An unmanned vehicle left turn decision method for urban intersections

Abstract

The application provides a left-turn decision-making method for an unmanned vehicle at an urban intersection, and belongs to the technical field of vehicle automatic driving control. The method comprises the following steps: acquiring self-vehicle state information and other-vehicle state information in a bird's-eye view; taking the self-vehicle state information as a query matrix, and taking the combination of the self-vehicle state information and the other-vehicle state information as a key matrix and a value matrix; inputting the key matrix and the value matrix into a multi-head cross attention network to calculate the attention weight of each state in the key matrix with respect to the self-vehicle state; screening target other-vehicle state information for left-turn decision-making from the other-vehicle state information according to the attention weight; fusing the target other-vehicle state information with the self-vehicle state information, and inputting the fused information into a decision-making model to generate an acceleration control instruction of the self-vehicle; determining the longitudinal behavior of the self-vehicle according to the instruction, and generating a lateral behavior through trajectory tracking to control the self-vehicle to complete left-turn traffic. The application significantly improves the traffic decision-making accuracy in a complex scene.

Description

Technical Field

[0001] This invention belongs to the field of vehicle autonomous driving control technology, and specifically relates to a method for left-turn decision-making of unmanned vehicles at urban intersections. Background Technology

[0002] With the rapid development of autonomous driving technology, the decision-making ability of driverless vehicles in complex traffic scenarios has become a research focus. Studies have found that most accidents occur at intersections, causing significant traffic delays in urban areas, especially in unprotected left-turn scenarios at urban intersections. This is because left turns require crossing the oncoming lane, directly clashing with oncoming straight-going vehicles, and necessitating a complex game of speed, distance, and intent among multiple lanes and vehicles within a limited time. Right turns, on the other hand, are usually unaffected by oncoming traffic and only require attention to non-motorized vehicles and pedestrians traveling in the same direction, making the decision-making difficulty and risk far lower than that of left turns.

[0003] Existing decision-making methods for unprotected left turns at intersections can be broadly categorized into two types: traditional decision-making and planning methods, and machine learning methods. Traditional methods are often limited to pre-designed scenarios, requiring continuous patching of decision rules as environments become more complex. However, real-world environments are difficult to exhaustively represent, hindering the practical application of autonomous driving. Machine learning methods, on the other hand, are insufficient in extracting and utilizing the key states of oncoming vehicles from upstream BEV perception information, and also suffer from poor interpretability.

[0004] Therefore, existing autonomous driving systems often suffer from insufficient traffic situation recognition and low traffic efficiency when handling left turns. These problems not only affect the traffic efficiency of driverless vehicles but also pose a potential threat to traffic safety. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a left-turn decision-making method for autonomous vehicles at urban intersections, which significantly improves the accuracy of traffic decisions in complex scenarios and provides a solution for the efficient passage of autonomous vehicles in complex traffic environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention proposes a left-turn decision-making method for autonomous vehicles at urban intersections, comprising the following steps: Obtain the status information of your own vehicle and other vehicles from a bird's-eye view; The vehicle's status information is used as the query matrix, and the combination of the vehicle's status information and other vehicles' status information is used as the key matrix and value matrix, respectively. These are then input into a multi-head cross-attention network to calculate the attention weights of the vehicle's status to each state in the key matrix. Based on the attention weight, the target other vehicle status information for left turn decision is filtered from the other vehicle status information; The selected target vehicle status information is fused with the vehicle's status information and input into the decision model to generate acceleration control commands for the vehicle. The longitudinal behavior of the vehicle is determined according to the acceleration control command, and the lateral behavior is generated through trajectory tracking to control the vehicle to complete the left turn.

[0007] Furthermore, the vehicle status information is represented as follows: ; in, This represents the x-coordinate of the vehicle in the global coordinate system; This represents the ordinate of the vehicle in the global coordinate system; Indicates the vehicle speed Quantity; Indicates the vehicle speed Quantity.

[0008] Furthermore, before obtaining the status information of other vehicles, one or more potential collision points are first determined based on the vehicle's preset left-turn trajectory and the driving trajectory of vehicles approaching from the opposite lane. The status information of other vehicles is represented as follows: ; ; ; in, This represents the x-coordinate of the vehicle in the global coordinate system; This represents the ordinate of the vehicle in the global coordinate system; Represents the x-coordinate of the corresponding collision point; Represents the ordinate of the corresponding collision point; Indicates the speed of his vehicle Quantity; Indicates the speed of his vehicle Quantity.

