control decision is obtained based on the product of the first control factor and the first control decision, plus the product of the second control factor and the second control decision.

[0118] Steps 501 to 503 are described in detail below.

[0119] In some embodiments, after obtaining the first control decision and the second control decision After that, get the preset first control factor , where the first control factor The value range of is (0, 0.5). Then the second control factor is obtained based on one minus the first control factor. .

[0120] Next, when the image information at the current moment indicates that the vehicle is in a straight-moving state, the control is relatively simple at this time, and the road is relatively less complicated. At this time, the first control decision will be given priority. Therefore, in this case, based on the first control factor and the second control decision The product of and the first control decision The product of As shown in the following formula (3).

[0121] (3)

[0122] When the image information at the current moment indicates that the vehicle is in a turning state, the road is relatively complex, so more careful control is required. At this time, the second control decision will be given priority. Therefore, in this case, based on the first control factor and the first control decision The product of and the second control decision The product of As shown in the following formula (4).

[0123] (4)

[0124] Reference Figure 6 , is a schematic diagram of a flow chart of generating a preliminary control decision provided by an embodiment of the present application. Figure 6As shown in , during the operation control of the vehicle, after acquiring the RGB image (i.e., image information) and navigation information (including speed, high-level control commands, and navigation target points) in real time, the image information is used to predict the trajectory branch to obtain the predicted trajectory, and the PID control is used to generate the first control decision corresponding to the predicted trajectory, and the navigation information is used, and the multi-step control branch is performed under the guidance of the predicted trajectory, so as to generate a reliable second control decision, and then based on the first control decision and the second control decision, a preliminary control decision is generated that meets both the user's control needs and more complex road conditions.

[0125] Through the above steps 501 to 503, the first control decision obtained by the trajectory branch and the second control decision obtained by the multi-step control branch are combined to significantly improve the performance of the autonomous driving system through complementary advantages. The trajectory branch provides long-term path planning information to help the model better understand future driving goals, while the multi-step control branch enhances the model's adaptability to dynamic environments and time consistency by predicting control actions for multiple time steps in the future. The two are closely combined through the trajectory-guided attention mechanism, which enables the control branch to dynamically adjust the focus area according to the trajectory information, thereby generating more accurate control signals. This combination not only reduces the sudden changes and discontinuities in vehicle actions, but also significantly reduces the incidence of collisions and violations, ultimately achieving safer and more stable autonomous driving performance.

[0126] Step 104: Input the image information into the collision correction model for data processing to obtain a safety control decision, and perform integration processing based on the safety control decision and the preliminary control decision to obtain a target control decision.

[0127] Step 104 is described in detail below.

[0128] In some embodiments, in order to further improve the control safety of the vehicle in the face of emergencies in complex road environments, after obtaining the image information, the image information is input into the collision correction model generated in advance according to the reinforcement learning algorithm for data processing to obtain a safety control decision for avoiding collisions and other safety issues based on the vehicle's operation at the current moment. , in order to facilitate subsequent safety control decisions and initial control decisions Through integration processing, a target control decision is obtained that can meet the user's control needs and improve the safety of operation control. .

[0129] The following first describes how to construct the collision correction model.

[0130] Reference Figure 7 The generation process of the collision correction model includes the following steps 701 to 705.

[0131] Step 701: Acquire multiple sample image information, and generate a state space based on the sample image information.

[0132] Step 702: Generate an action space based on the control parameters of the vehicle.

[0133] Step 703: Obtain an anti-collision objective function, and generate a reward model based on the anti-collision objective function.

[0134] Steps 701 to 703 are described in detail below.

[0135] The embodiment of the present application uses the auxiliary branch of the reinforcement learning model to optimize the action output of the original end-to-end method to solve the generalization performance of the vehicle in complex road environment scenes. Specifically, the PPO algorithm PPO in reinforcement learning uses the sample image information containing multiple continuous image frames as the state space, and then uses the vehicle's control parameters (including speed changes, steering changes) as the action space.

[0136] An anti-collision objective function matching the operation of the vehicle in a complex road environment is further utilized as a reward model, and the generation of the anti-collision objective function will be further described below.

[0137] Reference Figure 8 , obtaining the anti-collision objective function, including the following steps 801 to 804.

