Assistance system and trajectory generation method for assisted driving

Abstract

The application provides an assisted driving system and a trajectory generation method for assisted driving. The method is used in a feature fusion module of an assisted driving system deployed in a vehicle, the assisted driving system further comprising a trajectory generation module, and the method comprises: acquiring navigation semantic information and a bird's eye view (BEV) feature of the vehicle, the navigation semantic information comprising global trajectory trend information and / or local driving constraint information of a navigation trajectory, and the BEV feature being generated according to surrounding environment data perceived by the vehicle; generating navigation semantic features according to the navigation semantic information, and performing feature screening and fusion processing on the BEV feature by using the navigation semantic features, so as to suppress redundant information in the BEV feature and enhance target features related to the navigation semantic information; and inputting the processed semantic fusion features into the trajectory generation module, so as to generate a planned driving trajectory of the vehicle.

Description

Technical Field

[0001] One or more embodiments of the present invention relate to the field of driver assistance technology, and more particularly to a driver assistance system and a method for generating driver assistance trajectories. Background Technology

[0002] In assisted driving planning scenarios, the traditional segmented modular architecture of "perception → prediction → decision → planning" is still widely used. In this architecture, each module needs to interact through manually defined structured information. For example, the upstream perception module outputs semantic information such as target detection boxes and lane line coordinates, and the downstream prediction and planning module performs trajectory inference based on this semantic information. However, the independent design between modules leads to the propagation and amplification of errors at different levels; at the same time, independent optimization of each module can only achieve local optima and cannot guarantee the global optima of the entire driving decision-making chain. In addition, the process of independent research and development and debugging of multiple modules significantly prolongs the technology iteration cycle. As mid-to-high-level assisted driving functions evolve to scenarios such as complex urban road conditions, the industry urgently needs new technical solutions that can overcome the inherent risks of the above modular architecture and achieve global optimization.

[0003] In related technologies, a two-stage end-to-end approach is typically used to avoid the error accumulation and local optima problems of segmented architectures. This basic approach employs a split architecture of "upstream perception environment modeling" and "downstream predictive planning end-to-end": upstream, BEV (Bird's-Eye-View) generation technology is used to convert raw environmental data collected by onboard sensors into BEV features, and structured environmental semantic information is parsed through a dedicated output head; downstream, the planned trajectory is directly output based on this structured information. Building upon this, some improved approaches simply concatenate or element-wise add basic navigation information such as navigation commands or trajectory location points with the BEV features, allowing the downstream to output the planned trajectory using the processed BEV features.

[0004] However, the aforementioned basic approach suffers from feature redundancy due to a lack of task guidance: BEV features generated from raw environmental data typically contain all information about the current driving environment. This results in BEV features including not only information relevant to the current driving task but also a large amount of redundant background information (such as a stationary sky, distant buildings, and irrelevant vegetation). Therefore, directly using initial BEV features to generate trajectories in the aforementioned basic approach not only reduces the utilization of storage and computing resources but also interferes with subsequent prediction and planning's perception of task-related areas. The aforementioned improved approach suffers from a lack of intent expression due to incomplete navigation query construction: these approaches only introduce a single navigation command or trajectory location point, causing the navigation query to fail to fully convey the task intent, further resulting in lower quality of the subsequently generated planned driving trajectory, which urgently needs improvement. Summary of the Invention

[0005] In view of this, the present invention provides an assisted driving system and a trajectory generation method for assisted driving, to overcome the shortcomings of related technologies. Specifically, the present invention is achieved through the following technical solution: In a first aspect, the present invention provides a trajectory generation method for assisted driving, applied to a feature filtering module in an assisted driving system deployed in a vehicle, wherein the assisted driving system further includes a trajectory generation module, and the method includes: The navigation semantic information and bird's-eye view BEV features of the vehicle are obtained. The navigation semantic information includes global trajectory trend information and / or local driving constraint information of the navigation trajectory. The BEV features are generated based on the surrounding environment data perceived by the vehicle. Based on the navigation semantic information, navigation semantic features are generated, and the navigation semantic features are used to perform feature filtering and fusion processing on the BEV features in order to suppress redundant information in the BEV features and enhance the target features related to the navigation semantic information. The processed semantic fusion features are input into the trajectory generation module to generate the planned driving trajectory of the vehicle.