[0009] Furthermore, the multi-head cross-attention network is an improved multi-head cross-attention network, and its construction method includes: The vehicle status information is split into multiple feature dimensions and embedded separately, and the status information of other vehicles is also embedded. The dimensions of the embedding representation are determined according to the preset embedding dimensions. Through a multi-head attention mechanism, the attention weights of each feature dimension of the vehicle's state information to each feature dimension of the key matrix are independently calculated in each attention head, so as to obtain the weight coefficients of other vehicle state information relative to the vehicle's state information. The outputs of each attention head are concatenated and integrated through a linear transformation layer to obtain the comprehensive attention weights of each state information in the key matrix of each feature dimension of the vehicle state information.

[0010] Furthermore, based on the attention weights, target other vehicle status information for left-turn decisions is filtered from the other vehicle status information, specifically as follows: Obtain the attention weight matrix output by the multi-head cross-attention network, where each element in the attention weight matrix represents the attention weight of a feature dimension of the other vehicle's state information relative to a feature dimension of the own vehicle's state information. For each piece of information about another vehicle's status, a comprehensive weight value is calculated based on its corresponding attention weight. The comprehensive weight value is compared with a preset threshold. The status information of other vehicles with a comprehensive weight value greater than the preset threshold is retained as the target status information of other vehicles, while the status information of other vehicles with a comprehensive weight value not greater than the preset threshold is discarded.

[0011] Furthermore, for each other vehicle's state, the comprehensive weight value is calculated based on its corresponding attention weight using one of the following methods: The average value of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value; the maximum value of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value; and the weighted sum of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value.

[0012] Furthermore, the process of fusing the selected target vehicle state information with the vehicle's own state information and inputting it into the decision model to generate the vehicle's acceleration control command includes: The selected target vehicle status information is concatenated with the vehicle status information to form a fused status vector; the fused status vector contains all feature dimensions of the vehicle status information and all feature dimensions of the selected target vehicle status information. The fused state vector is input into the decision model, and the acceleration control command of the vehicle is output after calculation by the decision model.

[0013] Furthermore, the decision-making model is a reinforcement learning model based on the TD3 algorithm, including an Actor network and a Critic network; The Actor network outputs acceleration control commands for the vehicle based on the fused state vector; The Critic network evaluates the acceleration control commands output by the Actor network.

[0014] Furthermore, the method also includes the step of discretizing the input state space of the decision model: Different discretization granularities are used depending on the distance between the vehicle and the collision point; When the distance between the vehicle and the collision point is less than a preset distance, a first discretization granularity is used; when the distance between the vehicle and the collision point is greater than or equal to the preset distance, a second discretization granularity is used, wherein the first discretization granularity is smaller than the second discretization granularity; the preset threshold is set according to the influence range of the collision point.

[0015] Secondly, the present invention also proposes a left-turn decision-making device for unmanned vehicles at urban intersections, comprising at least one processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the at least one processor, implements the aforementioned left-turn decision-making method for unmanned vehicles at urban intersections.

[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects: This invention proposes a left-turn decision-making method for autonomous vehicles at urban intersections, belonging to the field of vehicle autonomous driving control technology. The method includes the following steps: acquiring the vehicle's state information and other vehicle state information from a bird's-eye view; inputting the vehicle's state information as a query matrix, and the combination of the vehicle's state information and other vehicle state information as both key and value matrices, into a multi-head cross-attention network to calculate the attention weight of the vehicle's state to each state in the key matrix; based on the attention weights, selecting target other vehicle state information for left-turn decision-making from the other vehicle state information; fusing the selected target other vehicle state information with the vehicle's state information, inputting it into a decision model to generate acceleration control commands for the vehicle; determining the vehicle's longitudinal behavior based on the acceleration control commands, and generating lateral behavior through trajectory tracking to control the vehicle to complete the left turn. This invention significantly improves the accuracy of traffic decision-making in complex scenarios, providing a solution for the efficient passage of autonomous vehicles in complex traffic environments.

