Automatic driving decision and control method and system based on environmental perception

Abstract

The invention relates to the technical field of automatic driving, and provides an automatic driving decision and control method and system based on environmental perception, and the method comprises the steps: employing a preset encoder and decoder architecture, and obtaining the driving perception information of a target object in an automatic driving process through a preset feature fusion optimization strategy; verifying the accuracy of the driving perception result to obtain a target encoder and decoder architecture; constructing an environment state based on driving perception information output by the target encoder and decoder architecture, and obtaining an initial execution action of the vehicle in an automatic driving process through an initial execution strategy; training the initial execution strategy to obtain an optimized decision process model; and constructing a mapping relation between an environment state and an execution action, and stably converging the decision process model to an optimal strategy. Compared with the prior art, the method has the advantages of high decision reliability, high control fineness and the like, and ensures safe and stable operation of automatic driving.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to an autonomous driving decision-making and control method and system based on environmental perception. Background Art

[0002] Currently, most research on object detection algorithms for autonomous driving focuses on path planning and collision avoidance for normal driving. However, traffic violations beyond collisions are also a key factor in autonomous driving, and these violations are more likely to lead to accidents. Existing designs rarely address other violations, such as off-road driving and failure to obey road signs. Advanced driving models often lack dynamic 3D object detection during the imitation learning phase, resulting in safety concerns in autonomous driving decision-making.

[0003] Based on the above content, this application provides an autonomous driving decision-making and control method based on environmental perception, which accurately extracts semantic, depth and motion attributes through multimodal feature fusion, and optimizes decision-making strategies by combining reinforcement learning and dual network mechanisms, thereby reducing collision risks and decision-making deviations, and improving the safety, decision-making reliability and control precision of autonomous driving in complex scenarios, ensuring safe driving. Summary of the Invention

[0004] The purpose of this invention is to provide an autonomous driving decision-making and control method and system based on environmental perception, which has the advantages of high decision-making reliability and strong control precision, thereby improving the safety of autonomous navigation driving. The specific solution is as follows:

[0005] An autonomous driving decision-making and control method based on environmental perception, the method comprising the following steps:

[0006] S1: Using a preset encoder and decoder architecture and a preset feature fusion optimization strategy, obtain driving perception information of a target object during autonomous driving; wherein the driving perception information includes at least semantic attribute features, depth features, and motion attributes;

[0007] S2: Verify the accuracy of the driving perception results and optimize the parameters of the preset encoder and decoder architecture based on the verification results to obtain the target encoder and decoder architecture;

[0008] S3: Constructs the environment state based on the driving perception information output by the target encoder and decoder architecture, and obtains the initial execution action of the vehicle during the autonomous driving process through the initial execution strategy;

[0009] S4: training the initial execution strategy and updating the decision-making process model weight parameters in the execution strategy to obtain an optimized decision-making process model;

[0010] S5: Based on the updated decision-making process model weight parameters, the mapping relationship between the environment state and the execution action is constructed through the preset network, and the decision-making process model is stably converged to the optimal strategy.

[0011] Optionally, the preset encoder and decoder architecture in step S1 adopts a U-Net3 architecture, which includes: an encoder, a dual decoder and a speed prediction branch;

[0012] The encoder uses MobileNetV2 as a pre-trained encoder model to extract common features of the target image;

[0013] Dual decoders, including:

[0014] Semantic decoder, used to upsample and semantically reconstruct the fused features and output the semantic attribute features of the target object;

[0015] Depth decoder, used to perform geometric feature analysis on the fused features and output the depth features of the target object;

[0016] The velocity prediction branch is used to predict the motion properties of the target object based on the universal features output by the encoder and convert the sparse velocity measurements into pixel-level motion features; the motion properties include the target object's velocity, acceleration, movement trajectory, and relative speed to the vehicle.