[0138] Step 801: Generate a centralization factor function based on the degree of deviation of the driving center of the vehicle during road driving.

[0139] Step 802: Generate an angle factor function based on the driving angle of the vehicle during road driving.

[0140] Step 803: Generate a target factor function based on the driving distance between the vehicle and the target driving position during the road driving process.

[0141] Step 804: Accumulate and multiply the centering factor function, the angle factor function, and the target factor function to generate an anti-collision target function.

[0142] Steps 801 to 804 are described in detail below.

[0143] In some embodiments, corresponding to the anti-collision objective function, the main factors considered include: (1) Collision penalty: If the autonomous driving vehicle collides with other vehicles or pedestrians, it will be penalized with -10, and this round of training iteration will end and enter the next round. (2) Distance to the destination: If the autonomous driving vehicle is always moving forward and is closer to the destination, it will be rewarded with +0.1, and if the autonomous driving vehicle is away from the destination, it will be penalized with -0.1. Whether it is far away from or close to the destination is determined by comparing the distance from the previous moment and the current moment to the destination. If it increases, it is far away, otherwise it is close. (3) Abnormal freezing: If the autonomous driving vehicle stops abnormally for a long time, it will be penalized with -10, and this round of training iteration will end and enter the next round. (4) Reaching the end point: If the autonomous driving vehicle successfully reaches the end point, it will be rewarded with +10, and this round of training iteration will end and enter the next round.

[0144] Based on the above ideas, a centralization factor function is generated based on the degree of deviation of the vehicle's driving center during road driving. , that is, when (distance to center) less than (maximum distance to the center), the centralization factor It decreases linearly from 1 to 0. Otherwise, the centralization factor remains at 1. Therefore, the centralization factor function is as shown in the following formula (5).

[0145] (5)

[0146] At the same time, based on the driving angle of the vehicle during road driving, an angle factor function is generated When When (angle) is less than or equal to ±40°, the angle factor It decreases linearly from 1 to 0. Otherwise, the angle factor remains at 1. Therefore, the angle factor function is as shown in the following formula (6).

[0147] (6)

[0148] In addition, based on the distance between the vehicle and the target driving position during the road driving process, the target factor function is generated That is, the target factor It is the distance The reciprocal of , but not exceeding 1. Therefore, the objective factor function is shown in the following formula (7).

[0149] (7)

[0150] Then, the centering factor function, the angle factor function and the target factor function are cumulatively multiplied to generate the anti-collision target function as shown in the following formula (8).

[0151] (8)

[0152] Through the above steps 801 to 804, by comprehensively quantifying the degree of centralization, angle deviation and target distance of vehicle driving, and generating a unified anti-collision objective function, the safety, stability and path planning accuracy of the autonomous driving system are significantly improved. The centralization factor function ensures that the vehicle drives smoothly near the center of the lane and reduces the risk of deviation; the angle factor function helps the vehicle maintain the correct driving direction and avoid loss of control; the target factor function guides the vehicle accurately to the target position and optimizes the path execution; thereby generating an anti-collision objective function that matches the vehicle's operation in a complex road environment, so that when the anti-collision objective function is used to train the initial collision correction model in the future, the safety and stability of the output decision can be improved, providing comprehensive safety protection for autonomous driving in complex road environments.

[0153] Step 704: Obtain initial action network parameters of the action network and initial evaluation network parameters of the evaluation network.

[0154] Step 705: Generate an initial collision correction model based on the state space, action space, reward model, initial action network parameters and initial evaluation network parameters, perform multiple rounds of training on the initial collision correction model, and obtain a collision correction model based on the trained initial collision correction model.

[0155] Steps 704 to 705 are described in detail below.

[0156] Reference Fig. 9 , is a schematic diagram of a training process of a deep reinforcement learning model provided in an embodiment of the present application. Fig. 9 As shown in , firstly, the policy network of the agent is designed; then the agent begins to interact with the environment through the actions output by the policy network, generating immediate rewards and observations, where the observations are generally consistent with the state; secondly, the trajectory is defined as the set of the current state, current action, current reward, and next state (observation), and the trajectories are continuously collected through the interaction between the agent and the environment; finally, when the number of trajectories reaches a certain value, the policy network (composed of a deep learning neural network) is updated, and the training is repeated in sequence until the cumulative reward of each round of training iteration is maximized and almost converged.