[0017] This invention introduces an improved multi-head cross-attention network, using the vehicle's own state information as the query matrix and the combination of the vehicle's own state information and other vehicle's state information as the key matrix and value matrix. This enables joint attention calculation of the vehicle's own state and other vehicle's state, which can more accurately extract key other vehicle state information that affects the vehicle's left-turn decision, improve the rationality of attention weight allocation, and enhance the interpretability of the network.

[0018] This invention significantly accelerates training convergence by incorporating the collision point into the state space and discretizing it, addressing the problem of training convergence difficulties caused by the complexity of the state space in unprotected left-turn scenarios at intersections. Different discretization granularities are employed based on the distance between the vehicle and the collision point. A finer discretization strategy is used in areas closer to the collision point to more accurately capture potential risks and improve decision-making safety; a coarser discretization strategy is used in areas farther from the collision point to reduce computational complexity and improve decision-making efficiency. Attached Figure Description

[0019] Figure 1 This is a flowchart of a left-turn decision-making method for unmanned vehicles at urban intersections, as proposed in Embodiment 1 of the present invention. Figure 2 This is an architecture diagram of an unmanned vehicle left-turn decision-making method for urban intersections proposed in Embodiment 1 of the present invention. Figure 3 This is a diagram of the improved multi-head cross-attention network structure proposed in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the crossroads model proposed in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of a left-turn decision-making device for unmanned vehicles at urban intersections, as proposed in Embodiment 2 of the present invention. Detailed Implementation

[0020] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0021] Example 1 Embodiment 1 of this invention proposes a left-turn decision method for unmanned vehicles at urban intersections, which addresses the technical problems of insufficient extraction of key states of oncoming vehicles from upstream BEV perception information, poor interpretability, and difficulty in adapting to the specific complex scenario of unprotected left turns at intersections, resulting in low traffic efficiency and insufficient safety.

[0022] The overall process of the left-turn decision-making method for unmanned vehicles at urban intersections proposed in Embodiment 1 of this invention is as follows: the position and speed information of surrounding vehicles perceived by the upstream BEV are obtained, the state information is discretized and key information affecting the vehicle is extracted through an improved cross-attention network, and a reasonable acceleration action is output. The longitudinal behavior of the left-turn process is generated by the decision system, and the lateral behavior is generated by trajectory tracking. Finally, the left-turning vehicle is controlled by the steering wheel and accelerator and brake pedals to obtain a safe and efficient left-turn behavior.

[0023] Figure 1 This is a flowchart of a left-turn decision-making method for unmanned vehicles at urban intersections, as proposed in Embodiment 1 of the present invention. Figure 2 This is an architecture diagram of a left-turn decision-making method for autonomous vehicles at urban intersections proposed in Embodiment 1 of the present invention; combined with Figure 1 and Figure 2 The process of implementing this invention will be described in detail below: In step S1, the status information of the vehicle itself and the status information of other vehicles are obtained from a bird's-eye view. In this invention, the vehicle status information is represented as follows: ; in, This represents the x-coordinate of the vehicle in the global coordinate system; This represents the ordinate of the vehicle in the global coordinate system; Indicates the vehicle speed Quantity; Indicates the vehicle speed Quantity.

[0024] Before obtaining the status information of other vehicles, one or more potential collision points are first determined based on the preset left-turn trajectory of the vehicle and the driving trajectory of vehicles approaching from the opposite lane. The status information of other vehicles is represented as follows: ; ; ; in, This represents the x-coordinate of the vehicle in the global coordinate system; This represents the ordinate of the vehicle in the global coordinate system; Represents the x-coordinate of the corresponding collision point; Represents the ordinate of the corresponding collision point; Indicates the speed of his vehicle Quantity; Indicates the speed of his vehicle Quantity.

[0025] Compared to scenarios involving mixed traffic of pedestrians and vehicles or overtaking, unprotected left turns at intersections are characterized by a fixed driving trajectory and a fixed collision area. Figure 4 This is a schematic diagram of the intersection modeling proposed in Embodiment 1 of the present invention. The left-turn trajectory is simplified to a combination of a straight line and a quarter circle. In the diagram, red represents the reference trajectory for the left-turning vehicle, consisting of three segments (entrance segment straight ab, left-turn semicircle bc with radius R=9m, exit segment straight cd), green represents the trajectory of the oncoming left-lane straight-going vehicle, and black represents the trajectory of the oncoming right-lane straight-going vehicle. The collision points P and Q can be obtained by calculating the intersection of the oncoming straight-going vehicle's straight trajectory and the left-turning vehicle's semicircular trajectory. Therefore, based on the characteristics of the unprotected left-turn scenario at the intersection, the status information of other vehicles is represented as follows: .