[0017] Optionally, the preset feature fusion optimization strategy in S1 includes:

[0018] The universal features output by the encoder are input to the initial layers of the dual decoder and the speed prediction branch respectively; the dual decoder preliminarily extracts semantic attribute features and depth features based on the universal features; the speed prediction branch generates preliminary pixel-level motion features based on the universal features;

[0019] Combine the encoder's universal features with the preliminary pixel-level motion features generated by the speed prediction branch in the channel dimension to obtain fused features.

[0020] The fused features are input into the deep layer of the dual decoder and the deep layer of the speed prediction branch respectively, and finally the driving perception information including semantic attribute features, depth features and motion attributes is obtained;

[0021] The parameters of the preset encoder and decoder architecture are jointly optimized through multiple loss functions to ensure that the driving perception information obtained meets the preset accuracy standards.

[0022] Optional, S2, specifically including:

[0023] Collect real image information of the target object;

[0024] Compare and analyze the real image information with the driving perception result output in step S1, and calculate the error index;

[0025] If all error indicators are less than the preset threshold, the driving perception result is determined to meet the design requirements, and the preset architecture is the target encoder and decoder architecture;

[0026] If any error indicator is greater than or equal to the preset threshold, it indicates that the perception result does not meet the requirements. The parameters of the encoder, dual decoder and speed prediction branch are optimized based on error backpropagation, and the parameters of the preset encoder and decoder architecture are continued to be optimized until all error indicators are less than the preset threshold, and the target encoder and decoder architecture is obtained.

[0027] Optional, including:

[0028] The decision-making process of the vehicle's execution action during the autonomous driving process is achieved through a reinforcement learning algorithm, wherein a decision-making process model is used as the basic framework of reinforcement learning to output the execution action under different environmental conditions; the decision-making process model adopts a Markov decision process model;

[0029] The input data of the Markov decision process model is the environmental state data; the environmental state data is constructed based on the driving perception information output by the target encoder and decoder architecture;

[0030] Based on the above environmental status data, the initial execution action is output through the initial execution strategy;

[0031] After the vehicle performs the initial action, the experience tuple generated by the interaction is stored in the experience replay pool;

[0032] The experience tuple includes: the vehicle's current environment state s, the executed action a, the immediate reward r, and the vehicle's environment state s' at the next moment; wherein, the next environment state s' and the immediate reward r are generated after the executed action a interacts with the environment.

[0033] Optional, including:

[0034] S401: Randomly sample experience tuples (s, a, r, s') from the experience replay pool based on the state-action value function Q n (s,a) Calculate the expected cumulative reward Q for taking action a in the current environment state s;

[0035] S402: Update the Q value in real time through the Bellman equation;

[0036] S403: Use the DDPG algorithm to process the continuous state-action space of autonomous driving, randomly sample samples through the experience replay pool, and update the weight parameters of the decision process model.

[0037] Optionally, step S5 specifically includes:

[0038] S501: Based on the updated decision process model weight parameters, a mapping relationship between the environment state and the execution action is constructed through a deep deterministic policy network;

[0039] S502: Based on the reward obtained for each executed action, the parameters of the training network are updated in real time;

[0040] S503: A dual-network mechanism is used to control the convergence of the target network weights to the training network by updating the coefficients to ensure stable optimization of the strategy; the dual-network mechanism includes: a training network and a target network, both of which are sub-networks of the decision process model.

[0041] An autonomous driving decision-making and control system based on environmental perception, comprising:

[0042] a fusion perception module configured to employ a preset encoder and decoder architecture and a preset feature fusion optimization strategy to obtain driving perception information of a target object during autonomous driving; wherein the driving perception information includes at least semantic attribute features, depth features, and motion attributes;

[0043] a perception verification module configured to verify the accuracy of the driving perception results and optimize parameters of the preset encoder and decoder architectures based on the verification results to obtain target encoder and decoder architectures;

[0044] a decision execution module configured to construct an environment state based on the driving perception information output by the target encoder and decoder architecture, and obtain an initial execution action of the vehicle during the autonomous driving process through an initial execution strategy;

[0045] a decision building module configured to train the initial execution strategy and update the decision process model weight parameters in the execution strategy to obtain an optimized decision process model;

[0046] The decision optimization module is configured to build a mapping relationship between the environment state and the execution action based on the updated decision process model weight parameters through a preset network, and stably converge the decision process model to the optimal strategy.

