Automatic driving environment sensing method and system based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based automatic driving environment perception method and system. The method comprises the steps of multi-source data acquisition, original data optimization, environment perception model construction, decision optimization algorithm design and automatic driving decision making. The invention relates to the technical field of automatic driving environment perception, in particular to an automatic driving environment perception method and system based on artificial intelligence, and the method comprises the steps: obtaining original data through multi-source data collection; an original data optimization method of sensor calibration, dynamic and static data separation, data normalization, data enhancement and data set segmentation is adopted; a double-branch transformer model is adopted as an environment sensing model, and an irregular behavior pattern of a moving target and geometric topology constraints of a road structure are captured through a double-branch architecture; an improved ant colony optimization algorithm is adopted as a decision optimization algorithm, a multi-dimensional risk field is converted into a computable optimization target, and an optimal decision is realized through elite ant colony global exploration and non-elite local optimization.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving environmental perception technology, specifically to an autonomous driving environmental perception method and system based on artificial intelligence. Background Technology

[0002] Autonomous driving environmental perception refers to the technical process of collecting real-time data on the dynamic and static environment around a vehicle using multiple sensors such as LiDAR, cameras, and millimeter-wave radar, and then using algorithms such as deep learning and multi-source data fusion to perform structured analysis and semantic understanding of the environmental information, thereby constructing a three-dimensional dynamic environment model that the vehicle can recognize. By accurately capturing multi-dimensional information in complex traffic scenarios, it provides key basis for vehicle decision-making and planning, which can significantly reduce the traffic accident rate caused by human error and improve the safety of the traffic system.

[0003] However, traditional autonomous driving environmental perception methods suffer from several technical problems. They lack the ability to extract features from irregular motion patterns, lack topological correlation in static environment modeling, making it difficult to adapt to roads with varying curvature. Furthermore, the interaction between dynamic and static elements is not explicitly modeled, resulting in blind spots in risk field assessment. Traditional autonomous driving environmental perception methods also suffer from several technical problems. During autonomous driving decision-making, they fail to establish a coupled optimization model between the risk field and vehicle kinematics, which may lead to obstacle avoidance strategies violating dynamic constraints or causing occupant discomfort. In addition, they lack global situation field guidance and adaptive exploration mechanisms, severely limiting decision stability. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based method and system for autonomous driving environment perception. Traditional autonomous driving environment perception methods suffer from insufficient feature extraction capabilities for irregular motion patterns, lack of topological correlation in static environment modeling, difficulty adapting to roads with varying curvature, and the absence of explicit modeling of the interaction between dynamic and static elements, resulting in blind spots in risk field assessment. This solution creatively employs a dual-branch transformer model as the environment perception model. Through a dual-branch architecture, it captures the irregular behavior patterns of moving targets and the geometric topological constraints of the road structure, achieving feature complementarity between dynamic risks and static constraints. Simultaneously, superpixels are introduced. Graph convolution parsing reveals the semantic continuity of roads, providing a precise environmental cognitive foundation for subsequent decision-making. Addressing the technical issues of traditional autonomous driving environmental perception methods, such as the lack of a coupled optimization model between the risk field and vehicle kinematics during autonomous driving decision-making, which may lead to obstacle avoidance strategies violating dynamic constraints or causing occupant discomfort, and the lack of global situation field guidance and adaptive exploration mechanisms severely limiting decision stability, this solution creatively adopts an improved ant colony optimization algorithm as the decision optimization algorithm. This transforms the multidimensional risk field output by environmental perception into a computable optimization objective. Through the synergistic effect of global exploration by elite ant colonies and local optimization by non-elite ant colonies, optimal decision-making is achieved while satisfying vehicle dynamic constraints.

[0005] The technical solution adopted by this invention is as follows: The autonomous driving environment perception method based on artificial intelligence provided by this invention includes the following steps:

[0006] Step S1: Multi-source data acquisition;

[0007] Step S2: Optimize the raw data;

[0008] Step S3: Constructing the environmental perception model;

[0009] Step S4: Design of decision optimization algorithm;

[0010] Step S5: Autonomous driving decision.

[0011] Further, in step S1, the multi-source data acquisition is used to collect the raw data required for autonomous driving environmental perception. Specifically, it involves data acquisition through multiple types of sensors to obtain a raw environmental perception dataset. The raw environmental perception dataset specifically includes a past perception raw dataset and a current perception raw dataset. Both the past perception raw dataset and the current perception raw dataset include vehicle data, dynamic environment data, and static environment data. The past perception raw dataset also includes past dynamic obstacle threat scores and past static area passage scores.

[0012] Further, in step S2, the raw data optimization is used to optimize the collected environmental perception raw data, specifically including the following steps:

[0013] Step S21: Sensor calibration, used to unify the coordinate system and timestamps of multiple sensors. Specifically, it involves converting data from different sensors to the vehicle coordinate system and aligning the timestamps using an interpolation method to obtain spatiotemporally aligned data.

[0014] Step S22: Separate dynamic and static data to distinguish moving targets and static backgrounds in the environment. Specifically, this involves separating dynamic and static environmental data using optical flow.

[0015] Step S23: Data normalization, used to eliminate the dimensional differences between data from different sensors, specifically by using the min-max normalization method to transform the data to a uniform range;

[0016] Step S24: Data augmentation, used to increase the diversity of training data, specifically by performing random transformations on the original data to obtain augmented data. The random transformations include rotation, translation, scaling, and adding noise that conforms to physical laws.