[0026] In step S2, the vehicle state information is used as the query matrix, and the combination of the vehicle state information and the other vehicle state information is used as the key matrix and value matrix, respectively. These are then input into the multi-head cross-attention network to calculate the attention weight of the vehicle state to each state in the key matrix. Multi-head cross-attention networks are an improved version of multi-head cross-attention networks, and their construction methods include: The vehicle status information is split into multiple feature dimensions and embedded separately, and the status information of other vehicles is also embedded. The dimensions of the embedding representation are determined according to the preset embedding dimensions. Through a multi-head attention mechanism, the attention weights of each feature dimension of the vehicle's state information to each feature dimension of the key matrix are independently calculated in each attention head, so as to obtain the weight coefficients of other vehicle state information relative to the vehicle's state information. The outputs of each attention head are concatenated and integrated through a linear transformation layer to obtain the comprehensive attention weights of each state information in the key matrix of each feature dimension of the vehicle state information.

[0027] Figure 3 This is the structure diagram of the improved multi-head cross-attention network proposed in Embodiment 1 of the present invention; the input matrix is ​​reconstructed and the vehicle state is added to the Key matrix and Value matrix to achieve a reasonable allocation of attention weights, thereby making the attention network suitable for intersection traffic scenarios. The vehicle state information is then used to... and his vehicle status information Together, they serve as input to the Key and Value matrices, allowing for a reasonable allocation of attention to the states of the vehicle itself and other vehicles.

[0028] Vehicle status information Vehicle status information is used as input to the query matrix. and his vehicle status information Merging is done using both the Key matrix and Value matrix as input; splitting... Think of it as four words in NLP. Take 512, perform embedding, and the specific dimensional changes are as follows: The vehicle state information is decomposed into multiple feature dimensions and represented by each dimension using the first embedding representation. The dimensional changes of the first embedding representation are as follows: ; in, 4 represents the batch size; 4 represents the number of feature dimensions for the vehicle status information; 512 represents the embedding dimension, which is the vector representation dimension that maps the input features. The status information of other vehicles is represented by a second embedding, and the dimensionality of the second embedding is changed as follows: ; Among them, 20 represents the number of feature dimensions for the status information of other vehicles.

[0029] After dot product and softmax, the matrix dimension is: , which represents the weighting coefficient of the intersection traffic environment state relative to the vehicle state.

[0030] To capture features in different subspaces, the attention mechanism is improved to a multi-head attention mechanism: the input Query, Key, and Value are linearly transformed into 8 heads, and each head independently calculates the attention. During the calculation, the dimensions of the Query and the matrix are... The dimensions of the Key and Value matrices are... Finally, the outputs of all heads are concatenated and integrated through a linear transformation layer. Multi-head attention allows the model to learn information in different representation subspaces, thereby significantly improving the model's expressive power.

[0031] In step S3, based on the attention weight, target other vehicle status information for left-turn decisions is filtered from the other vehicle status information; specifically: Obtain the attention weight matrix output by the multi-head cross-attention network, where each element in the attention weight matrix represents the attention weight of a feature dimension of the other vehicle's state information relative to a feature dimension of the own vehicle's state information. For each piece of information about another vehicle's status, a comprehensive weight value is calculated based on its corresponding attention weight. The comprehensive weight value is compared with a preset threshold. The status information of other vehicles with a comprehensive weight value greater than the preset threshold is retained as the target status information of other vehicles, while the status information of other vehicles with a comprehensive weight value not greater than the preset threshold is discarded.

[0032] In this invention, for each other vehicle state, the comprehensive weight value is calculated based on its corresponding attention weight in one of the following ways: The average value of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value; the maximum value of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value; and the weighted sum of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value.