[0157] In the PPO algorithm used in the embodiment of the present application, the policy network uses two fully connected neural networks, the action (Actor) network and the evaluation (Critic) network, to make decisions. Therefore, it is also necessary to generate the initial action network parameters of the action network and the initial evaluation network parameters of the evaluation network.

[0158] Then, an initial collision correction model is generated based on the state space, action space, reward model, initial action network parameters and initial evaluation network parameters that match the vehicle operation control process. The initial collision correction model is then trained for multiple rounds, and the actual collision correction model is obtained based on the action (Actor) network in the trained initial collision correction model.

[0159] Among them, the training samples, feedback, and result data obtained in each training process are all saved in the experience pool, and when storing obs, the action of the old_policy network and the logarithmic probability of the action logprob are calculated.

[0160] The process of the PPO algorithm includes the following steps.

[0161] (1) Sampling: Using the current strategy Interact with the environment and collect status ,action ,award .

[0162] (2) Calculating the advantage function: Evaluating an action In Status The performance relative to the average (the advantage function ),use Guide strategy improvements.

[0163] (3) Calculate the probability ratio : Compare the new strategy and the old strategy for action The probability of selection.

[0164] (4) Strategy update: If the update is too large (outside the clipping range arrive ), will be punished. It is necessary to ensure that the update range is moderate, neither too conservative nor too radical.

[0165] (5) Value function update: Use the following loss function to optimize the value function .

[0166] (6) Repeat the above steps: Through multiple rounds of iterations, the strategy is gradually optimized until convergence.

[0167] Among them, the training of the initial collision correction model based on the PPO algorithm includes the following steps.

[0168] (1) Discounted Reward: For each time step t, if t is a terminal state, then .otherwise, ,in is the discount factor, is the reward at time step t to compute the cumulative reward at each time step, taking into account the discount of future rewards.

[0169] (2) Evaluating Old Actions and Values: Use the policy network to evaluate the old states old_states: and old actions old_actions to obtain log probability logprobs, value values, and distribution entropy dist_entropy to evaluate the value of old actions and states, and obtain log probability, value estimate, and distribution entropy.

[0170] (3) Matching Tensor Dimensions: Compress the value tensor values ​​to match the dimensions of the reward tensor returns to adjust the dimensions of the value tensor to match the reward tensor.

[0171] (4) Finding the Ratio: ratios = exp(logprobs - old_logprobs), where old_logprobs is the old log probability detached from the computational graph and logprobs is the log probability of the action under the current strategy, to calculate the action probability ratio between the new and old strategies.

[0172] (5) Calculate the surrogate loss: advantages = returns - values, surrl = ratios × advantages, surr2 = max (ratios × advantages, clamp (ratios, 1-clip, 1+clip)), where clip is a hyperparameter used to limit the value of the ratio, advantages is the advantage function, indicating the pros and cons of the current action relative to the average performance, surr1 and surr2 represent the surrogate loss after unclipping and clipping, respectively. The surrogate loss is calculated by the clipping ratio to ensure the stability of the policy update.

[0173] (6) Final Clipped Objective PPO Loss: loss = -min(surrl, surr2) + 0.5 × MseLoss (values, returns) - loss_coeff_entropy × dist_entropy, where MseLoss is the mean square error loss function, loss_coeff_entropy is the coefficient of the degree loss, and dist_entropy is the entropy of the policy distribution, which is used to encourage exploration. The final PPO loss is calculated by combining the substitution loss, value loss, and entropy loss for optimizing the policy network.

[0174] In particular, in the present invention, the action space of reinforcement learning only considers the change in speed, which is a continuous action distribution, because complex traffic scenes are usually sudden scenes that require a fast speed response rather than a detour. In summary, by using the reinforcement learning algorithm, the intelligent agent is constantly interacting with complex autonomous driving scenes, allowing the intelligent agent to continuously learn from the scene, thereby achieving better generalization performance.

