Automatic driving safety control method based on double-neurotransmitter coordinated regulation

Abstract

The invention, which belongs to the technical field of automatic driving safety control, discloses an automatic driving safety control method based on double-neurotransmitter cooperative regulation and control, comprising the following steps: acquiring multi-modal environment data of a lane scene for feature fusion, generating a virtual scene through a hybrid strategy network, quantifying collision risk and environment uncertainty through a world model, and determining the collision risk and the environment uncertainty; obtaining a fear emotion stimulation value; the concentrations of noradrenaline and 5-hydroxytryptamine are dynamically adjusted on the basis of a fear emotion stimulation value, a stress reaction mechanism is triggered when the concentration of noradrenaline is high, a deep deterministic strategy gradient algorithm is combined with an action safety layer, and an instantaneous safety action is forcibly generated through mixed integer optimization. And triggering a long-term planning mechanism when the concentration of the 2, 5-hydroxytryptamine is high, and generating an optimal action with controllable global risk through trust domain strategy optimization based on a security constraint reinforcement learning framework. According to the method, the problem of collaborative optimization of instantaneous safety and long-term planning of automatic driving in a complex scene can be solved.

Description

TECHNICAL FIELD

[0001] The application relates to a dual-neurotransmitter synergistic regulation automatic driving safety control method and belongs to the technical field of automatic driving safety control. BACKGROUND

[0002] Existing automatic driving safety control technologies mainly rely on rule-driven or single reinforcement learning strategies, and have many limitations.

[0003] Firstly, the dynamic risk response capability is insufficient, and traditional methods are difficult to quantify the environmental uncertainty such as adversarial vehicle behavior and sudden obstacles, so that the risk avoidance strategy in the emergency scene is not good in robustness.

[0004] Secondly, the extreme scene generalization ability is weak, and the training based on real data is difficult to simulate the extreme interaction scene with high collision probability due to the insufficient coverage of long-tail risk.

[0005] Thirdly, the decision mode is single, and there is a lack of dynamic synergistic mechanism for instantaneous safety (such as emergency braking and obstacle avoidance) and long-term planning (such as path optimization), which easily leads to the imbalance of "over-conservative" or "blind risk" decision-making.

[0006] Finally, the biological explainability is missing, and most algorithms are implemented in a black box form without the bionic design of human driver risk response logic, which is difficult to meet the compliance verification needs of high safety level automatic driving. SUMMARY

[0007] The purpose of the present application is to provide a dual-neurotransmitter synergistic regulation automatic driving safety control method, which can solve the problem of synergistic optimization of instantaneous safety and long-term planning of automatic driving in complex scenes, realize the efficient fusion of emergency avoidance and global optimization, and improve the environmental adaptability, strategy generalization ability and all-time safety robustness of the automatic driving system.

[0008] To achieve the above purpose, the present application provides the following technical scheme: In a first aspect, the present application provides a dual-neurotransmitter synergistic regulation automatic driving safety control method, comprising: constructing a lane scene and obtaining multi-modal environment data of the lane scene; converting the multi-modal environment data to a bird's eye view feature space, and performing feature fusion through elastic unwinding technology and multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles; based on the scene representation, generating a virtual scene through a hybrid policy network to simulate the multi-element interaction behavior of surrounding vehicles, and quantifying the collision risk and environmental uncertainty through a world model to obtain a fear emotion stimulus value; The concentration of norepinephrine and serotonin is dynamically adjusted based on the fear emotional stimulus value, and the switching between the stress response mechanism and the long-term planning mechanism is realized through an S-shaped function; When the concentration of norepinephrine is greater than the concentration of serotonin, the stress response mechanism is triggered, which adopts a deep deterministic policy gradient algorithm combined with an action safety layer to generate instantaneous safe actions through mixed integer optimization; When the concentration of norepinephrine is less than the concentration of serotonin, the long-term planning mechanism is triggered, which is based on a safety-constrained reinforcement learning framework to generate globally risk-controllable optimal actions through trust region policy optimization; The instantaneous safe action or the globally risk-controllable optimal action is executed, and the network parameters of the stress response mechanism and the long-term planning mechanism are updated according to the environmental feedback to complete the closed-loop safety control.

[0009] In combination with the first aspect, further, the lane scene is constructed based on an open-source traffic simulation platform, and the lane scene includes urban roads, highways, and sudden obstacle situations, and the lane scene simulates continuous traffic situations from normal driving to extreme emergency situations by dynamically adjusting traffic density, intersection complexity, and pedestrian participation.

[0010] In combination with the first aspect, further, acquiring the multi-modal environment data of the lane scene includes: acquiring image data of the lane scene through a multi-angle camera; acquiring point cloud data of the lane scene through a laser radar; the multi-modal environment data is composed of the image data and the point cloud data.

[0011] In combination with the first aspect, further, the multi-modal environment data is converted to a bird's eye view feature space, and feature fusion is performed through elastic unwinding technology and multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles, including: the image data is input to a Swin-T network after image reduction and image cropping for multi-scale feature extraction, and the image is divided into multiple windows for local self-attention calculation through a hierarchical window attention mechanism, generating a multi-resolution feature map, and the multi-resolution feature map is fused through a feature pyramid network to generate a unified resolution feature map, and the probability of each pixel of the unified resolution feature map belonging to each depth interval is predicted through discrete depth prediction, combined with camera geometric projection to generate 3D pseudo-pixel features, and the 3D pseudo-pixel features are compressed through a space-to-channel operation to obtain a 2D bird's eye view image data feature map; the point cloud data is divided into uniform 3D voxel grids, and the features of the points in each non-empty voxel are aggregated to generate 3D voxel features, and the 3D voxel features are compressed through a maximum pooling operation to obtain a 2D bird's eye view point cloud data feature map. The image data feature map and the point cloud data feature map of the 2D bird's eye view are fused by a stitching technology, and a local alignment deviation is compensated by a full convolution bird's eye view encoder to obtain unified bird's eye view features; The unified bird's eye view features are disentangled into global panoramic features and local close-up features by an elastic disentangling technology, and cross-modal attention focusing is performed on the global panoramic features and the local close-up features to generate enhanced features; Based on the enhanced features, a scene representation focusing on surrounding vehicles and key obstacles is generated.

[0012] In combination with the first aspect, further, the hybrid policy network includes a normal policy subnetwork for generating normal actions of the surrounding vehicles and a dangerous policy subnetwork for generating dangerous actions of the surrounding vehicles, and the hybrid policy network controls the danger level of the virtual scene by dynamically adjusting the sampling proportion of the normal actions of the surrounding vehicles and the dangerous actions of the surrounding vehicles; The hybrid policy network is: ; Wherein, represents the hybrid policy network, represents a dynamic weight for adjusting the sampling proportion of the normal actions of the surrounding vehicles and the dangerous actions of the surrounding vehicles, represents the normal policy subnetwork, represents the dangerous policy subnetwork, represents an environment state, represents normal policy subnetwork parameters, represents dangerous policy subnetwork parameters.