[0006] Secondly, the present invention provides a driver assistance system, the driver assistance system comprising a feature filtering module and a trajectory generation module, wherein, The feature filtering module is used to: acquire vehicle navigation semantic information and bird's-eye view BEV features, wherein the navigation semantic information includes global trajectory trend information and / or local driving constraint information of the navigation trajectory, and the BEV features are generated based on the surrounding environment data perceived by the vehicle; generate navigation semantic features based on the navigation semantic information, and use the navigation semantic features to perform feature filtering and fusion processing on the BEV features to suppress redundant information in the BEV features and enhance target features related to the navigation semantic information; and input the processed semantic fusion features into the trajectory generation module. The trajectory generation module is used to generate the planned driving trajectory of the vehicle based on the semantic fusion module.

[0007] Thirdly, the present invention also provides a vehicle having an assisted driving system deployed thereon, the system being used to implement the method described in the first aspect above.

[0008] Fourthly, the present invention also provides an electronic device, including a communication interface, a processor, a memory, and a bus, wherein the communication interface, the processor, and the memory are interconnected via the bus; the memory stores machine-readable instructions, and the processor implements the method described in the first aspect above by invoking the machine-readable instructions.

[0009] Fifthly, the present invention also provides a machine-readable storage medium storing machine-readable instructions, which, when invoked and executed by a processor, implement the method described in the first aspect above.

[0010] Therefore, in the technical solution of the present invention, the BEV features generated based on the surrounding environment data are not directly input into the trajectory generation module to generate the driving trajectory. Instead, corresponding navigation semantic features are first generated based on the vehicle's global trajectory trend information. Then, the aforementioned BEV features are filtered and fused using these features to suppress redundant information in the BEV features and enhance the target features related to the navigation semantic information. Finally, the processed semantic fusion features are input into the trajectory generation module to generate the vehicle's planned driving trajectory.

[0011] As can be seen, this solution creatively introduces navigation semantic information, including global trajectory trends and / or local driving constraints, as a guiding signal. It utilizes navigation semantic features generated based on this information to "query" and "filter" BEV features, and generates a planned driving trajectory based on the filtered semantic fusion features. This solution achieves numerous technical effects: On the one hand, this solution enables accurate representation of navigation intent across all dimensions, enhancing the logic and safety of trajectory generation. By acquiring navigation semantic information containing "global trajectory trend information and / or local driving constraint information" and generating navigation semantic features accordingly, it overcomes the limitations of existing technologies that rely solely on single commands or location points. This fusion mechanism allows the generated navigation semantic features to convey high-order semantic information related to navigation, such as global driving trends (e.g., the "straight ahead" intent on a long, straight highway, the curvature characteristics of a roundabout) and / or local driving constraints (e.g., speed limits at urban intersections, turning commands). This high-order semantic information ensures that the system can accurately understand the driving task when facing complex road conditions, thereby effectively avoiding problems such as trajectory swaying or exceeding planned speed limits due to missing intent.

[0012] On the other hand, this solution enables task-driven adaptive feature focusing, significantly reducing interference from redundant information. Utilizing navigation semantic features carrying complete driving intent, it performs deep feature filtering and fusion processing on BEV features containing comprehensive environmental information. This process automatically identifies and enhances target features strongly relevant to the current navigation task (such as the outline of the target lane, the vehicle ahead in the current lane, and the passable area at intersections), while effectively suppressing redundant background information unrelated to the task (such as distant buildings not on the navigation trajectory, irrelevant vehicles in the oncoming lane, and stationary walls outside parking spaces). Through this proactive feature filtering, the feature weights of irrelevant areas are significantly diluted, ensuring that key driving environment information dominates the feature space, thereby significantly improving the accuracy and robustness of trajectory planning in complex scenarios.