[0175] Through the above steps 701 to 705, the adaptive learning characteristics of the reinforcement learning algorithm are utilized, combined with the detailed parameters in the vehicle operation control, and the anti-collision objective function that matches the vehicle operation in a complex road environment is used as the learning goal, so as to train a collision correction model that can effectively improve the operation safety of the vehicle in actual operation control, thereby effectively improving the safety and reliability of vehicle operation control in practical applications.

[0176] In some embodiments, after obtaining a preliminary control decision and safety control decisions , using the preset control combination factors , will initially control the decision and safety control decisions Combine to get the target control decision .

[0177] Step 105: Perform driving control on the vehicle based on the target control decision.

[0178] The following is a detailed description of step 105.

[0179] Reference Fig.10 , is a schematic flow chart of a vehicle driving control method provided by an embodiment of the present application. Fig.10As shown in , after acquiring the image information and navigation information during the operation of the vehicle in real time, the image information and navigation information are used to perform trajectory branching and multi-step control branching to generate a preliminary control decision that accurately corresponds to the original control information, and the collision correction model obtained by reinforcement learning is used to generate a safety control decision to avoid the vehicle from being hit, thereby combining the preliminary control decision with the safety control decision to obtain the target control decision, and then the control module on the vehicle is used to perform safe and reliable operation control of the vehicle's speed and steering.

[0180] The vehicle driving control method provided in the embodiment of the present application is a modular method and has the following advantages: 1) Reinforcement learning allows the agent to learn autonomously through interaction with the environment, rather than relying on pre-defined rules or direct mapping of multimodal data, which enables the agent to adapt to unknown or changing environments. 2) The goal of reinforcement learning is to find a strategy that maximizes the cumulative reward, which enables the agent to continuously optimize its decision-making process, while rule-based methods may not be flexible enough in complex or dynamic environments. 3) Strong generalization ability: By learning to interact with the environment, reinforcement learning agents can generalize to new or unseen situations, while multimodal end-to-end methods may be limited by the diversity of training data. 4) Interactive learning: Reinforcement learning is particularly suitable for tasks that require interactive learning, such as complex traffic scenarios in autonomous driving, while multimodal end-to-end methods may focus more on directly learning feature representations from data.

[0181] The vehicle driving control method, device, electronic device and storage medium proposed in the embodiments of the present application include: first, obtaining image information and navigation information of the vehicle at the current moment in the driving process; second, extracting features of the current image frame to obtain path prediction features, inputting the path prediction features and navigation information into a path prediction network to obtain a predicted trajectory, the predicted trajectory includes predicted path points at multiple consecutive moments, calculating the path vector between every two adjacent predicted path points, and obtaining a first control decision based on the path vector and PID control; then, performing attention processing on the path prediction features and navigation information to obtain initial fusion features, and at each consecutive moment, performing attention processing on the path prediction features and navigation information to obtain initial fusion features based on the probability distribution of the initial fusion features and the control of the path vector. The corresponding predicted path points are superimposed to obtain updated fusion features, and splicing processing is performed based on the updated fusion features and navigation information to obtain spliced ​​fusion features. Attention processing is performed on the spliced ​​fusion features to obtain the prediction decision features corresponding to the continuous moments. Based on multiple prediction decision features, a predicted action for each continuous moment is generated, and a second control decision is obtained based on all predicted actions, and a first control factor is obtained, and a second control factor is obtained based on one minus the first control factor, and the first control factor is less than one-half. When the image information represents that the vehicle is in a straight state, a preliminary control decision is obtained based on the product of the first control factor and the second control decision, plus the product of the second control factor and the first control decision. When the image information represents that the vehicle is in a straight state, a preliminary control decision is obtained based on the product of the first control factor and the second control decision, and the product of the second control factor and the first control decision. When the information characterizes that the vehicle is in a turning state, a preliminary control decision is obtained based on the product of the first control factor and the first control decision, plus the product of the second control factor and the second control decision; next, a plurality of sample image information is obtained, and a state space is generated based on the sample image information, each sample image information includes a plurality of continuous image frames, an action space is generated based on the control parameters of the vehicle, a centralization factor function is generated based on the degree of deviation of the driving center of the vehicle during road driving, an angle factor function is generated based on the driving angle of the vehicle during road driving, a target factor function is generated based on the driving distance of the vehicle from the target driving position during road driving, and the centralization factor function and the angle factor function are cumulatively multiplied. and the target factor function, generate an anti-collision target function, and generate a reward model based on the anti-collision target function, obtain initial action network parameters of the action network and initial evaluation network parameters of the evaluation network, generate an initial collision correction model based on the state space, action space, reward model, initial action network parameters and initial evaluation network parameters, perform multiple rounds of training on the initial collision correction model, and obtain a collision correction model based on the trained initial collision correction model, and input image information into the collision correction model for data processing to obtain a safety control decision, and integrate the safety control decision and the preliminary control decision to obtain a target control decision; finally, drive and control the vehicle based on the target control decision.