[0047] An electronic device includes: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the method.

[0048] A computer-readable storage medium stores a computer program executable by an electronic device. When the computer program is run on the electronic device, the electronic device executes the steps of the method.

[0049] Through the above solution, the following beneficial technical effects are achieved:

[0050] The present application provides an autonomous driving decision-making and control method and system based on environmental perception. First, a preset encoder and decoder architecture is adopted, and a preset feature fusion optimization strategy is used to achieve high-precision prediction of driving perception information of a target object during driving. Further, in order to ensure the accuracy of the prediction, the driving perception results are verified, and the parameters of the preset encoder and decoder architecture are optimized according to the verification results to obtain a target encoder and decoder architecture. Furthermore, the driving perception information output by the target encoder and decoder architecture is optimized through a reinforcement learning algorithm to optimize the execution strategy to obtain the optimal execution action during the autonomous driving process. At the same time, by constructing a mapping relationship between environmental state and execution action, and combining a dual-network mechanism, problems such as target value fluctuation, high data correlation, and strategy oscillation in reinforcement learning are overcome, ultimately ensuring the stability of the decision-making process model output in autonomous driving and improving the safety of autonomous driving in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flowchart of an autonomous driving decision-making and control method based on environmental perception;

[0052] Figure 2 This is a structural block diagram of an autonomous driving decision and control system based on environmental perception. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of this application clearer, the following Figure 1-Figure 2 This application is further described in detail. Obviously, the embodiments described are only a part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0054] The terms used in the examples of this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a," "the," and "the" used in the examples of this application and the appended claims are also intended to include plural forms, and unless the context clearly indicates otherwise, "a plurality" generally includes at least two.

[0055] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0056] It should be understood that although the terms first, second, third, etc. may be used to describe in the embodiments of the present application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, without departing from the scope of the embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first.

[0057] As used herein, the words "if" and "if" may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0058] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or device. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or device comprising the element.

[0059] It should be noted in particular that any symbols and / or numbers in the specification that are not marked in the accompanying drawings are not drawing marks.

[0060] The optional embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0061] Figure 1 An autonomous driving decision-making and control method based on environmental perception is shown, the method comprising the following steps:

[0062] S1: Using a preset encoder and decoder architecture and a preset feature fusion optimization strategy, obtain driving perception information of the target object during autonomous driving; wherein the driving perception information includes at least: semantic attribute features (such as identifying pedestrians and vehicles), depth features (such as the distance of the target object), and motion attributes (such as the movement trajectory, speed, and acceleration of pedestrians);

[0063] S2: Verify the accuracy of the driving perception results and optimize the parameters of the preset encoder and decoder architecture (such as adjusting the convolutional layer weights and feature fusion coefficients) based on the verification results (such as when the error exceeds the limit) to obtain the target encoder and decoder architecture;

[0064] S3: Constructs the environment state based on the driving perception information output by the target encoder and decoder architecture, and obtains the initial execution action of the vehicle during the autonomous driving process through the optimized execution strategy;

[0065] S4: Based on the interaction data between the environment state and the initial execution action, the execution strategy is trained and the weight parameters of the decision process model in the execution strategy are updated to obtain an optimized decision process model;

[0066] S5: Based on the updated decision-making process model weight parameters, the mapping relationship between the environment state and the execution action is constructed through the preset network, and the decision-making process model is stably converged to the optimal strategy.