[0017] Step S25: Dataset segmentation, specifically, the original past perception dataset is segmented to obtain a perception training set and a perception test set;

[0018] The current raw sensing dataset is optimized through sensor calibration, dynamic and static data separation, and data normalization to obtain a dataset to be processed. The past raw sensing dataset is optimized through sensor calibration, dynamic and static data separation, data normalization, data augmentation, and dataset segmentation to obtain a sensing training set and a sensing test set.

[0019] Further, in step S3, the environmental perception model construction is used to construct the model required for autonomous driving environmental perception, specifically to construct a dual-branch transformer model as the environmental perception model. The dual-branch transformer model specifically includes a dynamic environmental branch, a static environmental branch, a feature fusion module, and a stacked transformer module.

[0020] The construction of the environmental perception model specifically includes the following steps:

[0021] Step S31: Dynamic environment branch design, used to capture the spatiotemporal features of moving targets around the vehicle, including the following steps:

[0022] Step S311: Three-dimensional deformable convolution is used to extract irregular motion features of dynamic targets. Specifically, features are extracted adaptively in the time and space dimensions through a three-dimensional convolution kernel with learnable offsets to obtain preliminary dynamic spatiotemporal features.

[0023] Step S312: Constructing a time coherent unit to model the motion consistency of a dynamic target over continuous time. Specifically, by using gated cyclic residual connections and historical state decay factors, the long-term dependencies of the dynamic target are captured to obtain time-enhanced dynamic spatiotemporal features.

[0024] Step S313: Spatial coherence mask design, used to enhance the spatial structure features of dynamic targets, specifically by weighted focusing of dynamic features through a spatial attenuation mask based on the target center to obtain the final dynamic spatiotemporal features;

[0025] Step S32: Static environment branch design, used to analyze the relationship between road structure and spatial topology, includes the following steps:

[0026] Step S321: Oriented response convolution, used to extract orientation-sensitive features in the static environment. Specifically, it involves convolving the input static environment data with convolution kernels in multiple preset directions to obtain preliminary static spatial features.

[0027] Step S322: Constructing superpixel graph convolutional units to abstract from pixel-level features to region-level topology. Specifically, the input static environment data is segmented into superpixels using a simple linear iterative clustering algorithm. Each superpixel region is processed by the directional response convolution to obtain the preliminary static spatial features of each superpixel region. The superpixel region is used as a node, and the preliminary static spatial features of the superpixel region are used as node features. The edges between nodes represent the adjacency relationship between superpixel regions. Graph structure data is constructed, and graph convolution operation is performed to obtain the superpixel static spatial features.

[0028] Step S323: Road curvature encoding, used to embed road geometry information into features, specifically by calculating road curvature and combining it with B-spline interpolation, fusing curvature information with superpixel static spatial features to obtain the final static spatial features;

[0029] Step S33: Feature fusion module design, used to perform multi-scale adaptive fusion of features output from dynamic environment branch and static environment branch. Specifically, through learnable gating mechanism and channel attention weight, the final dynamic spatiotemporal features and the final static spatial features are dynamically weighted and fused at different scales to obtain dynamic-static fused features.

[0030] Step S34: Stacked transformer module design, used to model long-distance spatiotemporal dependencies in the environment, specifically by stacking multiple transformer layers to process dynamic-static fusion features. The stacked transformer module contains three output layers, which output the predicted trajectory, dynamic obstacle threat score and static area access score, respectively.

[0031] Step S35: Construct and train the model. Specifically, the dual-branch transformer model is constructed through the dynamic environment branch design, the static environment branch design, the feature fusion module design, and the stacked transformer module design. The model is trained based on the perception training set, and the model performance is verified based on the perception test set to obtain the dual-branch transformer model as the environment perception model.

[0032] Furthermore, in step S4, the decision optimization algorithm design is used to design an optimization algorithm required for optimizing autonomous driving decisions, specifically by adopting an improved ant colony optimization algorithm as the decision optimization algorithm;

[0033] The decision optimization algorithm design specifically includes the following steps:

[0034] Step S41: State space construction, used to quantify environmental risks and constraints, includes the following steps:

[0035] Step S411: Threat potential field construction, used to quantify the safety threats to the vehicle from dynamic and static environments. Specifically, the threat potential field is obtained by converting the predicted trajectory, dynamic obstacle threat score and static area passage score output by the environmental perception model into a dynamic potential field function.

[0036] Step S412: Lane constraint construction, used to express the driving constraints of the road structure on the vehicle, specifically by calculating the static potential field based on the vehicle's distance from the lane centerline and lane boundary to obtain the constraint potential field;

[0037] Step S413: State space definition, specifically, constructing a state vector that makes up the state space, whose elements include vehicle position data, vehicle speed data, vehicle acceleration data, vehicle heading angle data, estimated travel time, threat potential field and constraint potential field, constructing an ant cell set, the search path set of the ant cell set in the state space, as the autonomous driving decision set, and the optimal search path as the optimal autonomous driving decision.

[0038] Step S42: Design the objective function, which is used to design the path objective function required to optimize autonomous driving decisions;

[0039] Step S43: Design an ant colony state transition mechanism to achieve global search path optimization. The steps include:

[0040] Step S431: Pheromone field initialization, used to initialize the pheromone distribution for ant colony search, specifically by setting the initial pheromone concentration according to the risk cost to obtain the initial pheromone field for the guiding ant unit search;

[0041] Step S432: Elite ant global search, used to perform global optimal search path exploration, specifically by combining the transition probability formula of pheromone concentration, potential field and kinematic energy constraint to guide the movement of elite ant units. The elite ant units are specifically the ant units whose search path objective function value is the top 30% in ascending order.