[0013] Furthermore, this solution optimizes the integration performance of downstream tasks, improving trajectory generation efficiency and system operating efficiency. The semantic fusion features obtained after filtering and fusion processing suppress (e.g., remove or weaken) a large amount of noisy data that does not contribute to the current driving decision, resulting in extremely high information density. Inputting these high-quality features into the trajectory generation module not only avoids interference from high-dimensional redundant features in the downstream prediction and planning stages but also effectively reduces data transmission bandwidth pressure and the computational load of subsequent modules. This enables the system to achieve more efficient real-time inference with lower resource consumption while ensuring comprehensive perception, thereby improving the overall operating efficiency of the system. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the hardware architecture of an assisted driving system according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a trajectory generation method for assisted driving according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating another trajectory generation method for assisted driving, as shown in an embodiment of the present invention; Figure 4 This is a schematic structural diagram of an electronic device according to an embodiment of the present invention; Figure 5 This is a block diagram illustrating a trajectory generation system for assisted driving, as shown in an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0017] It should be noted that the steps of the corresponding methods in other embodiments are not necessarily performed in the order shown and described in this invention. In some other embodiments, the methods may include more or fewer steps than those described in this invention. Furthermore, a single step described in this invention may be broken down into multiple steps in other embodiments; and multiple steps described in this invention may be combined into a single step in other embodiments.

[0018] In assisted driving planning scenarios, the traditional segmented modular architecture of "perception → prediction → decision → planning" is still widely used. In this architecture, each module needs to interact through manually defined structured information. For example, the upstream perception module outputs semantic information such as target detection boxes and lane line coordinates, and the downstream prediction and planning module performs trajectory inference based on this semantic information. However, the independent design between modules leads to the propagation and amplification of errors at different levels; at the same time, independent optimization of each module can only achieve local optima and cannot guarantee the global optima of the entire driving decision-making chain. In addition, the process of independent research and development and debugging of multiple modules significantly prolongs the technology iteration cycle. As mid-to-high-level assisted driving functions evolve to scenarios such as complex urban road conditions, the industry urgently needs new technical solutions that can overcome the inherent risks of the above modular architecture and achieve global optimization.

[0019] In related technologies, a two-stage end-to-end approach is typically used to avoid the error accumulation and local optima problems of segmented architectures. This basic approach employs a split architecture of "upstream perception environment modeling" and "downstream predictive planning end-to-end": upstream, BEV (Bird's-Eye-View) generation technology is used to convert raw environmental data collected by onboard sensors into BEV features, and structured environmental semantic information is parsed through a dedicated output head; downstream, the planned trajectory is directly output based on this structured information. Building upon this, some improved approaches simply concatenate or element-wise add basic navigation information such as navigation commands or trajectory location points with the BEV features, allowing the downstream to output the planned trajectory using the processed BEV features.

[0020] However, the aforementioned basic solutions suffer from feature redundancy due to a lack of task guidance: BEV features generated from raw environmental data typically contain all information about the current driving environment. This results in BEV features including not only information relevant to the current driving task but also a large amount of redundant background information (such as a stationary sky, distant buildings, and irrelevant vegetation). Therefore, directly using initial BEV features to generate trajectories not only reduces the utilization of storage and computing resources but also interferes with subsequent prediction and planning's perception of task-related areas. For example, in parking scenarios, these solutions retain BEV features of distant walls outside the parking space. This redundant information dilutes the feature weights of key areas such as "parking space outline and distance to adjacent vehicles," leading to a decrease in the accuracy of parking trajectories. In highway scenarios, there is an overemphasis on vehicle features in oncoming lanes, diverting attention from the trajectories of vehicles ahead in the current lane and at the next intersection. The aforementioned improved solutions suffer from a lack of intent expression due to incomplete navigation query construction: these solutions only introduce a single navigation command or trajectory location point, resulting in the navigation query failing to fully convey the task intent, further leading to lower quality of subsequently generated planned driving trajectories, which urgently needs improvement. For example, on long, straight highway sections, existing navigation queries only include the trajectory location and do not incorporate the global trend of trajectory curvature (which is close to 0). As a result, the fused features cannot reflect the driving intention of "keeping straight" and may generate swaying trajectories. At urban intersections, the lack of local speed limit constraints (such as 30km / h) makes the generated trajectories prone to speeding issues.