[0033] In Embodiment 1 of this invention, based on the calculated attention weight matrix, for each piece of other vehicle status information, the average value of all attention weights corresponding to that other vehicle status information is taken as the comprehensive weight value. The comprehensive weight value is compared with a preset threshold of 0.2, and other vehicle status information with a comprehensive weight value greater than 0.2 is retained as the target other vehicle status information, while other vehicle status information with a comprehensive weight value less than or equal to 0.2 is discarded.

[0034] The scope of protection of this invention is not limited to the specific values ​​listed in Example 1, and those skilled in the art can make reasonable selections based on the actual situation.

[0035] In step S4, the selected target vehicle status information is fused with the vehicle status information and input into the decision model to generate the vehicle's acceleration control command. The selected target vehicle status information is concatenated with the vehicle status information to form a fused status vector. This fused status vector contains all four feature dimensions of the vehicle status information and all 20 feature dimensions of the selected target vehicle status information.

[0036] The fused state vector is input into the decision model. This embodiment employs a reinforcement learning model based on the TD3 algorithm, including an Actor network and a Critic network. The Actor network outputs acceleration control commands for the vehicle based on the fused state vector, and the Critic network evaluates the acceleration control commands output by the Actor network. Both the Actor and Critic networks are multi-layer neural networks, containing two hidden layers, each followed by a ReLU activation function for nonlinear transformation.

[0037] In step S5, the longitudinal behavior of the vehicle is determined according to the acceleration control command, and the lateral behavior is generated through trajectory tracking to control the vehicle to complete the left turn.

[0038] The longitudinal behavior of the vehicle is determined based on the generated acceleration control command, and the lateral behavior is generated through trajectory tracking. Finally, the left-turning vehicle is controlled by the steering wheel and accelerator and brake pedals to complete the left turn.

[0039] In implementing this invention, the problem of convergence difficulties in attention networks due to their numerous parameters, training challenges, and complex state spaces, compared to MLP networks, is addressed by referencing the application of Scaled Dot-Product Attention in NLP. In Natural Language Processing (NLP), Scaled Dot-Product Attention is widely used in tasks such as machine translation. Its input is a discrete sequence of the language to be translated, and its output is a discrete sequence of the translated language. This mechanism measures the similarity between the query and the key by calculating the dot product, thus obtaining attention weights. These weights are then multiplied by their corresponding values ​​to obtain the final output. This process can be analogized to the translation process, where the model searches for corresponding words in the vocabulary of the target language based on the context, where both the target and translated languages ​​have discrete and finite vocabulary. Although Scaled Dot-Product Attention networks are more complex than traditional NLP structures, with proper training and optimization, they can still achieve effective convergence, resulting in good performance in tasks such as machine translation.

[0040] The inputs of this invention include: vehicle status information. and his vehicle status information The output is the vehicle's acceleration. Considering driving comfort and the difference in the range of values ​​between input and output, the input state space is discretized. After processing by the TD3 algorithm, the output remains continuous. Furthermore, different levels of discretization are used for data sampling and processing based on the distance between the vehicle and surrounding vehicles from the collision point. For the area close to the collision point (0-15m), a more refined discretization strategy is used to capture more accurate dynamic changes and potential risks; while for the area far from the collision point (>15m), a relatively coarse discretization is used to reduce computational complexity and improve efficiency. For example, when the distance between the vehicle and the collision point is less than 15 meters, a distance discretization granularity of 0.5 meters and a velocity discretization granularity of 0.1 m / s are used; when the distance is greater than or equal to 15 meters, a distance discretization granularity of 2 meters and a velocity discretization granularity of 0.5 m / s are used.

[0041] To fully illustrate the effectiveness of the left-turn decision-making method for autonomous vehicles at urban intersections proposed in Embodiment 1 of this invention, 5000 left-turn simulations were conducted in the OpenAI Gym environment using the positions and speeds of three randomly opposing straight-going vehicles. The original and improved models were compared in terms of collision rate and average speed. Experimental results show that, compared to the baseline algorithm, the proposed method reduces the collision rate from 4.8% to 1.6% (a 67% reduction); and increases the average speed from 3.7 m / s to 4.2 m / s (a 14% improvement). This significantly improves the accuracy of traffic decisions in complex scenarios and provides a solution for the efficient passage of autonomous vehicles in complex traffic environments.