[0042] Step S44: Dual-population collaborative optimization, used for collaborative optimization between elite ant units and non-elite ant units, includes the following steps:

[0043] Step S441: Dynamic update of the pheromone field, used to update the pheromone field according to the search situation to provide feedback for the next search. Specifically, the pheromone field is updated based on the exponential decay of the path objective function value of the elite ant unit to obtain the updated pheromone field.

[0044] Step S442: Local optimization of non-elite ant units, used to perform fine search near the search path of elite ant units, specifically by applying a time-decreasing Gaussian perturbation to the search path of elite ant units to generate new candidate paths.

[0045] Step S45: Obtain the optimal search path. Specifically, this involves designing an improved ant colony optimization algorithm through the state space construction, the design objective function, the design ant colony state transition mechanism, and the dual-population collaborative optimization. The algorithm is continuously searched until the search stopping condition is met, and the optimal search path is obtained as the optimal autonomous driving decision. The search stopping condition specifically includes reaching the maximum number of search rounds and the path objective function value being less than a preset threshold.

[0046] Further, in step S5, the autonomous driving decision specifically involves using the dataset to be processed as input to the environmental perception model to obtain the predicted trajectory, dynamic obstacle threat score, and static area passage score, and then using the decision optimization algorithm to optimize the autonomous driving decision based on the predicted trajectory, the dynamic obstacle threat score, and the static area passage score, to obtain the optimal autonomous driving decision and execute it.

[0047] The autonomous driving environment perception system based on artificial intelligence provided by this invention includes a multi-source data acquisition module, a raw data optimization module, an environment perception model construction module, a decision optimization algorithm design module, and an autonomous driving decision module;

[0048] The multi-source data acquisition module is used to collect raw data, obtain an environmental perception raw dataset by collecting raw data, and send the environmental perception raw dataset to the raw data optimization module.

[0049] The raw data optimization module is used to optimize raw data. Through raw data optimization, a dataset to be processed, a perception training set, and a perception test set are obtained. The dataset to be processed is sent to the autonomous driving decision module, and the perception training set and the perception test set are sent to the environmental perception model construction module.

[0050] The environmental perception model construction module is used to construct an environmental perception model. It constructs a dual-branch transformer model as the environmental perception model and sends the environmental perception model to the autonomous driving decision module.

[0051] The decision optimization algorithm design module is used to design decision optimization algorithms. It adopts an improved ant colony optimization algorithm as the decision optimization algorithm and sends the decision optimization algorithm to the autonomous driving decision module.

[0052] The autonomous driving decision module is used for autonomous driving decision-making. It uses the environmental perception model to process data in real time to obtain predicted trajectory, dynamic obstacle threat score and static area passage score, and combines the decision optimization algorithm to optimize the autonomous driving decision in real time to obtain the optimal autonomous driving decision and execute it.

[0053] The beneficial effects achieved by the present invention using the above solution are as follows:

[0054] (1) In view of the technical problems of traditional autonomous driving environment perception methods, such as insufficient feature extraction capability for irregular motion patterns, lack of topological correlation in static environment modeling, difficulty in adapting to roads with curvature changes, and lack of explicit modeling of the interaction relationship between dynamic and static elements, resulting in blind spots in risk field assessment, this solution creatively adopts a dual-branch transformer model as the environment perception model. Through the dual-branch architecture, it captures the irregular behavior patterns of moving targets and the geometric topological constraints of road structures respectively, realizing feature complementarity between dynamic risks and static constraints. At the same time, it introduces superpixel map convolution to parse the semantic continuity of roads, providing an accurate environmental cognition basis for subsequent decision-making.

[0055] (2) In view of the technical problems of traditional autonomous driving environment perception methods, when making autonomous driving decisions, there is no coupled optimization model of risk field and vehicle kinematics, which may lead to obstacle avoidance strategies violating dynamic constraints or causing passenger discomfort. In addition, there is a lack of global situation field guidance and adaptive exploration mechanism, which severely limits the stability of decision-making. This solution creatively adopts an improved ant colony optimization algorithm as the decision optimization algorithm, which transforms the multidimensional risk field output by environmental perception into a computable optimization objective. Through the synergy of elite ant colony global exploration and non-elite local optimization, the optimal decision is achieved under the condition of satisfying vehicle dynamic constraints. Attached Figure Description

[0056] Figure 1A flowchart illustrating the artificial intelligence-based autonomous driving environment perception method provided by the present invention;

[0057] Figure 2 A schematic diagram of the modules of the artificial intelligence-based autonomous driving environment perception system provided by the present invention;

[0058] Figure 3 A flowchart illustrating the optimization of the raw data in step S2;

[0059] Figure 4 A flowchart illustrating the process of building the environmental perception model in step S3;

[0060] Figure 5 A flowchart illustrating the decision optimization algorithm design for step S4.

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0064] Example 1, see Figure 1 The technical solution adopted by this invention is as follows: The autonomous driving environment perception method based on artificial intelligence provided by this invention includes the following steps:

[0065] Step S1: Multi-source data acquisition;

[0066] Step S2: Optimize the raw data;

[0067] Step S3: Constructing the environmental perception model;

[0068] Step S4: Design of decision optimization algorithm;

[0069] Step S5: Autonomous driving decision.