[0021] To address this issue, this invention proposes a novel driver assistance system and a trajectory generation scheme implemented by this system. This scheme introduces navigation semantic information (including global trajectory trend information and / or local driving constraint information), utilizes this information to generate high-order navigation semantic features to filter BEV features, aiming to improve the information density of semantic fusion features, thereby accurately understanding the driving task and improving the accuracy and robustness of trajectory planning in complex scenarios, as well as the overall operating efficiency of the system. The specific implementation process of this scheme is described in detail below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a schematic diagram of the architecture of an assisted driving system provided in an exemplary embodiment. For example... Figure 1As shown, from a hardware perspective, the system can include only vehicle 11, or it can include both vehicle 11 and cloud server 13 (i.e., a "cloud-edge collaborative" or "vehicle-road-cloud integrated" architecture). From a software perspective, the driver assistance system can include the driver assistance system deployed in vehicle 11, and can also include a cloud service module deployed in cloud server 13 for interacting with the driver assistance system. Specifically, the driver assistance system of vehicle 11 can be deployed in the domain controller 116 (such as an intelligent driving domain controller ADCU) installed in vehicle 11. It is understood that the driver assistance system is an in-vehicle system, and this system can even operate offline.

[0023] In addition to the domain controller 116, the vehicle 11 is also equipped with various sensors, such as environmental sensing devices for collecting data on the external environment and behavior sensing devices for detecting the behavior of the driver 12 inside the vehicle. This application embodiment does not limit the number of these devices or their installation locations. For example, the environmental sensing devices may include image acquisition devices (such as a front-view camera 111, a side-view camera 112, a rear-view camera 113, etc.), ranging radar 114 (such as a lidar, millimeter-wave radar, ultrasonic radar, etc.); the behavior sensing devices may include a camera 115 (such as a visible light camera, an infrared camera, etc.) installed above the driver's seat to detect the direction of the driver 12's eyes, and / or a torque sensor, pressure sensor, etc., installed on the steering wheel to collect the driver 12's hand movements (this sensor is used to implement the hands-off detection (HOD) function). The aforementioned behavior sensing devices and corresponding control devices can constitute the vehicle's DMS system (Driver Monitoring System).

[0024] Of course, the vehicle may also be equipped with at least one other sensor in a suitable location, including but not limited to rain and fog sensors outside the vehicle (such as rain sensors, particulate matter sensors, haze sensors, etc.), cameras, microphones, biosensors, odor sensors, etc. inside the vehicle, to realize corresponding vehicle functions. This application embodiment does not limit this.

[0025] When the driver assistance system includes a cloud server 13 or there is a need to interact with the cloud server 13, the vehicle 11 can establish a network connection with the cloud server 13 through a wireless communication module to interact with the cloud server 13 for data exchange. For example, if the cloud server 13 locally maintains or has the right to access map information (such as high-precision maps) stored in other devices, the vehicle 11 can request the cloud server 13 to obtain the aforementioned map information. The cloud server 13 can be a physical cloud server containing an independent host, or it can be a virtual cloud server hosted by a host cluster, etc. Furthermore, this application embodiment does not limit the number, type, or specific interaction method between the cloud servers 13 and the vehicle 11. As for the network 10 for interaction between the vehicle 11 and the cloud server 13, it can be based on the communication methods supported by the corresponding devices, specifically selecting the appropriate type of wireless network for communication; this application does not limit this.

[0026] Furthermore, the vehicle described in this application (such as vehicle 11) can be a pickup truck, sedan, SUV (Sport Utility Vehicle), campervan, or truck, in terms of its functional form; and it can be a gasoline-powered vehicle or a new energy vehicle (such as a hybrid vehicle, pure electric vehicle, hydrogen fuel cell vehicle, or methanol fuel cell vehicle), in terms of its power source. This invention does not limit the specific form of the vehicle. The driver 12 described in this embodiment is a passenger sitting in the driver's seat of vehicle 11.