[0067] This application integrates multimodal driving perception information such as semantics, depth, and motion attributes, and realizes complementary enhancement of information of different dimensions through preset feature fusion strategies (such as using depth features to correct spatial ambiguity in semantic recognition, and using motion attributes to predict target behavior trends). After accuracy verification and parameter optimization, a robust target encoder and decoder architecture is formed, which solves the technical problems of one-sided information or insufficient perception accuracy of a single perception modality from the source, and provides a comprehensive and reliable environmental cognition basis for subsequent decision-making; furthermore, the environmental state constructed based on high-precision perception results is more in line with the real driving scene, making the initial action of the execution strategy more targeted; the decision model weights are continuously optimized through interactive data training, and the mapping construction between environmental state and execution action is combined to ensure that the strategy iteration goal is always: maximizing driving safety and efficiency; finally, the stable convergence mechanism is used to effectively avoid the strategy oscillation that is prone to occur in reinforcement learning in the continuous action space, so that the actions output by the decision process model are both accurately adapted to the real-time road conditions and maintain long-term stability.

[0068] In a specific embodiment, the preset encoder and decoder architecture in step S1 adopts a U-Net3 architecture, which includes: an encoder, a dual decoder and a speed prediction branch;

[0069] The encoder uses MobileNetV2 as a pre-trained encoder model to extract common features of the target image (such as target category, color, and texture);

[0070] Dual decoders, including:

[0071] Semantic decoder, used to upsample and semantically reconstruct the fused features and output the semantic attribute features of the target object;

[0072] Depth decoder, used to perform geometric feature analysis on the fused features and output the depth features of the target object;

[0073] The velocity prediction branch is used to predict the target object's motion properties based on the universal features output by the encoder and convert sparse velocity measurements (such as key target speeds obtained by radar or GPS) into pixel-level motion features. The motion properties include the target object's velocity, acceleration, movement trajectory, and relative speed to the vehicle.

[0074] Specifically, this embodiment uses a single encoder and multi-branch decoder architecture to achieve one-time extraction of common features and multi-task sharing, avoiding repeated calculations. At the same time, each decoder specifically processes different types of features (the semantic decoder focuses on category differentiation, the depth decoder focuses on spatial positioning, and the speed branch focuses on temporal motion), ensuring extraction efficiency and accuracy. Secondly, by adopting the skip connection of the U-Net3 architecture (fusing high- and low-dimensional common features) and cross-branch information complementarity (such as the area where semantic classification guides depth prediction), the problem of lack of spatial information in semantics or lack of category information in depth during single feature extraction is solved, making the fusion between features more efficient. At the same time, the speed prediction branch integrates common features with sparse measurements to expand discrete speed data (such as radar can only detect a few targets) into pixel-level motion features (covering targets in the entire scene), and introduces spatiotemporal modeling to capture motion trends (such as acceleration and trajectory), improving the integrity of dynamic environment perception. MobileNetV2 is used as the encoder to reuse feature extraction capabilities and lightweight, reducing the computing resources of the architecture and avoiding computational delays and slow response speeds.

[0075] In a specific embodiment, the preset feature fusion optimization strategy in S1 specifically includes:

[0076] The universal features output by the encoder are input to the initial layers of the dual decoder and the speed prediction branch respectively; the dual decoder preliminarily extracts semantic attribute features and depth features based on the universal features; the speed prediction branch generates preliminary pixel-level motion features based on the universal features;

[0077] Combine the encoder's universal features with the preliminary pixel-level motion features generated by the speed prediction branch in the channel dimension to obtain fused features.

[0078] The fused features are input into the deep layers of the dual decoder and the speed prediction branch, respectively. The initial semantic attribute features, depth features, and motion features are optimized (e.g., correcting semantic classification errors, improving depth measurement accuracy, and optimizing motion trend prediction). Ultimately, driving perception information including semantic attribute features, depth features, and motion attributes is obtained.

[0079] The parameters of the preset encoder and decoder architecture (such as attention mechanism parameters, transposed convolution kernel weights and biases) are jointly optimized through multiple loss functions to ensure that the obtained driving perception information meets the preset accuracy standards (such as semantic classification accuracy ≥95% and depth error ≤1m); among them, the multiple loss functions include: semantic loss, depth loss, speed loss, and acceleration loss.