[0070] Example 2, see Figure 1 and Figure 2 In step S1, the multi-source data acquisition is used to collect the raw data required for autonomous driving environmental perception. Specifically, it involves data acquisition through multiple types of sensors to obtain a raw environmental perception dataset. The raw environmental perception dataset specifically includes a past perception raw dataset and a current perception raw dataset. Both the past perception raw dataset and the current perception raw dataset include vehicle data, dynamic environment data, and static environment data. The past perception raw dataset also includes past dynamic obstacle threat scores and past static area passage scores. The vehicle data specifically includes vehicle mass data, vehicle position data, vehicle speed data, vehicle acceleration data, and vehicle heading angle data. The dynamic environment data specifically includes moving target detection data, including image data of surrounding vehicles, pedestrians, and animals, and retaining their position data, speed data, and acceleration data. The static environment data specifically includes curb geometric parameters, lane curvature, static obstacle data, road structure data, and traffic sign image data.

[0071] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the raw data optimization is used to optimize the collected environmental perception raw data, specifically including the following steps:

[0072] Step S21: Sensor calibration, used to unify the coordinate system and timestamps of multiple sensors. Specifically, it involves converting data from different sensors to the vehicle coordinate system and aligning the timestamps using an interpolation method to obtain spatiotemporally aligned data.

[0073] Step S22: Separate dynamic and static data to distinguish moving targets and static backgrounds in the environment. Specifically, this involves separating dynamic and static environmental data using optical flow.

[0074] Step S23: Data normalization, used to eliminate the dimensional differences between data from different sensors, specifically by using the min-max normalization method to transform the data to a uniform range;

[0075] Step S24: Data augmentation, used to increase the diversity of training data, specifically by performing random transformations on the original data to obtain augmented data. The random transformations include rotation, translation, scaling, and adding noise that conforms to physical laws.

[0076] Step S25: Dataset segmentation, specifically, the original past perception dataset is segmented to obtain a perception training set and a perception test set;

[0077] The current raw sensing dataset is optimized through sensor calibration, dynamic and static data separation, and data normalization to obtain a dataset to be processed. The past raw sensing dataset is optimized through sensor calibration, dynamic and static data separation, data normalization, data augmentation, and dataset segmentation to obtain a sensing training set and a sensing test set.

[0078] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the environmental perception model is constructed to construct the model required for autonomous driving environmental perception. Specifically, a dual-branch transformer model is constructed as the environmental perception model. The dual-branch transformer model specifically includes a dynamic environmental branch, a static environmental branch, a feature fusion module, and a stacked transformer module.

[0079] The construction of the environmental perception model specifically includes the following steps:

[0080] Step S31: Dynamic environment branch design, used to capture the spatiotemporal features of moving targets around the vehicle, including the following steps:

[0081] Step S311: Three-dimensional deformable convolution is used to extract irregular motion features of dynamic targets. Specifically, it uses a three-dimensional convolution kernel with learnable offsets to adaptively extract features in the temporal and spatial dimensions to obtain preliminary dynamic spatiotemporal features. The formula used is as follows:

[0082] ;

[0083] In the formula, Indicates preliminary dynamic spatiotemporal characteristics, Represents a three-dimensional deformable convolution function. This indicates the input of dynamic environment data. Indicates the learnable offset. This indicates a learnable modulated scalar;

[0084] Step S312: Construction of temporal coherence units, used to model the motion consistency of dynamic targets over continuous time. Specifically, by using gated cyclic residual connections and historical state decay factors, the long-term dependencies of dynamic targets are captured to obtain time-enhanced dynamic spatiotemporal features. The formula used is as follows:

[0085] ;

[0086] In the formula, This indicates the initial hidden state of the temporal coherent cell. This represents the initial dynamic spatiotemporal characteristics of time step t. Represents the historical state decay factor. This represents the sigmoid function. Represents the attenuation factor transformation matrix. This represents the initial hidden state of the temporal coherent cell at time step t-1. This represents the initial hidden state of the temporal coherence cell at time step t. Gated transformation matrix, Represents the hyperbolic tangent function. Represents the unit input transformation matrix. Indicates the cell input bias term. Represents tensor product operation;

[0087] Step S313: Spatial coherence mask design, used to enhance the spatial structural features of dynamic targets. Specifically, dynamic features are weighted and focused using a spatial attenuation mask based on the target center to obtain the final dynamic spatiotemporal features. The formula used is as follows:

[0088] ;

[0089] In the formula, Cm represents the spatial attenuation mask based on the target center, and le represents the learnable distance weight parameter. This represents the Euclidean distance from the vehicle pixel to the target center pixel. Indicates the final dynamic spatiotemporal characteristics, This indicates enhanced dynamic spatiotemporal characteristics of time. This represents the bilinear interpolation alignment operation function. This represents element-wise multiplication.