[0027] It is worth noting that the driver assistance system of the vehicle described in this application is used to implement driver assistance functions, and this invention does not limit the degree of automation of the driver assistance functions. For example, the driver assistance functions can be at least one of partial automation (i.e., Level 2 autonomous driving), conditional automation (i.e., Level 3 autonomous driving), highly automated driving (i.e., Level 4 autonomous driving), and fully automated driving (i.e., Level 5 autonomous driving). Level 2 to Level 3 autonomous driving functions are often also referred to as driver assistance functions. If the driver assistance system supports multiple levels of automation, functions with higher levels of automation can be compatible with functions with lower levels of automation. For example, if the vehicle simultaneously supports Level 2 and Level 3 autonomous driving, Level 3 autonomous driving can be compatible with Level 2 autonomous driving.

[0028] It is also important to emphasize that the driver assistance functions implemented by the aforementioned driver assistance system must comply with the relevant laws, regulations, and standards of the corresponding country and region (such as the place of sale and / or place of use of the vehicle), and provide relevant personnel (such as drivers, vehicle owners, and back-end administrators) with corresponding operation interfaces for them to choose to authorize or refuse use. Furthermore, the user information (including but not limited to user device information, user biometric information, and other personal information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved in this application are all information and data authorized by the user or fully authorized by all parties. The collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant country and region, and corresponding operation interfaces must be provided for users to choose to authorize or refuse use.

[0029] In one embodiment, the present invention proposes an assisted driving system, which includes a feature filtering module and a trajectory generation module. The feature filtering module is used to: acquire vehicle navigation semantic information and bird's-eye view BEV features, wherein the navigation semantic information includes global trajectory trend information and / or local driving constraint information, and the BEV features are generated based on surrounding environmental data perceived by the vehicle; generate navigation semantic features based on the navigation semantic information, and perform feature filtering and fusion processing on the BEV features using the navigation semantic features to suppress redundant information in the BEV features and enhance target features related to the navigation semantic information; input the processed semantic fusion features into the trajectory generation module; the trajectory generation module is used to: generate the planned driving trajectory of the vehicle based on the semantic fusion module.

[0030] The specific working principles and implementation methods of each module in this system can be found in the detailed description of the embodiments below, which will not be repeated here.

[0031] Figure 2 This is a flowchart illustrating a trajectory generation method for assisted driving according to an embodiment of the present invention. From a software perspective, this method is applied to an assisted driving system deployed in a vehicle, specifically for the feature filtering module within that system. From a hardware perspective, this method can be applied to the vehicle's assisted driving controller (such as...). Figure 1 Domain controller 116 (as shown). Figure 2 As shown, the method includes the following steps 202-206.

[0032] Step 202: Obtain the navigation semantic information and bird's-eye view BEV features of the vehicle. The navigation semantic information includes global trajectory trend information and / or local driving constraint information of the navigation trajectory. The BEV features are generated based on the surrounding environment data perceived by the vehicle.

[0033] The vehicle can automatically or in response to the driver's triggering behavior to activate the assisted driving system. The system then enters assisted driving mode, where it can intervene and control the vehicle's movement according to its own vehicle control logic, thereby enabling assisted driving functions. The trajectory generation scheme described in this invention can be implemented by the assisted driving system after the assisted driving function is activated, to automatically control the vehicle in assisted driving mode. It should be noted that at any given moment during the vehicle's operation (i.e., while in driving mode), its speed may be zero (e.g., while waiting at a red light) or non-zero (e.g., while moving forward). The vehicle being in driving mode can specifically be in manual driving mode (assisted driving system not activated or activated but not actually intervening in vehicle control) or automatic driving mode (assisted driving system activated and actively controlling the vehicle).