[0080] This application uses preliminary extraction at the initial layer, feature fusion enhancement, and deep-level refinement optimization to enable the fused features to play a targeted role at different levels. That is, the initial layer retains the original spatial details of the features (such as target edges and motion trajectory contours), the fusion layer integrates dynamic and static information (such as the category, distance and speed of the vehicle in front), and the deep layer focuses on key targets (such as pedestrians and vehicles at intersections) through the attention mechanism, achieving a perception upgrade from coarse to fine. Secondly, the fused features are based on general features and combined with dynamic preliminary motion features as the core, which not only provides a unified feature benchmark for the dual decoder and speed prediction branch, avoiding the semantic, depth, and motion features from fighting each other, but also feeds back the static features through motion information, making the depth measurement of dynamic targets more accurate. Furthermore, the four loss functions respectively constrain different features (such as semantic classification, depth distance, speed / acceleration), and jointly optimize the parameters of the entire architecture to ensure that the fused features can contain multi-dimensional information while more accurately matching the real scene.

[0081] In a specific embodiment, S2 verifies the accuracy of the driving perception result and optimizes the parameters of the preset encoder and decoder architecture based on the verification result to obtain the target encoder and decoder architecture, specifically including:

[0082] Collect real image information of the target object, including: real segmentation information, real depth information and real motion information;

[0083] Compare and analyze the real image information with the driving perception result output in step S1, and calculate the error index;

[0084] If all error indicators are less than the preset threshold, the driving perception result is determined to meet the design requirements, and the preset architecture is the target encoder and decoder architecture;

[0085] If any error indicator is greater than or equal to the preset threshold, it indicates that the perception result does not meet the requirements. The parameters of the encoder, dual decoder and speed prediction branch are optimized based on error backpropagation, and the parameters of the preset encoder and decoder architecture are continued to be optimized until all error indicators are less than the preset threshold, and the target encoder and decoder architecture is obtained.

[0086] It can be understood that the present application simultaneously verifies the accuracy of semantic attribute features (category), depth features (distance), and motion attributes (speed, acceleration), thereby achieving a complete match between the real image information and the driving perception information output by step S1, completing comprehensive coverage of the perception dimension, and avoiding the overall perception bias caused by verifying only some features, such as mistakenly identifying "nearby pedestrians" as "distant trees", which in turn leads to decision-making errors.

[0087] In a specific embodiment, S3: constructing an environment state based on the driving perception information output by the target encoder and decoder architecture, and generating an initial execution action in the autonomous driving process through an initial execution strategy, specifically includes:

[0088] The decision-making process of the vehicle's execution action during the autonomous driving process is implemented through a reinforcement learning algorithm, wherein a decision-making process model is used as the basic framework of reinforcement learning to output the execution action under different environmental conditions; the decision-making process model adopts a Markov decision process model;

[0089] The input data of the Markov decision process model is the environmental state data; the environmental state data is constructed based on the driving perception information output by the target encoder and decoder architecture;

[0090] Based on the above environmental state data, an initial execution strategy (such as predefined safe driving rules or basic strategy model) is used to output an initial execution action;

[0091] After the vehicle performs the initial action, the experience tuple generated by the interaction is stored in the experience replay pool;

[0092] The experience tuple includes: the vehicle's current environment state s, the executed action a, the immediate reward r, and the vehicle's environment state s' at the next moment; wherein, the next environment state s' and the immediate reward r are generated after the executed action a interacts with the environment.

[0093] Specifically, the environmental state is constructed based on the multimodal perception information output by the target encoder, ensuring that the decision input can fully reflect the actual road conditions and avoid inappropriate actions caused by the disconnection between traditional perception and decision-making; this application uses the Markov decision process model as the basic framework to make the decision process quantifiable and optimizable, thereby guiding it to safe and efficient goals through the reward function, and has the advantages of strong adaptability; then, actions are generated and experience tuples are stored through the initial execution strategy, which not only provides raw data for the strategy training of S4 later, but also provides a basic guarantee for subsequent optimization through the trial and error-feedback mechanism; ultimately, the output of the initial execution strategy ensures that the action will not deviate from the safety boundary, and has the advantages of accurate perception and reasonable decision output.