[0090] Step S32: Static environment branch design, used to analyze the relationship between road structure and spatial topology, includes the following steps:

[0091] Step S321: Oriented response convolution, used to extract orientation-sensitive features in the static environment. Specifically, it involves convolving the input static environment data with multiple pre-defined convolutional kernels to obtain preliminary static spatial features. The formula used is as follows:

[0092] ;

[0093] In the formula, Indicates the direction angle is Preliminary static spatial characteristics, The convolution kernel orientation angle is... directional convolution function, This indicates the input of static environment data. Indicates the kernel orientation angle. Indicates preliminary static spatial characteristics;

[0094] Step S322: Constructing superpixel graph convolutional units to abstract from pixel-level features to region-level topology. Specifically, the input static environment data is segmented into superpixels using a simple linear iterative clustering algorithm. Each superpixel region is processed by the directional response convolution to obtain the preliminary static spatial features of each superpixel region. The superpixel region is used as a node, and the preliminary static spatial features of the superpixel region are used as node features. The edges between nodes represent the adjacency relationship between superpixel regions. Graph structure data is constructed, and graph convolution operation is performed to obtain the superpixel static spatial features.

[0095] Step S323: Road curvature encoding, used to embed road geometry information into features. Specifically, by calculating the road curvature and combining it with B-spline interpolation, the curvature information is fused with the superpixel static spatial features to obtain the final static spatial features. The formula used is as follows:

[0096] ;

[0097] In the formula, Represents the final static spatial characteristics. This represents the function that runs the multilayer perceptron. Represents the static spatial features of superpixels. This represents the B-spline interpolation function. Indicates the curvature of the road;

[0098] Step S33: Feature fusion module design, used to perform multi-scale adaptive fusion of features output from dynamic and static environment branches. Specifically, through a learnable gating mechanism and channel attention weights, the final dynamic spatiotemporal features and the final static spatial features are dynamically weighted and fused at different scales to obtain dynamic-static fused features. The formula used is as follows:

[0099] ;

[0100] In the formula, Denotes the fusion gating weights at the k-th scale. Denotes the fully connected layer function at the k-th scale. This represents the final dynamic spatiotemporal characteristics at the k-th scale. Let represent the final static spatial features at the k-th scale, and Ch represent the channel attention weights. This represents the softmax function. Represents the channel attention transformation matrix. This represents the global average pooling function. Represents the dynamic-static fusion features at the k-th scale. This represents the dynamic-static fusion feature, where K represents the total number of scales;

[0101] Step S34: Stacked transformer module design, used to model long-distance spatiotemporal dependencies in the environment, specifically by stacking multiple transformer layers to process dynamic-static fusion features. The stacked transformer module contains three output layers, which output the predicted trajectory, dynamic obstacle threat score and static area access score, respectively.

[0102] Step S35: Construct and train the model. Specifically, the dual-branch transformer model is constructed through the dynamic environment branch design, the static environment branch design, the feature fusion module design, and the stacked transformer module design. The model is trained based on the perception training set, and the model performance is verified based on the perception test set to obtain the dual-branch transformer model as the environment perception model.

[0103] By performing the above operations, this solution addresses the technical problems of traditional autonomous driving environmental perception methods, such as insufficient feature extraction capabilities for irregular motion patterns, lack of topological correlation in static environment modeling, difficulty in adapting to roads with varying curvature, and lack of explicit modeling of the interaction between dynamic and static elements, resulting in blind spots in risk field assessment. This solution creatively adopts a dual-branch transformer model as the environmental perception model. Through the dual-branch architecture, it captures the irregular behavior patterns of moving targets and the geometric topological constraints of road structures, achieving feature complementarity between dynamic risks and static constraints. At the same time, it introduces superpixel map convolution to parse the semantic continuity of roads, providing an accurate environmental cognition foundation for subsequent decision-making.

[0104] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the decision optimization algorithm design is used to design the optimization algorithm required for optimizing autonomous driving decisions. Specifically, an improved ant colony optimization algorithm is adopted as the decision optimization algorithm.

[0105] The decision optimization algorithm design specifically includes the following steps:

[0106] Step S41: State space construction, used to quantify environmental risks and constraints, includes the following steps:

[0107] Step S411: Threat potential field construction, used to quantify the safety threats to the vehicle from dynamic and static environments. Specifically, this involves converting the predicted trajectory, dynamic obstacle threat score, and static area passage score output by the environmental perception model into a dynamic potential field function to obtain the threat potential field. The formula used is as follows:

[0108] ;

[0109] In the formula, Tf represents the threat potential field. Indicates dynamic threat weights. This represents the static threat weight, where Sa represents the sample size. This represents the dynamic obstacle threat score of the sa-th sample. Let represent the static region access score of the sa-th sample. Let represent the Euclidean distance from the vehicle to the predicted trajectory of the sa-th sample. This indicates a minimum value to prevent division by zero.

[0110] Step S412: Lane constraint construction, used to express the driving constraints of the road structure on vehicles. Specifically, it calculates the static potential field by the vehicle's distance from the lane centerline and lane boundary, and obtains the constraint potential field. The formula used is as follows:

[0111] ;

[0112] In the formula, Lf represents the constraint potential field. Indicates the weight of the central constraint. Indicates the boundary constraint weights. This represents the Euclidean distance from the vehicle to the center line of the lane. This represents the Euclidean distance from the vehicle to the lane boundary;

[0113] Step S413: State space definition, specifically, constructing a state vector that makes up the state space, whose elements include vehicle position data, vehicle speed data, vehicle acceleration data, vehicle heading angle data, estimated travel time, threat potential field and constraint potential field, constructing an ant cell set, the search path set of the ant cell set in the state space, as the autonomous driving decision set, and the optimal search path as the optimal autonomous driving decision.