[0013] In combination with the first aspect, further, the collision risk and the environmental uncertainty are quantified by a world model to obtain a fear emotion stimulus value, including: The environment state and the ego vehicle action and the surrounding vehicle action generated by the hybrid policy network are input into the world model, the collision risk and the environmental uncertainty in the vehicle interaction process are predicted, and the predicted value is taken as the fear emotion stimulus value of the ego vehicle; The world model is: ; Wherein, represents the th independently trained diagonal Gaussian model in the world model, represents the total number of independently trained diagonal Gaussian models in the world model, represents an environment state, represents an ego vehicle action, represents a state-action pair, represents to a predicted average of collision risk in the vehicle interaction process, denotes a predicted standard deviation of environmental uncertainty in the vehicle interaction process, denotes a normal distribution; a fear emotion stimulus value of the ego vehicle is: ; wherein, denotes a fear emotion stimulus value of the ego vehicle, denotes a trade-off coefficient for measuring the proportion of collision risk and environmental uncertainty in the fear emotion stimulus value, denotes a collision risk in the vehicle interaction process, , denotes an environmental uncertainty in the vehicle interaction process, .

[0014] In combination with the first aspect, further, a dynamic adjustment formula of the concentrations of norepinephrine and 5-hydroxytryptamine is: ; wherein, , respectively denote , the concentration of norepinephrine at the moment, , respectively denote , the concentration of 5-hydroxytryptamine at the moment, , respectively denote a weight for controlling the change of the concentrations of norepinephrine and 5-hydroxytryptamine, , respectively denote a weight for controlling the release of the fear emotion stimulus of norepinephrine and 5-hydroxytryptamine, denotes a fear emotion stimulus value of the ego vehicle, denotes an environmental state, denotes an ego vehicle action; a switching formula of the stress reaction mechanism and the long-term planning mechanism is: ; wherein, denotes a switching value, in the range of denotes triggering the stress reaction mechanism, in the range of denotes triggering the long-term planning mechanism, denotes an S-shaped function for constraining to the range of . , , respectively, represent the linear combination weights of the degree of influence of norepinephrine and serotonin concentrations.

[0015] Building upon the first aspect, the stress response mechanism further employs a deep deterministic strategy gradient algorithm combined with an action safety layer. Through mixed-integer optimization, it forces the generation of instantaneous safe actions, including: The original actions output by the performer network of the deep deterministic policy gradient algorithm are parameterized by the action safety layer to obtain the action parameters of the original actions, and a safe reachable set is constructed based on the action parameters. By using a mixed-integer nonlinear programming model to control the non-intersection between the safe reachable set and the dangerous region, and transforming the non-intersection between the safe reachable set and the dangerous region into a polynomial constraint on the action parameters, a mixed-integer optimization problem is constructed, and the original action is projected onto an instantaneous safe action that satisfies the polynomial constraint. Parameterization is as follows: ; in, This represents the original action output by the performer network in the deep deterministic policy gradient algorithm. Indicates used for Perform parameterized generator matrix, Indicates used to Convert the action parameters into parameter form. This represents the initial state used to construct a safe and reachable set; The secure reachable set is: ; in, express The corresponding secure reachable set, express The One portion, express Dimensions The first term used to describe the nonlinear effect of an action on a state is... Generator matrices corresponding to polynomials This represents the total number of polynomials used to describe the nonlinear effect of an action on a state. The first parameter in the noise parameter vector represents the second parameter. One portion, express The corresponding generator matrix, The dimension of the noise parameter vector; The danger zone is: ; in, Indicates a danger zone. , These represent the x-coordinate and y-coordinate of the vehicle, respectively. , Let x and y represent the x and y coordinates of the centers of the surrounding vehicles, respectively. Indicates the safety radius; The polynomial constraints on the motion parameters are: ; in, This represents a polynomial constraint on the action parameters. The first term used to describe the nonlinear effect of an action on a state is... The polynomials corresponding to the Relevant index parameters, This represents the threshold for action parameter constraints.

[0016] Building upon the first aspect, the long-term planning mechanism, based on a security-constrained reinforcement learning framework, generates optimal actions with controllable global risk through trust domain policy optimization, including: Based on a safety-constrained reinforcement learning framework, the cumulative expected fear cost constraint is transformed into an unconstrained optimization dual problem through the Lagrange multiplier method, and the optimal action with global risk control is generated based on the update magnitude of the policy neural network constrained by KL divergence. The cumulative expected fear cost constraint is: ; in, Indicates the environmental state. Indicates the action of the vehicle. Represents a state-action pair. , They represent The corresponding reward function and fear cost function, i.e. In the future Accumulated expected reward and accumulated expected fear cost within a time step Indicates in Choose below The probability distribution of actions, express The corresponding expectation, express The state distribution corresponding to the generated trajectory, express The corresponding expectation, Indicates in Choose below The distribution of parameterized strategies express and The degree of difference Denotes KL divergence, Indicates the safety constraint threshold. Indicates the trust domain constraint threshold; Introducing the Lagrange multiplier and The cumulative expected fear cost constraint is transformed into an unconstrained optimization dual problem: ; The KL divergence constraint formula is: ; in, , These represent the parameters of the policy neural network before and after the update, respectively. , They represent , Corresponding to Choose below strategy, express and The degree of difference Indicates in Choose below The optimal action probability distribution, express The state distribution corresponding to the generated trajectory, express The corresponding expectation, Indicates in Down The corresponding expectation.

[0017] Building upon the first aspect, furthermore, by updating the network parameters of the stress response mechanism and long-term planning mechanism based on environmental feedback, closed-loop safety control is achieved, including: The environmental feedback is stored in the experience pool, which includes the current environmental state, the current vehicle action, the current reward, and the next environmental state. For the stress response mechanism, the action value function is optimized by minimizing the temporal difference error, and the network parameters of the commentator network of the deep deterministic policy gradient algorithm are updated. The network parameters of the performer network of the deep deterministic policy gradient algorithm are then updated in conjunction with the action safety layer. For long-term planning mechanisms, the network parameters of the reward function and fear cost function are updated through gradient descent. The loss function for updating the network parameters of the critic network in the deep deterministic policy gradient algorithm is: ; in, The loss represents the loss of the critic network used to update the network parameters of the deep deterministic policy gradient algorithm. , , , They represent The environmental state at any time, The constant movement of the vehicle Momentary rewards The environmental state at any given time, Represents the experience pool. express Down The corresponding expectation, The network parameters of the critic network represent the deep deterministic policy gradient algorithm. express A network of critics for the deep deterministic policy gradient algorithm and The value estimate, The target network parameters of the critic network represent the parameters of the deep deterministic policy gradient algorithm. The performer network representing the deep deterministic policy gradient algorithm in The actions generated below, express A network of critics for the deep deterministic policy gradient algorithm and The value estimate, Indicates the discount factor; The update formula for the network parameters of the performer network in the deep deterministic policy gradient algorithm is as follows: ; in, The network parameters of the performer network in the deep deterministic policy gradient algorithm are represented. This represents the learning rate of the performer network in the deep deterministic policy gradient algorithm. Describe the objective function gradient, express Down The corresponding expectation, The performer network representing the deep deterministic policy gradient algorithm in The actions generated below gradient, express Performer networks for deep deterministic policy gradient algorithms and Value estimation The gradient; The loss function for updating the network parameters of the reward function and the fear cost function is: ; wherein, 、 respectively represent the network parameters of the updating reward function, the fear cost function 、 loss, 、 respectively represent the reward value estimation, the fear cost value estimation of the reward function, the fear cost function on 、 、 respectively represent the reward target value, the fear cost target value.