[0182] The embodiment of the present application utilizes the predicted trajectory obtained by performing trajectory prediction based on the image information around the vehicle to generate a corresponding first control decision, and further utilizes the predicted trajectory for guidance to generate a predicted multi-step control decision that matches the control navigation information, thereby generating a preliminary control decision that accurately corresponds to the original control information; in addition, a collision correction model that can be trained and generated in advance based on reinforcement learning for complex road environments is used to generate a safety control decision to prevent the vehicle from being hit, so as to utilize the high learning adaptability of reinforcement learning to avoid safety accidents as much as possible when encountering sudden risks, thereby improving the control reliability of the predetermined route and the operational control safety when the vehicle is controlled using the preliminary control decision and the safety control decision; in addition, the path prediction feature obtained by extracting features from the current image frame is used to characterize the current image information feature data during the vehicle operation process, and then combined with the navigation information to perform accurate path prediction to obtain a predicted path that matches the current operation of the vehicle and also matches the user's needs, thereby A PID controller is combined to generate a first control decision containing longitudinal and lateral control data that matches the predicted path, thereby improving the reliability of the vehicle driving control method; and, by processing the path prediction features and navigation information with attention, static information and dynamic information are integrated, and the first predicted path is further used as guidance information to capture the dependency between consecutive time steps, and the future position information obtained from the predicted path is used to optimize the control prediction, so that multiple future actions can be predicted more accurately, so that the output second control decision can be planned in advance to avoid short-sighted decisions, further improving the control safety of the vehicle in the face of more complex road environments during turning, while also ensuring the control reliability of following the user's original control strategy corresponding to the navigation information; and, by comprehensively quantifying the degree of centralization, angle deviation and target distance of vehicle driving, and generating a unified anti-collision objective function, the safety, stability and path planning accuracy of the autonomous driving system are significantly improved. The centering factor function ensures that the vehicle travels smoothly near the center of the lane and reduces the risk of deviation; the angle factor function helps the vehicle maintain the correct driving direction and avoids loss of control; the target factor function guides the vehicle accurately to the target position and optimizes path execution; thereby generating an anti-collision target function that matches the vehicle's operation in a complex road environment, so that when the anti-collision target function is subsequently used to train the initial collision correction model, the safety and stability of the output decision can be improved, providing comprehensive safety protection for autonomous driving in complex road environments; in addition, the adaptive learning characteristics of the reinforcement learning algorithm are utilized, combined with the detailed parameters in the vehicle operation control, and the anti-collision target function that matches the vehicle's operation in a complex road environment is used as the learning target, so as to train a collision correction model that can effectively improve the vehicle's operating safety in actual operation control, thereby effectively improving the safety and reliability of vehicle operation control in practical applications.

[0183] The present application also provides a vehicle driving control device, which can implement the above vehicle driving control method. Fig.11 , the vehicle driving control device 1100 comprises:

[0184] The information acquisition module 1110 is used to acquire the image information and navigation information of the vehicle at the current moment during the driving process;

[0185] A first control decision generating module 1120, configured to perform trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtain a first control decision based on the predicted trajectory;

[0186] A second control decision generating module 1130 is used to perform multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision, and to obtain a preliminary control decision based on the combination of the first control decision and the second control decision;

[0187] The target control decision generating module 1140 is used to input the image information into the collision correction model for data processing to obtain a safety control decision, and to integrate the safety control decision and the preliminary control decision to obtain a target control decision;

[0188] The vehicle control module 1150 is used to control the driving of the vehicle based on the target control decision.