[0114] Step S42: Design the objective function. The objective function used to design the path objective function required for optimizing autonomous driving decisions is as follows:

[0115] ;

[0116] In the formula, Fun represents the path objective function value. Indicates the time cost weight. Indicates the threat cost weight. This indicates the weight of the comfort cost. This represents the time cost, specifically the estimated travel time. This represents the risk cost, along the path of the threat field. The points, This represents the vehicle's acceleration data. The comfort cost is represented as the integral of the square of the rate of change of acceleration;

[0117] Step S43: Design an ant colony state transition mechanism to achieve global search path optimization. The steps include:

[0118] Step S431: Pheromone field initialization, used to initialize the pheromone distribution for ant colony search. Specifically, it involves setting the initial pheromone concentration based on the risk cost to obtain the initial pheromone field for the guide ant unit search. The formula used is as follows:

[0119] ;

[0120] In the formula, This represents the initial pheromone concentration from node i to node j. Indicates the initial pheromone risk weight. Indicates the initial pheromone base weights. Let $\frac{i}{j}$ represent the average risk cost between node $i$ and node $j$. Indicates the basic pheromone concentration;

[0121] Step S432: Elite ant global search, used to perform global optimal search path exploration. Specifically, by combining the transition probability formula of pheromone concentration, potential field and kinematic energy constraints, the elite ant units are guided to move. The elite ant units are specifically the ant units whose search path objective function value is in ascending order. The transition probability formula is expressed as follows:

[0122] ;

[0123] In the formula, Let represent the transition probability of an elite ant unit from node i to node j. This represents the pheromone concentration from node i to node j. Let represent the mean of the threat potential fields of nodes i and j. Let represent the mean of the constraint potential fields at nodes i and j. This represents the change in kinetic energy from node i to node j. This represents the pheromone concentration from node i to node l. Let represent the mean of the threat potential fields of node i and node l. Let represent the mean of the constraint potential fields between node i and node l. This represents the change in kinetic energy from node i to node l. Let i represent the set of neighboring nodes. Indicates the pheromone index weight. Indicates the potential field index weight. Indicates the weight of the kinematic energy index. This indicates the vehicle's weight data. This represents the vehicle speed data of node j. This represents the vehicle speed data at node i;

[0124] Step S44: Dual-population collaborative optimization, used for collaborative optimization between elite ant units and non-elite ant units, includes the following steps:

[0125] Step S441: Dynamic update of the pheromone field, used to update the pheromone field according to the search situation and provide feedback for the next search. Specifically, the pheromone field is updated based on the exponential decay of the path objective function value of the elite ant unit, resulting in the updated pheromone field. The formula used is as follows:

[0126] ;

[0127] In the formula, This indicates the strength of the search path of the elite ant unit. This represents the path objective function value of the historical best search path. The path objective function value represents the search path of an elite ant unit. This represents the change in pheromone concentration from node i to node j. This represents the decay coefficient, and m represents the current search round. This represents the pheromone concentration from node i to node j during the (m+1)th search. Indicates the volatility coefficient. The pheromone concentration from node i to node j is represented by N, where N represents the number of elite ant units. This represents the change in pheromone concentration from node i to node j along the search path of the nth elite ant unit.

[0128] Step S442: Local optimization of non-elite ant units, used to perform fine search near the search path of elite ant units, specifically by applying a time-decreasing Gaussian perturbation to the search path of elite ant units to generate new candidate paths.

[0129] Step S45: Obtain the optimal search path. Specifically, this involves designing an improved ant colony optimization algorithm through the state space construction, the design objective function, the design ant colony state transition mechanism, and the dual-population collaborative optimization. The algorithm is continuously searched until the search stopping condition is met, and the optimal search path is obtained as the optimal autonomous driving decision. The search stopping condition specifically includes reaching the maximum number of search rounds and the path objective function value being less than a preset threshold.

[0130] By performing the above operations, this solution addresses the technical problems of traditional autonomous driving environmental perception methods, which lack a coupled optimization model between the risk field and vehicle kinematics during autonomous driving decision-making. This leads to obstacle avoidance strategies potentially violating dynamic constraints or causing occupant discomfort. Furthermore, the lack of a global situation field guidance and adaptive exploration mechanism severely limits decision-making stability. This solution creatively adopts an improved ant colony optimization algorithm as the decision optimization algorithm. It transforms the multidimensional risk field output by environmental perception into a computable optimization objective. Through the synergy of elite ant colony global exploration and non-elite local optimization, the optimal decision is achieved while satisfying vehicle dynamic constraints.

[0131] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S5, the autonomous driving decision specifically involves using the dataset to be processed as the input of the environmental perception model to obtain the predicted trajectory, dynamic obstacle threat score and static area passage score. Based on the predicted trajectory, the dynamic obstacle threat score and the static area passage score, the decision optimization algorithm is used to optimize the autonomous driving decision to obtain the optimal autonomous driving decision and execute it.

[0132] Example 7, see Figure 1 and Figure 2 This embodiment is based on the above embodiments. The artificial intelligence-based autonomous driving environment perception system provided by the present invention includes a multi-source data acquisition module, a raw data optimization module, an environment perception model construction module, a decision optimization algorithm design module, and an autonomous driving decision module.

[0133] The multi-source data acquisition module is used to collect raw data, obtain an environmental perception raw dataset by collecting raw data, and send the environmental perception raw dataset to the raw data optimization module.