[0018] In a second aspect, the present application provides a dual-neurotransmitter synergistic regulation automatic driving safety control device, comprising: a data processing module, configured to construct a lane scene and acquire multi-modal environment data of the lane scene; convert the multi-modal environment data to a bird's eye view feature space, and perform feature fusion through an elastic unwinding technology and a multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles; based on the scene representation, generate a virtual scene through a hybrid policy network to simulate multi-element interaction behaviors of the surrounding vehicles, and through a world model, quantify collision risks and environmental uncertainties to obtain a fear emotion stimulus value; a control switching module, configured to dynamically adjust concentrations of norepinephrine and 5-hydroxytryptamine based on the fear emotion stimulus value, and realize switching of a stress response mechanism and a long-term planning mechanism through an S-shaped function; when the concentration of norepinephrine is greater than the concentration of 5-hydroxytryptamine, the stress response mechanism is triggered, the stress response mechanism adopts a deep deterministic policy gradient algorithm combined with an action safety layer to generate an instantaneous safe action through mixed integer optimization; when the concentration of norepinephrine is less than the concentration of 5-hydroxytryptamine, the long-term planning mechanism is triggered, the long-term planning mechanism is based on a safety-constrained reinforcement learning framework to generate an optimal action with controllable global risks through trust region policy optimization; an action execution module, configured to execute the instantaneous safe action or the optimal action with controllable global risks, and update network parameters of the stress response mechanism and the long-term planning mechanism according to environmental feedback to complete a closed-loop safety control.

[0019] In a third aspect, the present application provides a computer device, comprising: a storage medium: configured to store a computer program; a processor: configured to execute the computer program to realize the dual-neurotransmitter synergistic regulation automatic driving safety control method of the first aspect.

[0020] ​​In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the dual-neurotransmitter coordinated automatic driving safety control method of the first aspect.

[0021] In a fifth aspect, the present application provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the dual-neurotransmitter coordinated automatic driving safety control method of the first aspect.

[0022] Compared with the prior art, the present application has the following beneficial effects: The dual-neurotransmitter coordinated automatic driving safety control method provided by the present application can realize the adaptive switching of the instantaneous safety hard constraint in the emergency scene and the long-term planning soft constraint in the normal scene by simulating the release rules of norepinephrine (NE) and 5-hydroxytryptamine (5-HT), and constructing a "fear value-neurotransmitter concentration-decision mode" closed loop link, which can significantly improve the decision flexibility and safety of the automatic driving system in a dynamic environment. Based on the mixed policy network and the world model, the "normal-extreme" continuous interaction scene is generated, and the environment uncertainty and collision risk are quantified by the fear value feedback, which can break through the real data dependence bottleneck and enhance the generalization ability of the automatic driving system to the long-tail risk scene. By using the S-shaped function to constrain the dynamic range of the neurotransmitter concentration, the abstract decision process is mapped to the quantifiable neurotransmitter signal (NE / 5-HT), which can support the traceability of the policy behavior and the safety compliance verification, and provide transparent decision basis for high safety level automatic driving. The design of the neurotransmitter regulation layer and the safety constraint module can be seamlessly integrated into the existing automatic driving perception-planning link, which can ensure the safety of instantaneous action and long-term risk control while improving the technical landing feasibility. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a flow chart of the dual-neurotransmitter coordinated automatic driving safety control method provided by the embodiment of the present application; Figure 2 is a multi-modal data fusion flow chart provided by the embodiment of the present application; Figure 3 is a virtual scene generation and risk assessment schematic diagram provided by the embodiment of the present application; Figure 4 is a dual-neurotransmitter regulation schematic diagram provided by the embodiment of the present application; Figure 5 is a stress response mechanism safety layer optimization schematic diagram provided by the embodiment of the present application; Figure 6 is a long-term planning mechanism constraint optimization schematic diagram provided by the embodiment of the present application; Figure 7 is a dual-neurotransmitter regulation strategy switching schematic diagram provided by the embodiment of the present application. DETAILED DESCRIPTION

[0024] The technical solutions of the present application will be further described in detail below in combination with the specific embodiments.

[0025] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application. The technical features in the embodiments of the present application and the embodiments can be combined with each other without conflict.

[0026] The embodiment of the present application provides an automatic driving safety control method based on dual-neurotransmitter synergistic regulation, comprising: constructing a lane scene and obtaining multi-modal environment data of the lane scene; converting the multi-modal environment data to a bird's-eye view feature space, and performing feature fusion through an elastic unwinding technology and a multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles; based on the scene representation, generating a virtual scene through a hybrid policy network to simulate multi-element interaction behaviors of surrounding vehicles, and quantifying collision risks and environmental uncertainties through a world model to obtain a fear emotion stimulus value; based on the fear emotion stimulus value, dynamically adjusting concentrations of norepinephrine and 5-hydroxytryptamine, and realizing switching of a stress response mechanism and a long-term planning mechanism through an S-shaped function; when the concentration of norepinephrine is greater than the concentration of 5-hydroxytryptamine, triggering the stress response mechanism, the stress response mechanism adopts a deep deterministic policy gradient algorithm combined with an action safety layer to generate instantaneous safe actions through mixed integer optimization; when the concentration of norepinephrine is less than the concentration of 5-hydroxytryptamine, triggering the long-term planning mechanism, the long-term planning mechanism is based on a safety-constrained reinforcement learning framework to generate optimal actions with controllable global risks through trust region policy optimization; executing the instantaneous safe actions or the optimal actions with controllable global risks, and updating network parameters of the stress response mechanism and the long-term planning mechanism according to environmental feedback to complete closed-loop safety control.

[0027] The automatic driving safety control method based on dual neurotransmitter synergistic regulation provided in the embodiment of the present application can solve the problem of synergistic optimization of instantaneous safety and long-term planning of automatic driving in a complex scene, and improve the environmental adaptability, strategy generalization ability and all-time safety robustness of the automatic driving system.

[0028] Figure 1 FIG. 1 is a flowchart of the automatic driving safety control method based on dual neurotransmitter synergistic regulation provided in the embodiment of the present application. The flowchart only shows the logical order of the method of the present embodiment, and the steps shown or described can be completed in an order different from that shown in the case of no conflict. Figure 1

[0029] The automatic driving safety control method based on dual neurotransmitter synergistic regulation provided in the embodiment of the present application can be applied to a terminal and can be executed by an automatic driving safety control device based on dual neurotransmitter synergistic regulation. The device can be realized by software and / or hardware, and can be integrated in the terminal, for example, any tablet computer or computer device with communication function.