[0042] Example 2 Based on the method for left-turn decision-making of unmanned vehicles at urban intersections proposed in Embodiment 1 of the present invention, Embodiment 2 of the present invention also proposes a device for left-turn decision-making of unmanned vehicles at urban intersections. Figure 5 This is a schematic diagram of a left-turn decision-making device for unmanned vehicles at urban intersections, as proposed in Embodiment 2 of the present invention.

[0043] At the hardware level, electronic device 500 includes a processor 510, and optionally, an internal bus 520, a network interface 530, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or it may also include non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations. The processor 510, network interface 530, and memory can be interconnected via an internal bus 520. This internal bus 520 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be categorized as an address bus, data bus, control bus, etc. For ease of illustration, only a single bidirectional arrow is used in this diagram, but this does not imply that there is only one bus or one type of bus. The memory is used to store programs. Specifically, the program can include program code, which includes computer operation instructions. The memory can include main memory 540 and non-volatile memory 550, and provides instructions and data to the processor 510. Processor 510 reads the corresponding computer program from non-volatile memory 550 into memory 540 and then runs it, forming a device for locating the target user at the logical level. Processor 510 executes the program stored in memory and specifically performs the following: In step S1, the status information of the vehicle itself and the status information of other vehicles are obtained from a bird's-eye view. In step S2, the vehicle state information is used as the query matrix, and the combination of the vehicle state information and the other vehicle state information is used as the key matrix and value matrix, respectively. These are then input into the multi-head cross-attention network to calculate the attention weight of the vehicle state to each state in the key matrix. In step S3, based on the attention weight, the target other vehicle status information for left turn decision is filtered from the other vehicle status information; In step S4, the selected target vehicle status information is fused with the vehicle status information and input into the decision model to generate the vehicle's acceleration control command. In step S5, the longitudinal behavior of the vehicle is determined according to the acceleration control command, and the lateral behavior is generated through trajectory tracking to control the vehicle to complete the left turn.

[0044] This device can be integrated into the onboard computer of a drone to achieve real-time target detection; it can also be used as a ground station to process images transmitted back by the drone offline.

[0045] Figure 1It can be applied to processor 510, or implemented by processor 510. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the processor or by instructions in the form of software. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. Of course, in addition to software implementation, the electronic device of this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0046] Example 3 The present invention also proposes a readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the following method steps: In step S1, the status information of the vehicle itself and the status information of other vehicles are obtained from a bird's-eye view. In step S2, the vehicle state information is used as the query matrix, and the combination of the vehicle state information and the other vehicle state information is used as the key matrix and value matrix, respectively. These are then input into the multi-head cross-attention network to calculate the attention weight of the vehicle state to each state in the key matrix. In step S3, based on the attention weight, the target other vehicle status information for left turn decision is filtered from the other vehicle status information; In step S4, the selected target vehicle status information is fused with the vehicle status information and input into the decision model to generate the vehicle's acceleration control command. In step S5, the longitudinal behavior of the vehicle is determined according to the acceleration control command, and the lateral behavior is generated through trajectory tracking to control the vehicle to complete the left turn.

[0047] Embodiment 3 of this application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory that stores a computer program, which can be executed by a processor to complete the steps described in the aforementioned method. The computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0048] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks. Alternatively, if the integrated units of this application are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0049] The application provides a device for left-turn decision-making of unmanned vehicles at urban intersections in Embodiment 2 of this application. For the description of the relevant part of the storage medium for left-turn decision-making of unmanned vehicles at urban intersections proposed in Embodiment 3 of this application, please refer to the detailed description of the corresponding part of the method for left-turn decision-making of unmanned vehicles at urban intersections in Embodiment 1 of this application. It will not be repeated here.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0051] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for left-turn decision-making by autonomous vehicles at urban intersections, characterized in that, Includes the following steps: Obtain the status information of your own vehicle and other vehicles from a bird's-eye view; The vehicle's status information is used as the query matrix, and the combination of the vehicle's status information and other vehicles' status information is used as the key matrix and value matrix, respectively. These are then input into a multi-head cross-attention network to calculate the attention weights of the vehicle's status to each state in the key matrix. Based on the attention weight, the target other vehicle status information for left turn decision is filtered from the other vehicle status information; The selected target vehicle status information is fused with the vehicle's status information and input into the decision model to generate acceleration control commands for the vehicle. The longitudinal behavior of the vehicle is determined according to the acceleration control command, and the lateral behavior is generated through trajectory tracking to control the vehicle to complete the left turn.

2. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 1, characterized in that, The vehicle status information is represented as follows: ； in, This represents the x-coordinate of the vehicle in the global coordinate system; This represents the ordinate of the vehicle in the global coordinate system; Indicates the vehicle speed Quantity; Indicates the vehicle speed Quantity.

3. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 1, characterized in that, Before obtaining the status information of other vehicles, one or more potential collision points are first determined based on the vehicle's preset left-turn trajectory and the travel trajectory of vehicles approaching from the opposite lane; the status information of other vehicles is represented as follows: ； ； ； in, This represents the x-coordinate of the vehicle in the global coordinate system; This represents the ordinate of the vehicle in the global coordinate system; Represents the x-coordinate of the corresponding collision point; Represents the ordinate of the corresponding collision point; Indicates the speed of his car Quantity; Indicates the speed of his car Quantity.

4. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 1, characterized in that, The multi-head cross-attention network is an improved multi-head cross-attention network, and its construction method includes: The vehicle status information is split into multiple feature dimensions and embedded separately, and the status information of other vehicles is also embedded. The dimensions of the embedding representation are determined according to the preset embedding dimensions. Through a multi-head attention mechanism, the attention weights of each feature dimension of the vehicle's state information to each feature dimension of the key matrix are independently calculated in each attention head, so as to obtain the weight coefficients of other vehicle state information relative to the vehicle's state information. The outputs of each attention head are concatenated and integrated through a linear transformation layer to obtain the comprehensive attention weights of each state information in the key matrix of each feature dimension of the vehicle state information.

5. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 1, characterized in that, Based on the attention weights, target other vehicle status information for left-turn decisions is filtered from the other vehicle status information, specifically as follows: Obtain the attention weight matrix output by the multi-head cross-attention network, where each element in the attention weight matrix represents the attention weight of a feature dimension of the other vehicle's state information relative to a feature dimension of the own vehicle's state information. For each piece of information about another vehicle's status, a comprehensive weight value is calculated based on its corresponding attention weight. The comprehensive weight value is compared with a preset threshold. The status information of other vehicles with a comprehensive weight value greater than the preset threshold is retained as the target status information of other vehicles, while the status information of other vehicles with a comprehensive weight value not greater than the preset threshold is discarded.

6. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 5, characterized in that, For each other vehicle's status, the comprehensive weight value is calculated based on its corresponding attention weight, including one of the following methods: The average value of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value; the maximum value of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value; and the weighted sum of all attention weights corresponding to the state of another vehicle is taken as the comprehensive weight value.

7. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 1, characterized in that, The process of fusing the selected target vehicle status information with the vehicle's own status information and inputting it into the decision model to generate acceleration control commands for the vehicle includes: The selected target vehicle status information is concatenated with the vehicle status information to form a fused status vector; the fused status vector contains all feature dimensions of the vehicle status information and all feature dimensions of the selected target vehicle status information. The fused state vector is input into the decision model, and the acceleration control command of the vehicle is output after calculation by the decision model.

8. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 7, characterized in that, The decision-making model is a reinforcement learning model based on the TD3 algorithm, which includes an Actor network and a Critic network. The Actor network outputs acceleration control commands for the vehicle based on the fused state vector; The Critic network evaluates the acceleration control commands output by the Actor network.

9. The method for left-turn decision-making of unmanned vehicles at urban intersections according to claim 1, characterized in that, The method further includes the step of discretizing the input state space of the decision model: Different discretization granularities are used depending on the distance between the vehicle and the collision point; When the distance between the vehicle and the collision point is less than a preset distance, a first discretization granularity is used; when the distance between the vehicle and the collision point is greater than or equal to the preset distance, a second discretization granularity is used, wherein the first discretization granularity is smaller than the second discretization granularity; the preset threshold is set according to the influence range of the collision point.

10. A left-turn decision-making device for an unmanned vehicle at an urban intersection, comprising at least one processor and a memory, the memory storing a computer program, characterized in that, When the computer program is executed by the at least one processor, it implements a left-turn decision method for unmanned vehicles at urban intersections as described in any one of claims 1 to 9.

Images