[0189] In some embodiments, the first control decision generating module 1120 is further configured to:

[0190] Extract features from the current image frame to obtain path prediction features;

[0191] Inputting the path prediction features and navigation information into the path prediction network to obtain a predicted trajectory, the predicted trajectory including predicted path points at multiple consecutive moments;

[0192] A path vector between every two adjacent predicted path points is calculated, and a first control decision is obtained based on the path vector and PID control.

[0193] In some embodiments, the second control decision generating module 1130 is further configured to:

[0194] Perform attention processing on the path prediction features and navigation information to obtain the initial fusion features;

[0195] Based on the initial fusion features and multiple predicted path points, the prediction decision features of multiple consecutive moments are generated;

[0196] Based on multiple prediction decision features, a prediction action at each continuous moment is generated, and a second control decision is obtained based on all the prediction actions.

[0197] In some embodiments, the second control decision generating module 1130 is further configured to:

[0198] At each continuous moment, the updated fusion feature is obtained based on the probability distribution of the initial fusion feature and the superposition processing of the corresponding predicted path points;

[0199] Perform splicing processing based on the updated fusion feature and the navigation information to obtain the spliced ​​fusion feature;

[0200] Attention processing is performed on the concatenated fusion features to obtain the prediction decision features corresponding to consecutive moments.

[0201] In some embodiments, the second control decision generating module 1130 is further configured to:

[0202] Obtaining a first control factor and obtaining a second control factor based on one minus the first control factor, the first control factor being less than one-half;

[0203] When the image information indicates that the vehicle is in a straight-moving state, a preliminary control decision is obtained based on the product of the first control factor and the second control decision, plus the product of the second control factor and the first control decision;

[0204] When the image information represents that the vehicle is in a turning state, a preliminary control decision is obtained based on the product of the first control factor and the first control decision, plus the product of the second control factor and the second control decision.

[0205] In some embodiments, the target control decision generation module 1140 is further configured to:

[0206] Acquire multiple sample image information and generate a state space based on the sample image information, each sample image information including multiple continuous image frames;

[0207] Generate an action space based on the vehicle's control parameters;

[0208] Obtaining an anti-collision objective function and generating a reward model based on the anti-collision objective function;

[0209] Obtaining initial action network parameters of the action network and initial evaluation network parameters of the evaluation network;

[0210] An initial collision correction model is generated based on the state space, the action space, the reward model, the initial action network parameters and the initial evaluation network parameters, and the initial collision correction model is trained for multiple rounds, and a collision correction model is obtained based on the trained initial collision correction model.

[0211] In some embodiments, the target control decision generation module 1140 is further configured to:

[0212] Generate a centralization factor function based on the degree of deviation of the driving center of the vehicle during road driving;

[0213] Generate an angle factor function based on the driving angle of the vehicle during the road driving process;

[0214] Generate a target factor function based on the driving distance of the vehicle from the target driving position during the road driving process;

[0215] The centering factor function, the angle factor function and the target factor function are accumulated to generate the anti-collision target function.

[0216] In the above embodiments, the description of each embodiment has its own emphasis. For the part that is not described in detail in a certain embodiment, the specific implementation method of the vehicle driving control device is basically the same as the specific implementation method of the above vehicle driving control method, and will not be repeated here.