[0030] In one possible embodiment, the lane scene is constructed based on an open-source traffic simulation platform. The lane scene includes urban roads, highways and emergency obstacle situations. The lane scene simulates continuous traffic situations from normal driving to extreme emergency situations by dynamically adjusting the traffic flow density, intersection complexity and pedestrian participation.

[0031] Specifically, the present application constructs a multi-dimensional and multi-complexity simulated lane scene using an open-source traffic simulation platform, including typical traffic environments such as urban roads, highways and emergency events. By dynamically adjusting parameters such as traffic flow density, intersection complexity and pedestrian participation, continuous traffic situations from normal driving to extreme emergency situations are simulated. The performance of the agent in the high-fidelity simulation environment is evaluated from the task success rate, risk response time and dynamic behavior indicators, thereby optimizing and improving the safety robustness of the automatic driving system.

[0032] In one possible embodiment, acquiring the multi-modal environment data of the lane scene specifically includes the following steps: Step 1: acquiring image data of the lane scene through a multi-angle camera; ​Step 2: Obtain point cloud data of the lane scene by laser radar; Step 3: Form multi-modal environment data jointly by image data and point cloud data.

[0033] In one possible embodiment, the multi-modal environment data is converted to a bird's eye view feature space, and feature fusion is performed through an elastic unwinding technology and a multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles, which specifically includes the following steps: Step 1: After image reduction and image cropping, the image data is input into a Swin-T network for multi-scale feature extraction, and a layered window attention mechanism is used to divide the image into multiple windows for local self-attention calculation, generating multi-resolution feature maps, and a feature pyramid network is used to fuse the multi-resolution feature maps to generate a unified resolution feature map, and a discrete depth prediction is used to predict the probability of each pixel of the unified resolution feature map belonging to each depth interval, and a camera geometric projection is used to generate a 3D pseudo-pixel feature, and a space-to-channel operation is used to compress the 3D pseudo-pixel feature to obtain a 2D bird's eye view image data feature map; Specifically, as shown in Figure 2 , the RGB image data collected by the multi-angle camera is input into the Swin-T network after image reduction and image cropping to extract multi-scale features, a layered window attention mechanism is used to divide the image into multiple windows for local self-attention calculation to generate multi-resolution feature maps, and a feature pyramid network (FPN) is used to fuse multi-resolution feature maps of different scales to generate a unified resolution feature map with rich semantics , where represents the RGB image data after image reduction and image cropping, represents the Swin-T network, represents the feature pyramid network.

[0034] Through discrete depth prediction (dividing continuous depth into probabilistic discrete intervals), the probability of each pixel of the unified resolution feature map belonging to each depth interval is predicted, and a camera geometric projection (using camera internal and external parameters, mapping 2D features to 3D space according to depth probability along the optical axis) is used to generate a 3D pseudo-pixel feature, and a space-to-channel operation is used to compress the 3D pseudo-pixel feature to obtain a 2D bird's eye view (BEV) image data feature map .

[0035] Step 2: Divide the point cloud data into a uniform 3D voxel grid, and aggregate the features of the points in each non-empty voxel to generate a 3D voxel feature, and compress the 3D voxel feature through a maximum pooling operation to obtain a 2D bird's eye view point cloud data feature map; Specifically, as shown in Figure 2As shown, for point cloud data, i.e. lidar data, first, voxel division processing is performed to divide the original point cloud data into a uniform 3D voxel grid, and then the features of the points in each non-empty voxel are aggregated: , 3D voxel features are generated , 3D voxel features are compressed along the axis using a max-pooling operation: , the point cloud data feature map of the 2D bird's eye view is obtained , wherein, represents the coordinates of the th point cloud data falling into the same voxel grid, represents the voxelization function, represents the average pooling operation, represents the max-pooling operation for compressing 3D voxel features along the axis.

[0036] Step 3: Fuse the image data feature map and the point cloud data feature map of the 2D bird's eye view through a stitching technique, and compensate for local alignment deviations through a fully convolutional bird's eye view encoder to obtain unified bird's eye view features; Specifically, by fusing and , input the fully convolutional bird's eye view encoder (ResNet block) to compensate for local alignment deviations, and output the unified bird's eye view features .

[0037] Step 4: decompose the unified bird's eye view features into global panoramic features and local close-up features through an elastic unwinding technique, and perform cross-modal attention focusing on the global panoramic features and the local close-up features to generate enhanced features; Specifically, by decomposing into global panoramic features and local close-up features , wherein, represents a hyperparameter for controlling the decomposition granularity, represents a stride pooling decomposition, represents a local separation, extracts a large-span global background, focuses on local details, and are alternately stacked to enhance semantic robustness.

[0038] Perform cross-modal attention focusing on the unwound global panoramic features and the local close-up features, and calculate the query-key value attention weight : ; wherein, represents a query, representing the features of the ego vehicle, represents a key, representing the features of all elements in the environment, represents a value, representing detailed feature information of all elements in the environment, represents a position and velocity embedding, adding the spatial position and motion state encoding of each element in the environment, represents the dimension of the key vector, which is a scaling factor for stable training, represents a normalized exponential function, represents a transpose.

[0039] Output enhanced features of focused surrounding vehicles and key obstacles by multi-head attention mechanism .

[0040] Step 5: Based on the enhanced features, generate a scene representation focusing on surrounding vehicles and key obstacles.

[0041] In one possible embodiment, as shown in Figure 3 , the hybrid policy network includes a normal policy subnetwork for generating normal actions of surrounding vehicles (such as safe lane changing and following) and a dangerous policy subnetwork for generating dangerous actions of surrounding vehicles (such as forced lane changing and sudden braking), the hybrid policy network controls the danger level of the virtual scene by dynamically adjusting the sampling proportion of normal actions of surrounding vehicles and dangerous actions of surrounding vehicles, and generates a virtual scene containing multi-element interaction behaviors (normal actions and dangerous actions).

[0042] Specifically, the hybrid policy network is: ; wherein, represents the hybrid policy network, represents a dynamic weight for adjusting the sampling proportion of normal actions of surrounding vehicles and dangerous actions of surrounding vehicles, represents the normal policy subnetwork, represents the dangerous policy subnetwork, represents the environment state, represents the normal policy subnetwork parameters, represents the dangerous policy subnetwork parameters.

[0043] In one possible embodiment, as shown in Figure 3 , the fear emotion stimulus value is quantified by the world model based on collision risk and environmental uncertainty, specifically including: inputting the environment state and ego vehicle action and the surrounding vehicle action generated by the hybrid policy network into the world model, predicting the collision risk and environmental uncertainty in the vehicle interaction process, and taking the predicted value as the fear emotion stimulus value of the ego vehicle.