[0094] Furthermore, S4: training the initial execution strategy and updating the decision process model weight parameters in the initial execution strategy to obtain an optimized decision process model, specifically including:

[0095] S401: Randomly sample experience tuples (s, a, r, s') from the experience replay pool based on the state-action value function Q n (s,a) Calculate the expected cumulative reward Q for taking action a in the current environment state s;

[0096] S402: Update the Q value in real time through the Bellman equation;

[0097] S403: Use the DDPG algorithm to process the continuous state-action space of autonomous driving, randomly sample samples through the experience replay pool, continuously iterate the above process, update the weight parameters of the decision process model, and obtain the optimized decision process model.

[0098] It can be understood that by processing continuous state and action space through the DDPG algorithm, it can output more precise control quantities and adapt to the requirements of driving scenarios for action smoothness; the random sampling of the experience replay pool breaks the temporal correlation of the data, avoiding parameter bias caused by continuous similar scene samples. Combined with the iterative update mechanism of DDPG, it effectively alleviates policy oscillations in reinforcement learning.

[0099] Furthermore, the average reward value after N consecutive rounds tends to be stable or the fluctuation range is ≤ the preset threshold;

[0100] Or the update amplitude of the weight parameters of the decision-making process model is less than the preset threshold, indicating that the parameters have converged;

[0101] Or the performance indicators of the initial execution strategy in the preset evaluation scenario meet the preset standards (such as collision rate ≤ 0.1%, average vehicle speed ≥ 40 km / h);

[0102] After the above conditions are met, the iteration is stopped and the optimized decision process model is obtained.

[0103] Furthermore, S5: Based on the updated decision-making process model weight parameters, a mapping relationship between the environment state and the execution action is established through a preset network, and the decision-making process model is stably converged to the optimal strategy, specifically including:

[0104] S501: Based on the updated decision process model weight parameters, the deep deterministic policy network of the DDPG algorithm is used to construct a mapping relationship between the environment state and the execution action;

[0105] S502: During policy iteration, the training network updates the parameters of the training network in real time based on the immediate reward and cumulative reward obtained for each executed action;

[0106] S503: Using a dual-network mechanism, the weights of the target network are controlled to converge to the training network by updating coefficients to ensure stable optimization of the strategy. The dual-network mechanism includes a training network and a target network, both of which are subnetworks of the decision-making process model.

[0107] Among them, the training network is an online network that receives environmental feedback in real time and dynamically updates weights, which is used to directly generate current actions and quickly learn new experiences; the target network is a lagged copy of the training network. By slowly following the weight updates of the training network, it eventually converges from the weights of the training network to the long-term training network, realizing the optimal strategy of maximizing the accumulated rewards in the long term.

[0108] Specifically, a precise mapping relationship between environmental states and execution actions is constructed through a deep deterministic policy network, achieving a direct connection between perception information and control instructions, and ensuring that the action output fits the real-time road conditions; at the same time, it relies on a dual-network mechanism to achieve stable convergence, that is, the training network dynamically adjusts parameters to adapt to new scenarios, and the target network slowly synchronizes parameters through soft updates to avoid drastic fluctuations, effectively eliminating policy oscillations; ultimately, guided by the long-term maximization of cumulative rewards, the decision-making process model converges smoothly from dynamic trial and error to stable output, ensuring action flexibility in complex scenarios while taking into account the safety and efficiency of long-term driving.

[0109] In summary, this application first achieves high-precision extraction of perception information, and then dynamically and finely optimizes the decision-making process model until the stable convergence of the output action, ultimately ensuring the control accuracy of the execution strategy. The final output execution action can not only adapt to subtle changes in continuous states and action spaces, but also maintain stable and precise control effects in complex scenarios, reducing the decision-making deviation of the execution strategy and improving the safety of automatic navigation driving.