[0217] In an embodiment of the present application, a vehicle driving control device uses a predicted trajectory obtained by performing trajectory prediction based on image information around the vehicle to generate a corresponding first control decision, and further uses the predicted trajectory for guidance to generate a predicted multi-step control decision that matches the control navigation information, thereby generating a preliminary control decision that accurately corresponds to the original control information; in addition, a collision correction model that can be trained and generated in advance based on reinforcement learning for complex road environments is used to generate a safety control decision to prevent the vehicle from being hit, so as to utilize the high learning adaptability of reinforcement learning to avoid safety accidents as much as possible when encountering sudden risks, thereby improving the control reliability of the predetermined route and the safety of operation control when the vehicle is controlled using preliminary control decisions and safety control decisions; in addition, the path prediction features obtained by feature extraction from the current image frame are used to characterize the current image information feature data during the vehicle operation process, and then combined with the navigation information to perform accurate path prediction to obtain a prediction that matches the current operation of the vehicle and also matches the user's needs. path, thereby combining with the PID controller to generate a first control decision containing longitudinal and lateral control data matching the predicted path, thereby improving the reliability of the vehicle driving control method; and, by processing the path prediction features and navigation information with attention, the static information and the dynamic information are integrated, and the first predicted path is further used as guidance information to capture the dependency between consecutive time steps, and the future position information obtained from the predicted path is used to optimize the control prediction, so that multiple actions in the future can be predicted more accurately, so that the output second control decision can be planned in advance to avoid short-sighted decisions, further improving the control safety of the vehicle in the face of more complex road environments during turning, and at the same time ensuring the control reliability of the original control strategy corresponding to the navigation information of the user; and, by comprehensively quantifying the degree of centralization, angle deviation and target distance of vehicle driving, and generating a unified anti-collision objective function, the safety, stability and path planning accuracy of the autonomous driving system are significantly improved.The centering factor function ensures that the vehicle travels smoothly near the center of the lane and reduces the risk of deviation; the angle factor function helps the vehicle maintain the correct driving direction and avoids loss of control; the target factor function guides the vehicle accurately to the target position and optimizes path execution; thereby generating an anti-collision target function that matches the vehicle's operation in a complex road environment, so that when the anti-collision target function is subsequently used to train the initial collision correction model, the safety and stability of the output decision can be improved, providing comprehensive safety protection for autonomous driving in complex road environments; in addition, the adaptive learning characteristics of the reinforcement learning algorithm are utilized, combined with the detailed parameters in the vehicle operation control, and the anti-collision target function that matches the vehicle's operation in a complex road environment is used as the learning target, so as to train a collision correction model that can effectively improve the vehicle's operating safety in actual operation control, thereby effectively improving the safety and reliability of vehicle operation control in practical applications.

[0218] The present application also provides an electronic device, including:

[0219] at least one memory;

[0220] at least one processor;

[0221] at least one program;

[0222] The program is stored in the memory, and the processor executes the at least one program to implement the vehicle driving control method implemented in the present application. The electronic device can be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a vehicle-mounted computer, etc.

[0223] See also Fig.12 , Fig.12 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0224] The processor 1201 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0225] The memory 1202 can be implemented in the form of ROM (Read Only Memory), static storage device, dynamic storage device or RAM (Random Access Memory). The memory 1202 can store operating systems and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program codes are stored in the memory 1202, and the processor 1201 is used to call and execute the vehicle driving control method of the embodiment of this application;

[0226] Input / output interface 1203, used to implement information input and output;

[0227] The communication interface 1204 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.);

[0228] A bus 1205 that transmits information between various components of the device (e.g., the processor 1201, the memory 1202, the input / output interface 1203, and the communication interface 1204);

[0229] The processor 1201 , the memory 1202 , the input / output interface 1203 and the communication interface 1204 are connected to each other in communication within the device via the bus 1205 .

[0230] An embodiment of the present application also provides a storage medium, which is a computer-readable storage medium. The storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned vehicle driving control method is implemented.

[0231] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0232] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0233] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0234] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0235] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0236] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0237] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0238] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0239] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0240] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0241] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0242] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A vehicle driving control method, characterized in that: The method comprises: Obtain the image information and navigation information of the vehicle at the current moment during driving; Performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory; Performing a multi-step control process based on the navigation information and the predicted trajectory to obtain a second control decision, and obtaining a preliminary control decision based on a combination of the first control decision and the second control decision; Inputting the image information into a collision correction model for data processing to obtain a safety control decision, and integrating the safety control decision and the preliminary control decision to obtain a target control decision; Performing driving control on the vehicle based on the target control decision; The image information includes a current image frame, and performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory, includes: Extracting features from the current image frame to obtain path prediction features; Inputting the path prediction feature and the navigation information into a path prediction network to obtain the predicted trajectory, wherein the predicted trajectory includes predicted path points at multiple consecutive moments; Calculating a path vector between every two adjacent predicted path points, and obtaining the first control decision based on the path vector and PID control; The performing multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision includes: Performing attention processing on the path prediction feature and the navigation information to obtain an initial fusion feature; Based on the initial fusion feature and the plurality of predicted path points, generating prediction decision features at a plurality of consecutive moments; Based on the plurality of prediction decision features, a prediction action at each continuous moment is generated, and the second control decision is obtained based on all the prediction actions.