[0044] Specifically, the world model is: ; wherein, denotes the i-th independently trained diagonal Gaussian model in the world model, denotes the total number of independently trained diagonal Gaussian models in the world model, denotes the environment state, denotes the ego action, denotes the state-action pair, denotes the predicted average collision risk in the vehicle interaction process, representing the probability of collision or safety violation, denotes the predicted standard deviation of environmental uncertainty in the vehicle interaction process, representing the uncertainty of , the higher the uncertainty of ,denotes a normal distribution.

[0045] By averaging the average collision risk of all diagonal Gaussian models, a more reliable collision probability estimate can be obtained.

[0046] By averaging the standard deviation of environmental uncertainty of all diagonal Gaussian models, the overall uncertainty of the environment can be obtained.

[0047] Under the stimulation of fear emotion, two neurotransmitters, norepinephrine and serotonin, are released.

[0048] Specifically, the fear emotion stimulation value of the ego vehicle is: ; wherein, denotes the fear emotion stimulation value of the ego vehicle, denotes the trade-off coefficient for measuring the proportion of collision risk (known risk) and environmental uncertainty (unknown risk) in the fear emotion stimulation value, denotes the collision risk in the vehicle interaction process, , denotes the environmental uncertainty in the vehicle interaction process, .

[0049] In one possible embodiment, the dynamic adjustment formula of the concentrations of norepinephrine and serotonin is: ; wherein, , They represent , The concentration of norepinephrine at all times. , They represent , The concentration of 5-hydroxytryptamine at a given time, , These represent the weights used to control changes in the concentrations of norepinephrine and serotonin, respectively. , These represent the weights of norepinephrine and serotonin release used to control fear-induced emotional responses. This indicates the level of fear associated with the vehicle. Indicates the environmental state. It indicates the movement of the vehicle.

[0050] like Figure 4 As shown, norepinephrine and serotonin, two neurotransmitters, act on the prefrontal cortex (PFC). In the PFC, the switching of decision-making modes (stress response mechanism and long-term planning mechanism) of the PFC is regulated according to the real-time concentration of norepinephrine and serotonin.

[0051] Specifically, the formula for switching between stress response mechanisms and long-term planning mechanisms is as follows: ; in, Indicates the toggle value. exist The range indicates the triggering of the stress response mechanism. exist The range indicates that a long-term planning mechanism has been triggered. Indicates used to Constrained by S-shaped functions within a certain range, specifically the Sigmoid function. , , respectively, represent the linear combination weights of the degree of influence of norepinephrine and serotonin concentrations.

[0052] like Figure 7 As shown, if exist Within this range, it indicates that under fear stimulation, the concentration of norepinephrine is much higher than that of serotonin, leading to hypervigilance and temporary functional failure of the prefrontal cortex. In this case, the autonomous driving system prioritizes the activation of the fear-driven stress response mechanism, rapidly generating instantaneous safety actions to address the threat. If... exist In the range, it is explained that at this time under the fear stimulus, the concentration of norepinephrine is much lower than that of serotonin, and the prefrontal cortex returns to normal and fully exerts the long-term planning ability, at this time the automatic driving system formulates and executes more safe optimal action through comprehensive evaluation of the environment and risk.

[0053] In one possible embodiment, the stress response mechanism adopts a deep deterministic policy gradient algorithm combined with an action safety layer to generate instantaneous safe actions through mixed integer optimization, which specifically includes the following steps: Step 1: The action safety layer is used to parameterize the original action output by the actor network of the deep deterministic policy gradient algorithm to obtain the action parameters of the original action, and based on the action parameters, a safety reachable set is constructed. Step 2: The mixed integer nonlinear programming model is used to control the safety reachable set and the dangerous area to have no intersection, and the safety reachable set and the dangerous area without intersection are converted into polynomial constraints on the action parameters, and a mixed integer optimization problem is constructed to project the original action to the instantaneous safe action that satisfies the polynomial constraints.

[0054] In this embodiment, as shown in Figure 5 , the deep deterministic policy gradient (DDPG) algorithm is used to enhance the real-time safety verification capability of the action generation module. After the actor network outputs the original action, an action safety layer is introduced to use parameterization projection and mixed integer optimization technology to forcibly limit the reachable set of the action to have no intersection with the dangerous area, ensuring that the current action has absolutely no collision risk. The action after safety verification interacts with the environment to obtain the immediate state-reward data, and the action value evaluation model of the critic network is updated based on the time difference (TD) error to form a closed-loop decision link of "safety constraint-action execution-feedback learning".

[0055] Specifically, the parameterization processing is as follows: ; Among them, represents the original action output by the actor network of the deep deterministic policy gradient algorithm, represents the generating matrix used for parameterizing , which defines the change direction and range of the action, represents the action parameters used to convert into a parameter form, , represents the initial state used to construct the safety reachable set.

[0056] Based on the parameterization processing, convert into a parameter form: .

[0057] The safety reachable set is: ; in, express The corresponding secure reachable set, express The One portion, express Dimensions The first term used to describe the nonlinear effect of an action on a state is... Generator matrices corresponding to polynomials This represents the total number of polynomials used to describe the nonlinear effect of an action on a state. The first parameter in the noise parameter vector represents the second parameter. One portion, express The corresponding generator matrix, This represents the dimension of the noise parameter vector.

[0058] In this embodiment, the location of obstacles is obtained in real time through sensors (LiDAR, camera) at the sensing end, and dangerous areas are dynamically generated.

[0059] Specifically, the danger zone is: ; in, Indicates a danger zone. , These represent the x-coordinate and y-coordinate of the vehicle, respectively. , Let x and y represent the x and y coordinates of the centers of the surrounding vehicles, respectively. Indicates the safety radius.

[0060] The polynomial constraints on the motion parameters are: ; in, This represents a polynomial constraint on the action parameters. The first term used to describe the nonlinear effect of an action on a state is... The polynomials corresponding to the Relevant index parameters, This indicates the threshold for action parameter constraints.

[0061] Based on polynomial constraints on the action parameters, an optimization problem is constructed, projecting the original action onto actions that satisfy the safety constraints: ; in, This indicates the motion parameters after safety verification and projection. make sure satisfy all safety constraints of dangerous regions.

[0062] solved by nonlinear programming (MINLP) , converted into safety actions based on parameterization . Perform interact with the environment to get the state and reward of the next moment.

[0063] In one possible embodiment, the long-term planning mechanism is based on a safety constraint reinforcement learning framework, and the globally risk-controllable optimal action is generated by trust region policy optimization, which specifically includes: based on the safety constraint reinforcement learning framework, the cumulative expected fear cost constraint is converted into an unconstrained optimization dual problem by the Lagrange multiplier method, and the update amplitude of the policy neural network is constrained based on the KL divergence, and the globally risk-controllable optimal action is generated.