[0110] Figure 2 As shown, this application provides an autonomous driving decision and control system based on environmental perception, including:

[0111] a fusion perception module configured to employ a preset encoder and decoder architecture and a preset feature fusion optimization strategy to obtain driving perception information of a target object during autonomous driving; wherein the driving perception information includes at least semantic attribute features, depth features, and motion attributes;

[0112] a perception verification module configured to verify the accuracy of the driving perception results and optimize parameters of the preset encoder and decoder architectures based on the verification results to obtain target encoder and decoder architectures;

[0113] a decision execution module configured to construct an environment state based on the driving perception information output by the target encoder and decoder architecture, and obtain an initial execution action of the vehicle during the autonomous driving process through an initial execution strategy;

[0114] a decision building module configured to train the initial execution strategy and update the decision process model weight parameters in the execution strategy to obtain an optimized decision process model;

[0115] The decision optimization module is configured to build a mapping relationship between the environment state and the execution action based on the updated decision process model weight parameters through a preset network, and stably converge the decision process model to the optimal strategy.

[0116] It is worth noting that although this system only discloses a fusion perception module, a perception verification module, a decision execution module, a decision construction module and a decision optimization module, it does not mean that this device is limited to the above-mentioned basic functional modules. On the contrary, what the present invention wants to express is that, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with the existing technology to form an infinite number of embodiments or technical solutions. In other words, this system is open rather than closed. Just because this embodiment only discloses individual basic functional modules, it cannot be considered that the scope of protection of the claims of the present invention is limited to the above-mentioned basic functional modules.

[0117] On the other hand, the present application provides an electronic device, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0118] A computer program is stored in the memory, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method.

[0119] On the other hand, the present application provides a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of the method when the computer program / instruction is executed by a processor.

[0120] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0121] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program may be stored in a non-volatile computer-readable storage medium. When executed, the computer program may include the processes of the above-described method embodiments. Any reference to memory, database, or other media used in the embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRRAM), etc. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM may take various forms. The databases referred to in the embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic devices based on quantum computing, etc., but are not limited thereto.

[0122] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0123] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. An autonomous driving decision-making and control method based on environmental perception, characterized in that: The method comprises the following steps: S1: Using a preset encoder and decoder architecture and a preset feature fusion optimization strategy, obtain driving perception information of a target object during autonomous driving; wherein the driving perception information includes at least semantic attribute features, depth features, and motion attributes; S2: Verify the accuracy of the driving perception results and optimize the parameters of the preset encoder and decoder architecture based on the verification results to obtain the target encoder and decoder architecture; S3: Constructs the environment state based on the driving perception information output by the target encoder and decoder architecture, and obtains the initial execution action of the vehicle during the autonomous driving process through the initial execution strategy; S4: training the initial execution strategy and updating the decision-making process model weight parameters in the execution strategy to obtain an optimized decision-making process model; S5: Based on the updated decision-making process model weight parameters, the mapping relationship between the environment state and the execution action is constructed through the preset network, and the decision-making process model is stably converged to the optimal strategy.

2. The method according to claim 1, characterized in that The preset encoder and decoder architecture in step S1 adopts the U-Net3 architecture, which includes: an encoder, a dual decoder and a speed prediction branch; The encoder uses MobileNetV2 as a pre-trained encoder model to extract common features of the target image; Dual decoders, including: Semantic decoder, used to upsample and semantically reconstruct the fused features and output the semantic attribute features of the target object; Depth decoder, used to perform geometric feature analysis on the fused features and output the depth features of the target object; The velocity prediction branch is used to predict the motion properties of the target object based on the universal features output by the encoder and convert the sparse velocity measurements into pixel-level motion features; the motion properties include the target object's velocity, acceleration, movement trajectory, and relative speed to the vehicle.

3. The method according to claim 2, characterized in that The preset feature fusion optimization strategy in S1 specifically includes: The universal features output by the encoder are input to the initial layers of the dual decoder and the speed prediction branch respectively; the dual decoder preliminarily extracts semantic attribute features and depth features based on the universal features; the speed prediction branch generates preliminary pixel-level motion features based on the universal features; Combine the encoder's universal features with the preliminary pixel-level motion features generated by the speed prediction branch in the channel dimension to obtain fused features. The fused features are input into the deep layer of the dual decoder and the deep layer of the speed prediction branch respectively, and finally the driving perception information including semantic attribute features, depth features and motion attributes is obtained; The parameters of the preset encoder and decoder architecture are jointly optimized through multiple loss functions to ensure that the driving perception information obtained meets the preset accuracy standards.