[0134] The raw data optimization module is used to optimize raw data. Through raw data optimization, a dataset to be processed, a perception training set, and a perception test set are obtained. The dataset to be processed is sent to the autonomous driving decision module, and the perception training set and the perception test set are sent to the environmental perception model construction module.

[0135] The environmental perception model construction module is used to construct an environmental perception model. It constructs a dual-branch transformer model as the environmental perception model and sends the environmental perception model to the autonomous driving decision module.

[0136] The decision optimization algorithm design module is used to design decision optimization algorithms. It adopts an improved ant colony optimization algorithm as the decision optimization algorithm and sends the decision optimization algorithm to the autonomous driving decision module.

[0137] The autonomous driving decision module is used for autonomous driving decision-making. It uses the environmental perception model to process data in real time to obtain predicted trajectory, dynamic obstacle threat score and static area passage score, and combines the decision optimization algorithm to optimize the autonomous driving decision in real time to obtain the optimal autonomous driving decision and execute it.

[0138] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0140] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based method for environmental perception in autonomous driving, characterized by: The method includes the following steps: S1: Multi-source data acquisition, data is acquired through multiple types of sensors to obtain the environmental perception raw dataset, which specifically includes the past perception raw dataset and the current perception raw dataset; S2: Raw data optimization, which optimizes the collected raw data to obtain the dataset to be processed, the perception training set, and the perception test set; S3: Environmental perception model construction, used to construct the model required for autonomous driving environmental perception, specifically to construct a dual-branch transformer model as the environmental perception model. The dual-branch transformer model specifically includes a dynamic environmental branch, a static environmental branch, a feature fusion module, and a stacked transformer module. S4: Decision optimization algorithm design, used to design optimization algorithms required for autonomous driving decisions, specifically adopting an improved ant colony optimization algorithm as the decision optimization algorithm; S5: Autonomous driving decision-making, specifically, taking the dataset to be processed as the input of the environmental perception model to obtain the predicted trajectory, dynamic obstacle threat score and static area passage score, and using the decision optimization algorithm to optimize the autonomous driving decision based on the predicted trajectory, the dynamic obstacle threat score and the static area passage score, to obtain the optimal autonomous driving decision and execute it.

2. The artificial intelligence-based autonomous driving environment perception method according to claim 1, characterized in that: The construction of the environmental perception model specifically includes the following steps: Step S31: Dynamic environment branch design, used to capture the spatiotemporal features of moving targets around the vehicle, including the following steps: Step S311: Three-dimensional deformable convolution is used to extract irregular motion features of dynamic targets. Specifically, features are extracted adaptively in the time and space dimensions through a three-dimensional convolution kernel with learnable offsets to obtain preliminary dynamic spatiotemporal features. Step S312: Constructing a time coherent unit to model the motion consistency of a dynamic target over continuous time. Specifically, by using gated cyclic residual connections and historical state decay factors, the long-term dependencies of the dynamic target are captured to obtain time-enhanced dynamic spatiotemporal features. Step S313: Spatial coherence mask design, used to enhance the spatial structure features of dynamic targets, specifically by weighted focusing of dynamic features through a spatial attenuation mask based on the target center to obtain the final dynamic spatiotemporal features; Step S32: Static environment branch design, used to analyze the relationship between road structure and spatial topology, includes the following steps: Step S321: Oriented response convolution, used to extract orientation-sensitive features in the static environment. Specifically, it involves convolving the input static environment data with convolution kernels in multiple preset directions to obtain preliminary static spatial features. Step S322: Constructing superpixel graph convolutional units to abstract from pixel-level features to region-level topology. Specifically, the input static environment data is segmented into superpixels using a simple linear iterative clustering algorithm. Each superpixel region is processed by the directional response convolution to obtain the preliminary static spatial features of each superpixel region. The superpixel region is used as a node, and the preliminary static spatial features of the superpixel region are used as node features. The edges between nodes represent the adjacency relationship between superpixel regions. Graph structure data is constructed, and graph convolution operation is performed to obtain the superpixel static spatial features. Step S323: Road curvature encoding, used to embed road geometry information into features, specifically by calculating road curvature and combining it with B-spline interpolation, fusing curvature information with superpixel static spatial features to obtain the final static spatial features; Step S33: Feature fusion module design, used to perform multi-scale adaptive fusion of features output from dynamic environment branch and static environment branch. Specifically, through learnable gating mechanism and channel attention weight, the final dynamic spatiotemporal features and the final static spatial features are dynamically weighted and fused at different scales to obtain dynamic-static fused features. Step S34: Stacked transformer module design, used to model long-distance spatiotemporal dependencies in the environment, specifically by stacking multiple transformer layers to process dynamic-static fusion features. The stacked transformer module contains three output layers, which output the predicted trajectory, dynamic obstacle threat score and static area access score, respectively. Step S35: Construct and train the model. Specifically, the dual-branch transformer model is constructed through the dynamic environment branch design, the static environment branch design, the feature fusion module design, and the stacked transformer module design. The model is trained based on the perception training set, and the model performance is verified based on the perception test set to obtain the dual-branch transformer model as the environment perception model.