[0064] In this embodiment, as shown in Figure 6 , under the long-term planning mechanism, the automatic driving system is based on a safety constraint reinforcement learning framework, with balancing long-term reward maximization and cumulative risk controllability as the core goal, by constructing a constraint optimization problem, limiting the cumulative fear cost to be less than the safety threshold, using the critic network to evaluate the long-term value and fear cost of the action, guiding the actor network to generate the optimal action, and using trust region policy optimization (TRPO) to ensure that the policy update is within the safety boundary, while dynamically adjusting the constraint strength combined with the environment feedback, to realize efficient decision-making under the global risk controllable.

[0065] Specifically, the cumulative expected fear cost constraint is: ; Wherein, represents the environment state, represents the action of the ego vehicle, represents the state-action pair, , respectively represent corresponding reward function, fear cost function, i.e. accumulated expected reward, accumulated expected fear cost in the future time steps, represents the action probability distribution selected under , represents corresponding expectation, represents the state distribution corresponding to the trajectory generated by , represents corresponding expectation, represents the action probability distribution selected under , ​the parameterized policy distribution, denotes the difference between , denotes the KL divergence, denotes the safety constraint threshold, denotes the trust region constraint threshold.

[0066] In this embodiment, the Lagrange multiplier and are introduced to convert the cumulative expected fear cost constraint into an unconstrained optimization dual problem: .

[0067] The derivative of is taken and set to zero to obtain the action probability distribution in the current state: ; where is the partition function.

[0068] Substituting into the unconstrained optimization dual problem, we obtain: .

[0069] By solving , we obtain the optimal dual variables and .

[0070] Substituting the optimal dual variables and into , we obtain the unique optimal action probability distribution in the current state: ; where denotes the optimal action probability distribution of under .

[0071] Under the guidance of the optimal action probability distribution, the policy is improved, i.e., the parameterized policy distribution is fitted to the non-parametric optimal distribution while limiting the policy change within the trust region.

[0072] Specifically, the KL divergence constraint formula is: ; where , denote the parameters of the policy neural network after updating and before updating, respectively, , denote , Corresponding to Choose below strategy, express and The degree of difference Indicates in Choose below The optimal action probability distribution, express The state distribution corresponding to the generated trajectory, express The corresponding expectation, Indicates in Down The corresponding expectation.

[0073] In this embodiment, the Lagrange multiplier is introduced. ,get: .

[0074] right Find the derivative, given ,in, express The corresponding expectation is: .

[0075] Parameter updates are performed using the gradient ascent method. ; in, express The corresponding learning rate.

[0076] calculate If the value is close to or exceeds Then increase Strengthen constraints, if much smaller Then decrease Accelerate optimization by iterating repeatedly until the iteration termination condition is met, thus obtaining the optimal parameters of the policy neural network that satisfy the constraints. Under the premise of satisfying safety constraints (controlling costs) and trust domain constraints (ensuring stable policy updates), maximize the expected reward to achieve safe and efficient policy optimization.

[0077] In one possible embodiment, updating the network parameters of the stress response mechanism and the long-term planning mechanism based on environmental feedback to complete closed-loop safety control specifically includes the following steps: Step 1: Store the environmental feedback in the experience pool. The environmental feedback includes the current environmental state, the current vehicle action, the current reward, and the next environmental state. Step 2: for the stress response mechanism, updating the network parameters of the critic network of the deep deterministic policy gradient algorithm by minimizing the error of time difference, updating the network parameters of the actor network of the deep deterministic policy gradient algorithm in combination with the action safety layer; Specifically, the loss function for updating the network parameters of the critic network of the deep deterministic policy gradient algorithm is: ; Wherein, represents the loss of updating the network parameters of the critic network of the deep deterministic policy gradient algorithm, 、 、 、 respectively represent the environment state at time t, the ego vehicle action at time t, the reward at time t, the environment state at time t, represents the experience pool, represents the corresponding expectation, represents the value estimation of the critic network of the deep deterministic policy gradient algorithm for and , represents the target network parameters of the critic network of the deep deterministic policy gradient algorithm, represents the action generated by the actor network of the deep deterministic policy gradient algorithm under , represents the value estimation of the critic network of the deep deterministic policy gradient algorithm for and , represents the target network parameters of the critic network of the deep deterministic policy gradient algorithm, represents the action generated by the actor network of the deep deterministic policy gradient algorithm under , represents the value estimation of the critic network of the deep deterministic policy gradient algorithm for and ,

[0078] represents the discount factor. The update formula of the network parameters of the actor network of the deep deterministic policy gradient algorithm is: ; Wherein, represents the network parameters of the actor network of the deep deterministic policy gradient algorithm, represents the learning rate of the actor network of the deep deterministic policy gradient algorithm, represents the gradient of the objective function , represents the corresponding expectation under the actor network representing the deep deterministic policy gradient algorithm generates an action under the gradient of the action , the gradient of the value estimate of the actor network representing the deep deterministic policy gradient algorithm under and . the gradient of the value estimate of the actor network representing the deep deterministic policy gradient algorithm under

[0079] and .

[0080] Specifically, the loss function for updating the network parameters of the reward function and the fear cost function is: ; wherein, , respectively represent the loss of the network parameters of the reward function and the fear cost function , , , respectively represent the reward value estimate and the fear cost value estimate of the reward function and the fear cost function under , , , , respectively represent the reward target value and the fear cost target value.

[0081] The automatic driving safety control method based on dual neurotransmitter synergistic regulation provided in the embodiment converts multi-source data into bird's-eye view features through multi-modal sensor fusion technology, converts multi-source data into bird's-eye view features, and obtains high-confidence input; based on the scene imagination and risk assessment functions of the brain amygdala, a virtual scene imagination module is designed, a virtual interactive scene is generated through a hybrid policy network, the environmental uncertainty and collision risk during interaction are described using a world model, and a fear emotion stimulus is generated; based on the neural regulation functions of the brain locus ceruleus and raphe nucleus, a dual neurotransmitter (NE-alerting factor, 5-HT-calming factor) synergistic model is constructed, the neurotransmitter dynamic balance is triggered through the fear value, and the decision mode is switched in real time; based on the functions of the prefrontal cortex, a synergistic decision model of stress response (instantaneous safety) and long-term planning (risk control) is constructed, full-time safety control is realized, and the environmental adaptability and safety robustness of the automatic driving decision system can be significantly improved.

[0082] The automatic driving safety control device based on dual neurotransmitter synergistic regulation provided in the embodiment comprises: The data processing module is configured to construct a lane scene and acquire multi-modal environment data of the lane scene; convert the multi-modal environment data to a bird's eye view feature space, and perform feature fusion through an elastic unwinding technology and a multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles; based on the scene representation, generate a virtual scene through a hybrid policy network to simulate multi-element interaction behaviors of the surrounding vehicles, and quantify collision risks and environmental uncertainties through a world model to obtain a fear emotion stimulus value; The control switching module is configured to dynamically adjust concentrations of norepinephrine and 5-hydroxytryptamine based on the fear emotion stimulus value, and switch a stress response mechanism and a long-term planning mechanism through an S-shaped function; when the concentration of norepinephrine is greater than the concentration of 5-hydroxytryptamine, the stress response mechanism is triggered, the stress response mechanism adopts a deep deterministic policy gradient algorithm combined with an action safety layer to generate a transient safe action through mixed integer optimization; when the concentration of norepinephrine is less than the concentration of 5-hydroxytryptamine, the long-term planning mechanism is triggered, the long-term planning mechanism is based on a safety constraint reinforcement learning framework to generate an optimal action with controllable global risks through trust region policy optimization; The action execution module is configured to execute the transient safe action or the optimal action with controllable global risks, and update network parameters of the stress response mechanism and the long-term planning mechanism according to environmental feedback to complete closed-loop safety control.