4. The method according to claim 3, characterized in that S2, specifically including: Collect real image information of the target object; Compare and analyze the real image information with the driving perception result output in step S1, and calculate the error index; If all error indicators are less than the preset threshold, the driving perception result is determined to meet the design requirements, and the preset architecture is the target encoder and decoder architecture; If any error indicator is greater than or equal to the preset threshold, it indicates that the perception result does not meet the requirements. The parameters of the encoder, dual decoder and speed prediction branch are optimized based on error backpropagation, and the parameters of the preset encoder and decoder architecture are continued to be optimized until all error indicators are less than the preset threshold, and the target encoder and decoder architecture is obtained.

5. The method according to claim 4, characterized in that S3, specifically including: The decision-making process of the vehicle's execution action during the autonomous driving process is achieved through a reinforcement learning algorithm, wherein a decision-making process model is used as the basic framework of reinforcement learning to output the execution action under different environmental conditions; the decision-making process model adopts a Markov decision process model; The input data of the Markov decision process model is the environmental state data; the environmental state data is constructed based on the driving perception information output by the target encoder and decoder architecture; Based on the above environmental status data, the initial execution action is output through the initial execution strategy; After the vehicle performs the initial action, the experience tuple generated by the interaction is stored in the experience replay pool; The experience tuple includes: the vehicle's current environment state s, the executed action a, the immediate reward r, and the vehicle's environment state s' at the next moment; wherein, the next environment state s' and the immediate reward r are generated after the executed action a interacts with the environment.

6. The method according to claim 5, characterized in that Step S4 specifically includes: S401: Randomly sample experience tuples (s, a, r, s') from the experience replay pool based on the state-action value function Q n (s,a) Calculate the expected cumulative reward Q for taking action a in the current environment state s; S402: Update the Q value in real time through the Bellman equation; S403: Use the DDPG algorithm to process the continuous state-action space of autonomous driving, randomly sample samples through the experience replay pool, continuously iterate the above process, and update the weight parameters of the decision process model.

7. The method according to claim 6, characterized in that Step S5 specifically includes: S501: Based on the updated decision process model weight parameters, a mapping relationship between the environment state and the execution action is constructed through a deep deterministic policy network; S502: Based on the reward obtained for each executed action, the parameters of the training network are updated in real time; S503: A dual-network mechanism is used to control the convergence of the target network weights to the training network by updating the coefficients to ensure stable optimization of the strategy; the dual-network mechanism includes: a training network and a target network, both of which are sub-networks of the decision process model.

8. An autonomous driving decision and control system based on environmental perception, characterized in that: include: a fusion perception module configured to employ a preset encoder and decoder architecture and a preset feature fusion optimization strategy to obtain driving perception information of a target object during autonomous driving; wherein the driving perception information includes at least semantic attribute features, depth features, and motion attributes; a perception verification module configured to verify the accuracy of the driving perception results and optimize parameters of the preset encoder and decoder architectures based on the verification results to obtain target encoder and decoder architectures; a decision execution module configured to construct an environment state based on the driving perception information output by the target encoder and decoder architecture, and obtain an initial execution action of the vehicle during the autonomous driving process through an initial execution strategy; a decision building module configured to train the initial execution strategy and update the decision process model weight parameters in the execution strategy to obtain an optimized decision process model; The decision optimization module is configured to build a mapping relationship between the environment state and the execution action based on the updated decision process model weight parameters through a preset network, and stably converge the decision process model to the optimal strategy.

9. An electronic device comprising: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; characterized in that a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the method described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that A computer program executable by an electronic device is stored, and when the computer program is run on the electronic device, the electronic device executes the steps of the method according to any one of claims 1 to 8.

Images