3. The artificial intelligence-based autonomous driving environment perception method according to claim 1, characterized in that: The decision optimization algorithm design specifically includes the following steps: Step S41: State space construction, used to quantify environmental risks and constraints, includes the following steps: Step S411: Threat potential field construction, used to quantify the safety threats to the vehicle from dynamic and static environments. Specifically, the threat potential field is obtained by converting the predicted trajectory, dynamic obstacle threat score and static area passage score output by the environmental perception model into a dynamic potential field function. Step S412: Lane constraint construction, used to express the driving constraints of the road structure on the vehicle, specifically by calculating the static potential field based on the vehicle's distance from the lane centerline and lane boundary to obtain the constraint potential field; Step S413: State space definition, specifically, constructing a state vector that makes up the state space, whose elements include vehicle position data, vehicle speed data, vehicle acceleration data, vehicle heading angle data, estimated travel time, threat potential field and constraint potential field, constructing an ant cell set, the search path set of the ant cell set in the state space, as the autonomous driving decision set, and the optimal search path as the optimal autonomous driving decision. Step S42: Design the objective function, which is used to design the path objective function required to optimize autonomous driving decisions; Step S43: Design an ant colony state transition mechanism to achieve global search path optimization. The steps include: Step S431: Pheromone field initialization, used to initialize the pheromone distribution for ant colony search, specifically by setting the initial pheromone concentration according to the risk cost to obtain the initial pheromone field for the guiding ant unit search; Step S432: Elite ant global search, used to perform global optimal search path exploration, specifically by combining the transition probability formula of pheromone concentration, potential field and kinematic energy constraint to guide the movement of elite ant units. The elite ant units are specifically the ant units whose search path objective function value is the top 30% in ascending order. Step S44: Dual-population collaborative optimization, used for collaborative optimization between elite ant units and non-elite ant units, includes the following steps: Step S441: Dynamic update of the pheromone field, used to update the pheromone field according to the search situation to provide feedback for the next search. Specifically, the pheromone field is updated based on the exponential decay of the path objective function value of the elite ant unit to obtain the updated pheromone field. Step S442: Local optimization of non-elite ant units, used to perform fine search near the search path of elite ant units, specifically by applying a time-decreasing Gaussian perturbation to the search path of elite ant units to generate new candidate paths. Step S45: Obtain the optimal search path. Specifically, this involves designing an improved ant colony optimization algorithm through the state space construction, the design objective function, the design ant colony state transition mechanism, and the dual-population collaborative optimization. The algorithm is continuously searched until the search stopping condition is met, and the optimal search path is obtained as the optimal autonomous driving decision. The search stopping condition specifically includes reaching the maximum number of search rounds and the path objective function value being less than a preset threshold.

4. The artificial intelligence-based autonomous driving environment perception method according to claim 1, characterized in that: Both the past perception raw dataset and the current perception raw dataset include vehicle data, dynamic environment data, and static environment data. The past perception raw dataset also includes past dynamic obstacle threat scores and past static area passage scores.

5. The artificial intelligence-based autonomous driving environment perception method according to claim 1, characterized in that: The optimization of the original data specifically includes the following steps: Step S21: Sensor calibration, used to unify the coordinate system and timestamps of multiple sensors. Specifically, it involves converting data from different sensors to the vehicle coordinate system and aligning the timestamps using an interpolation method to obtain spatiotemporally aligned data. Step S22: Separate dynamic and static data to distinguish moving targets and static backgrounds in the environment. Specifically, this involves separating dynamic and static environmental data using optical flow. Step S23: Data normalization, used to eliminate the dimensional differences between data from different sensors, specifically by using the min-max normalization method to transform the data to a uniform range; Step S24: Data augmentation, used to increase the diversity of training data, specifically by performing random transformations on the original data to obtain augmented data. The random transformations include rotation, translation, scaling, and adding noise that conforms to physical laws. Step S25: Dataset segmentation, specifically, the original past perception dataset is segmented to obtain a perception training set and a perception test set; The current raw sensing dataset is optimized through sensor calibration, dynamic and static data separation, and data normalization to obtain a dataset to be processed. The past raw sensing dataset is optimized through sensor calibration, dynamic and static data separation, data normalization, data augmentation, and dataset segmentation to obtain a sensing training set and a sensing test set.

6. An artificial intelligence-based autonomous driving environment perception system, used to implement the artificial intelligence-based autonomous driving environment perception method as described in any one of claims 1-5, characterized in that: It includes a multi-source data acquisition module, a raw data optimization module, an environmental perception model construction module, a decision optimization algorithm design module, and an autonomous driving decision module.

7. The artificial intelligence-based autonomous driving environment perception system according to claim 6, characterized in that: The multi-source data acquisition module is used to collect raw data, obtain an environmental perception raw dataset by collecting raw data, and send the environmental perception raw dataset to the raw data optimization module. The raw data optimization module is used to optimize raw data. Through raw data optimization, a dataset to be processed, a perception training set, and a perception test set are obtained. The dataset to be processed is sent to the autonomous driving decision module, and the perception training set and the perception test set are sent to the environmental perception model construction module. The environmental perception model construction module is used to construct an environmental perception model. It constructs a dual-branch transformer model as the environmental perception model and sends the environmental perception model to the autonomous driving decision module. The decision optimization algorithm design module is used to design decision optimization algorithms. It adopts an improved ant colony optimization algorithm as the decision optimization algorithm and sends the decision optimization algorithm to the autonomous driving decision module. The autonomous driving decision module is used for autonomous driving decision-making. It uses the environmental perception model to process data in real time to obtain predicted trajectory, dynamic obstacle threat score and static area passage score, and combines the decision optimization algorithm to optimize the autonomous driving decision in real time to obtain the optimal autonomous driving decision and execute it.

Images