[0083] The automatic driving safety control device with dual-neurotransmitter collaborative regulation provided by the embodiment of the application can execute the automatic driving safety control method with dual-neurotransmitter collaborative regulation provided by the embodiment of the application, and has the corresponding function modules and beneficial effects of the execution method.

[0084] The embodiment of the application provides a computer device, comprising: A storage medium is configured to store a computer program; A processor is configured to execute the computer program to implement the automatic driving safety control method with dual-neurotransmitter collaborative regulation provided by the embodiment of the application.

[0085] The embodiment of the application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the automatic driving safety control method with dual-neurotransmitter collaborative regulation provided by the embodiment of the application.

[0086] The embodiment of the application provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the automatic driving safety control method with dual-neurotransmitter collaborative regulation provided by the embodiment of the application.

[0087] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0088] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the present application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing device or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0089] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0091] The above merely preferred embodiments of the present application and it is to be understood that those skilled in the technical field of the present application can make a number of improvements and variations without departing from the technical principles of the present application, and these improvements and variations shall be considered as falling within the scope of protection of the present application.

Claims

1. A dual-neurotransmitter co-regulation autonomous driving safety control method, characterized in that, The method comprises the following steps: construct a lane scene and obtain multi-modal environment data of the lane scene; convert the multi-modal environment data to a bird's eye view feature space, and perform feature fusion through elastic unwinding technology and a multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles; based on the scene representation, generate a virtual scene through a hybrid policy network to simulate the multi-element interaction behavior of surrounding vehicles, and quantify collision risk and environmental uncertainty through a world model to obtain a fear emotion stimulus value; based on the fear emotion stimulus value, dynamically adjust the concentration of norepinephrine and 5-hydroxytryptamine, and realize the switching of the stress response mechanism and the long-term planning mechanism through an S-shaped function; when the concentration of norepinephrine is greater than the concentration of 5-hydroxytryptamine, trigger the stress response mechanism, which adopts a deep deterministic policy gradient algorithm combined with an action safety layer to generate a transient safe action through mixed integer optimization; when the concentration of norepinephrine is less than the concentration of 5-hydroxytryptamine, trigger the long-term planning mechanism, which is based on a safety-constrained reinforcement learning framework and generates an optimal action with controllable global risk through trust region policy optimization; execute the transient safe action or the optimal action with controllable global risk, and update the network parameters of the stress response mechanism and the long-term planning mechanism according to the environmental feedback to complete the closed-loop safety control.

2. The dual neurotransmitter co-regulated autonomous driving safety control method according to claim 1, characterized in that, The lane scene is constructed based on an open-source traffic simulation platform, and the lane scene includes urban roads, highways, and sudden obstacle situations. The lane scene simulates continuous traffic situations from normal driving to extreme emergency situations by dynamically adjusting traffic flow density, intersection complexity, and pedestrian participation. 3.The dual-neurotransmitter co-regulated autonomous driving safety control method of claim 1, wherein, Obtaining multi-modal environment data of the lane scene comprises: acquiring image data of the lane scene through multi-angle cameras; acquiring point cloud data of the lane scene through a laser radar; the multi-modal environment data is composed of image data and point cloud data.

4. The dual neurotransmitter co-regulation autonomous driving safety control method of claim 1, wherein, Converting the multi-modal environment data to a bird's eye view feature space and performing feature fusion through elastic unwinding technology and a multi-head attention mechanism to generate a scene representation focusing on surrounding vehicles and key obstacles comprises: after image reduction and image cropping, input the image data into a Swin-T network for multi-scale feature extraction, divide the image into multiple windows through a hierarchical window attention mechanism for local self-attention calculation, generate a multi-resolution feature map, fuse the multi-resolution feature map through a feature pyramid network to generate a uniform resolution feature map, predict the probability of each pixel of the uniform resolution feature map belonging to each depth interval through discrete depth prediction, and generate a 3D pseudo-pixel feature combined with camera geometric projection, compress the 3D pseudo-pixel feature through a space-to-channel operation to obtain a 2D bird's eye view image data feature map; divide the point cloud data into uniform 3D voxel grids, and aggregate the features of the points in each non-empty voxel to generate a 3D voxel feature, and compress the 3D voxel feature through a maximum value pooling operation to obtain a 2D bird's eye view point cloud data feature map. The image data feature map and the point cloud data feature map of the 2D bird's eye view are fused by a stitching technology, and a full convolution bird's eye view encoder is used to compensate for local alignment deviation to obtain unified bird's eye view features; The unified bird's eye view features are disentangled into global panoramic features and local close-up features by an elastic disentangling technology, and cross-modal attention focusing is performed on the global panoramic features and the local close-up features to generate enhanced features; Based on the enhanced features, a scene representation focusing on surrounding vehicles and key obstacles is generated.

5. The dual neurotransmitter co-regulation autonomous driving safety control method of claim 1, wherein, The hybrid policy network includes a normal policy subnetwork for generating normal actions of surrounding vehicles and a dangerous policy subnetwork for generating dangerous actions of surrounding vehicles, and the hybrid policy network adjusts the sampling proportion of the normal actions of surrounding vehicles and the dangerous actions of surrounding vehicles by dynamic weights to control the danger level of the virtual scene; The hybrid policy network is: ； wherein, represents a hybrid policy network, represents a dynamic weight for adjusting the sampling ratio of the normal action of the surrounding vehicle and the dangerous action of the surrounding vehicle, represents a normal policy subnetwork, represents a dangerous policy subnetwork, represents an environment state, represents normal policy subnetwork parameters, represents dangerous policy subnetwork parameters.

6. The dual neurotransmitter co-regulation autonomous driving safety control method of claim 1, wherein, The fear emotion stimulus value is obtained by quantifying the collision risk and environmental uncertainty based on the world model, including: The environmental state and the ego vehicle action and the surrounding vehicle action generated by the hybrid policy network are input into the world model to predict the collision risk and environmental uncertainty in the vehicle interaction process, and the predicted value is taken as the fear emotion stimulus value of the ego vehicle; The world model is: ； in, In the world model, the first An independently trained diagonal Gaussian model, This represents the total number of independently trained diagonal Gaussian models in the world model. Indicates the environmental state. Indicates the action of the vehicle. Represents a state-action pair. express right The predicted average collision risk during vehicle interaction. express right The predicted standard deviation of environmental uncertainty during vehicle interaction. Represents a normal distribution; The fear emotion stimulus value of the ego vehicle is: ； wherein, represents a fear emotion stimulus value of the ego vehicle, represents a trade-off coefficient for measuring the proportion of collision risk and environmental uncertainty in the fear emotion stimulus value, represents a collision risk in the vehicle interaction process, , represents an environmental uncertainty in the vehicle interaction process, .