2. The vehicle driving control method according to claim 1, characterized in that: The generating a plurality of prediction decision features at consecutive moments based on the initial fusion features and the plurality of predicted path points comprises: At each continuous moment, based on the probability distribution of the initial fusion feature and the superposition processing of the corresponding predicted path point, an updated fusion feature is obtained; Performing splicing processing based on the updated fusion feature and the navigation information to obtain a spliced ​​fusion feature; Attention processing is performed on the spliced ​​fusion features to obtain the prediction decision features corresponding to the continuous moments.

3. The vehicle driving control method according to claim 1, characterized in that: The obtaining a preliminary control decision based on the combination of the first control decision and the second control decision comprises: Obtaining a first control factor, and obtaining a second control factor based on one minus the first control factor, the first control factor being less than one-half; When the image information indicates that the vehicle is in a straight-moving state, the preliminary control decision is obtained based on the product of the first control factor and the second control decision, plus the product of the second control factor and the first control decision; When the image information represents that the vehicle is in a turning state, the preliminary control decision is obtained based on the product of the first control factor and the first control decision plus the product of the second control factor and the second control decision.

4. The vehicle driving control method according to claim 1, characterized in that: The steps of constructing the collision correction model include: Acquire a plurality of sample image information, and generate a state space based on the sample image information, each of the sample image information including a plurality of continuous image frames; generating an action space based on control parameters of the vehicle; Obtaining an anti-collision objective function, and generating a reward model based on the anti-collision objective function; Obtaining initial action network parameters of the action network and initial evaluation network parameters of the evaluation network; An initial collision correction model is generated based on the state space, the action space, the reward model, the initial action network parameters and the initial evaluation network parameters, and the initial collision correction model is trained for multiple rounds, and the collision correction model is obtained based on the trained initial collision correction model.

5. The vehicle driving control method according to claim 4, characterized in that: The obtaining of the anti-collision objective function comprises: generating a centralization factor function based on the degree of deviation of the driving center of the vehicle during road driving; generating an angle factor function based on a driving angle of the vehicle during road driving; generating a target factor function based on a travel distance between the vehicle and a target travel position during road travel; The centering factor function, the angle factor function and the target factor function are cumulatively multiplied to generate the anti-collision target function.

6. A vehicle driving control device, characterized in that: The device comprises: An information acquisition module is used to acquire image information and navigation information of the vehicle at the current moment during driving; A first control decision generating module, configured to perform trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtain a first control decision based on the predicted trajectory; a second control decision generating module, configured to perform a multi-step control process based on the navigation information and the predicted trajectory to obtain a second control decision, and to obtain a preliminary control decision based on a combination of the first control decision and the second control decision; a target control decision generating module, configured to input the image information into a collision correction model for data processing to obtain a safety control decision, and to integrate the safety control decision and the preliminary control decision to obtain a target control decision; A vehicle control module, configured to perform driving control on the vehicle based on the target control decision; The image information includes a current image frame, and performing trajectory prediction processing based on the image information to obtain a predicted trajectory, and obtaining a first control decision based on the predicted trajectory, includes: Extracting features from the current image frame to obtain path prediction features; Inputting the path prediction feature and the navigation information into a path prediction network to obtain the predicted trajectory, wherein the predicted trajectory includes predicted path points at multiple consecutive moments; Calculating a path vector between every two adjacent predicted path points, and obtaining the first control decision based on the path vector and PID control; The performing multi-step control processing based on the navigation information and the predicted trajectory to obtain a second control decision includes: Performing attention processing on the path prediction feature and the navigation information to obtain an initial fusion feature; Based on the initial fusion feature and the plurality of predicted path points, generating prediction decision features at a plurality of consecutive moments; Based on the plurality of prediction decision features, a prediction action at each continuous moment is generated, and the second control decision is obtained based on all the prediction actions.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the vehicle driving control method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the vehicle driving control method according to any one of claims 1 to 5 is implemented.

Images