7. The dual neurotransmitter co-regulated autonomous driving safety control method of claim 1, wherein, The dynamic adjustment formula of the concentrations of norepinephrine and serotonin is: ； wherein, , respectively represent , a concentration of norepinephrine at a time point, , respectively represent , a concentration of 5-hydroxytryptamine at a time point, , respectively represent a weight for controlling a change in a concentration of norepinephrine and 5-hydroxytryptamine, , respectively represent a weight for controlling a release of norepinephrine and 5-hydroxytryptamine by a fear emotion stimulus, represents a fear emotion stimulus value of the ego vehicle, represents an environment state, represents an ego vehicle action; The switching formula of the stress response mechanism and the long-term planning mechanism is: ； wherein, represents a switching value, in the range of represents triggering the stress reaction mechanism, in the range of represents triggering the long-term planning mechanism, represents an S-shaped function for constraining to the range of , represent linear combination weights for adjusting the degree of influence on the concentration of norepinephrine, serotonin, respectively.​ 8. The dual neurotransmitter co-regulated autonomous driving safety control method of claim 1, wherein, The stress response mechanism adopts a deep deterministic policy gradient algorithm combined with an action safety layer to forcibly generate instantaneous safe actions including: The action parameters of the original action are obtained by parameterizing the original action output by the actor network of the deep deterministic policy gradient algorithm through the action safety layer, and based on the action parameters, a safety reachable set is constructed; The safety reachable set and the dangerous area are controlled to have no intersection by a mixed integer nonlinear programming model, and the safety reachable set and the dangerous area are converted into polynomial constraints on the action parameters to construct a mixed integer optimization problem, and the original action is projected to an instantaneous safe action satisfying the polynomial constraints; The parameterization processing is: ； in, This represents the original action output by the performer network in the deep deterministic policy gradient algorithm. Indicates used for Perform parameterized generator matrix, Indicates used to Convert the action parameters into parameter form. This represents the initial state used to construct a safe and reachable set; The safety reachable set is: ； in, express The corresponding secure reachable set, express The One portion, express Dimensions The first term used to describe the nonlinear effect of an action on a state is... Generator matrices corresponding to polynomials This represents the total number of polynomials used to describe the nonlinear effect of an action on a state. The first parameter in the noise parameter vector represents the second parameter. One portion, express The corresponding generator matrix, The dimension of the noise parameter vector; The dangerous area is: ； wherein, represents a dangerous area, , respectively represent a lateral coordinate and a longitudinal coordinate of the host vehicle, , respectively represent a lateral coordinate and a longitudinal coordinate of the center of the surrounding vehicle, represents a safety radius; The polynomial constraint on the action parameter is: ； wherein, represents a polynomial constraint on the action parameters, represents an index parameter related to the polynomial corresponding to the non-linear impact of the action on the state, represents an action parameter constraint threshold.

9. The dual neurotransmitter co-regulated autonomous driving safety control method of claim 1, wherein, The long-term planning mechanism generates an optimal action with controllable global risk based on a safety constraint reinforcement learning framework by trust domain policy optimization including: Based on the safety constraint reinforcement learning framework, the cumulative expected fear cost constraint is converted into an unconstrained optimization dual problem by the Lagrange multiplier method, and the update amplitude of the policy neural network is constrained based on the KL divergence to generate an optimal action with controllable global risk; The cumulative expected fear cost constraint is: ； wherein, represents an environment state, represents a self-vehicle action, represents a state-action pair, , respectively represent corresponding reward function, fear cost function, i.e. in the future accumulated expected reward, accumulated expected fear cost, within a time step, represents an action probability distribution, selected under , represents corresponding expectation, represents generated trajectory corresponding state distribution, represents corresponding expectation, represents a parameterized policy distribution, selected under , represents difference from , represents KL divergence, represents safety constraint threshold, represents trust region constraint threshold; Introducing Lagrange multipliers and convert the cumulative expected fear cost constraint into an unconstrained optimization dual problem: ； The KL divergence constraint formula is: ； wherein, , respectively represent the parameters of the policy neural network after and before update, , respectively represent , the policy corresponding to the selection of , represent the difference between , represent the optimal action probability distribution of the selection of , represent the state distribution corresponding to the trajectory generated by , represent the corresponding expectation of , the corresponding expectation of​​​ 10. The dual neurotransmitter co-regulated autonomous driving safety control method of claim 1, wherein, The network parameters of the stress response mechanism and the long-term planning mechanism are updated according to the environmental feedback to complete the closed-loop safety control including: The environmental feedback is stored in the experience pool, wherein the environmental feedback includes the environmental state at the current time, the ego vehicle action at the current time, the reward at the current time, and the environmental state at the next time. For the stress response mechanism, the network parameters of the critic network of the deep deterministic policy gradient algorithm are updated by minimizing the time difference error, and the network parameters of the actor network of the deep deterministic policy gradient algorithm are updated in combination with the action safety layer; For the long-term planning mechanism, the network parameters of the reward function and the fear cost function are updated by gradient descent; The loss function for updating the network parameters of the critic network of the deep deterministic policy gradient algorithm is: ； wherein represents a loss of the critic network of the deep deterministic policy gradient algorithm, , , , respectively represent an environment state at time t, an ego vehicle action at time t, a reward at time t, an environment state at time t, represents an experience pool, represents a corresponding expectation, a corresponding expectation, represents a network parameter of the critic network of the deep deterministic policy gradient algorithm, represents a value estimate of the critic network of the deep deterministic policy gradient algorithm for and , represents a target network parameter of the critic network of the deep deterministic policy gradient algorithm, represents an action generated by the actor network of the deep deterministic policy gradient algorithm for , represents a value estimate of the critic network of the deep deterministic policy gradient algorithm for and , represents a discount factor; The network parameters of the actor network of the deep deterministic policy gradient algorithm are updated as follows: ； wherein denotes a network parameter of the actor network of the deep deterministic policy gradient algorithm, denotes a learning rate of the actor network of the deep deterministic policy gradient algorithm, denotes a gradient of the objective function denotes the corresponding expectation, denotes a gradient of the action generated by the actor network of the deep deterministic policy gradient algorithm under denotes a gradient of the value estimates and by the actor network of the deep deterministic policy gradient algorithm under ;​​ The loss function for updating the network parameters of the reward function and the fear cost function is: ； wherein, , respectively represent the network parameters of the update reward function, the fear cost function , the loss of , respectively represent the reward value estimate, the fear cost value estimate of , by the reward function, the fear cost function , , respectively represent the reward target value, the fear cost target value.

Images