[0034] During vehicle operation, real-time data on current driving conditions and surrounding environment can be collected. The current driving data refers to data reflecting / characterizing the vehicle's own driving state, which can be collected by sensors mounted on the vehicle (such as wheel speed sensors, inertial measurement units (IMUs), steering angle sensors, etc.). For example, this may include basic motion state data, such as real-time vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate (angular velocity of vehicle steering), etc.; chassis and attitude data, such as steering wheel angle (or front wheel steering angle), vehicle pitch angle, roll angle, tire speed, etc.; position and trajectory data, such as real-time position coordinates (latitude and longitude), heading angle (vehicle orientation), and mileage; and operation command status data, such as accelerator pedal opening (throttle depth), brake pedal status (whether the brake is applied and the force applied), gear status, etc., which will not be elaborated further. The surrounding environment data refers to data reflecting the state of the physical world outside the vehicle, which can be collected in real-time by the aforementioned onboard environmental perception devices (such as lidar, millimeter-wave radar, cameras, ultrasonic radar, etc.) or obtained by querying high-precision maps based on the vehicle's location. Examples may include dynamic obstacle data, such as the position, speed, acceleration, heading angle, size, and category of other vehicles, pedestrians, and non-motorized vehicles (bicycles / electric vehicles); static road element data, such as the type (solid / dashed) and position of lane lines, road boundaries (curb, median), traffic signs (speed limit signs, directional signs), traffic lights (red and green light status), and road markings (zebra crossings, guide lines); road geometry and topology data, such as road curvature (sharpness of curves), slope, lane width, and the direction of the road ahead; and environmental and road condition data, such as weather conditions (rain, snow, fog), light intensity (day / night / backlight), and road surface adhesion conditions (dry, slippery, icy, waterlogged), etc., which will not be elaborated further.

[0035] While the vehicle is in motion, navigation can be activated, allowing the system to guide the vehicle towards its destination according to the navigation instructions provided. To achieve this, the system needs to acquire the vehicle's navigation semantic information, which may include global trajectory trend information and local driving constraint information. This navigation semantic information refers to a high-dimensional feature vector representing the vehicle's intent, geometric trends, and environmental constraints within the global driving task and local traffic rules, aiming to provide task-specific contextual guidance to downstream modules. This information is not simply a "left turn / right turn" instruction, but a comprehensive set of information integrating macroscopic route trends and microscopic driving constraints. For example, it not only includes the global trajectory guiding the vehicle to its destination but can also incorporate traffic rule restrictions and driving behavior guidance elements for the current or upcoming road segment.

[0036] Global trajectory trend information refers to the overall driving route and geometric characteristics of a vehicle over a relatively long distance (macro scale). This information indicates the "general direction" of travel, focusing on the continuity and overall shape of the navigation path, primarily guiding the vehicle to maintain the correct macroscopic driving trajectory. For example, it can include the curvature of the navigation path (whether it's straight or curved, the specific degree of the curve, etc.), the direction of extension, and lane topology connections. For instance, in a long straight highway scenario, if the system extracts a path curvature close to 0 within 5 kilometers ahead, the global path trend could include "stay straight," and the vehicle doesn't need to frequently adjust the steering wheel. In roundabout or sharp curve scenarios, if the system identifies a path with continuous arc-shaped curvature, the global trajectory trend information could include "continuous detour" or "sharp turn," requiring the vehicle to prepare for continuous turns in advance. In a ramp merging into the main road scenario, if the system identifies that the current lane is about to merge with another lane, the global trajectory trend information could include "path intersection," requiring the vehicle to adjust its position in advance to prepare for the merge.

[0037] Local driving constraint information refers to the specific traffic rules, operational instructions, or environmental restrictions that a vehicle must comply with within a short distance (at a micro scale, typically the current road segment or the next intersection). This information instructs how to drive "at the moment," focusing on safety, compliance, and specific driving actions. Examples include speed limits, turning instructions, lane-changing requirements, and no-entry / restricted rules. For instance, in an urban intersection with a speed limit, the system learns from navigation maps or roadside perception (such as OCR recognition of an image captured by a forward-facing camera) that the speed limit at the upcoming intersection is 30 km / h. In this case, the local driving constraint information could include "speed constraint (specifically, the intersection speed limit is 30 km / h)" to prevent the vehicle from exceeding the speed limit. In a highway intersection turning scenario, the navigation instruction might suggest "exit the highway to the right in 500 meters." In this case, the local driving constraint information could include "action constraint," requiring the vehicle to perform lane-changing and exit actions at specific locations. In special road restriction scenarios (such as road construction ahead, requiring vehicles to use the oncoming lane, or a lane being a bus lane where other vehicles are prohibited), the local driving constraint information can include these "right-of-way or spatial constraints".

[0038] Understandably, if navigation semantic information only includes global trajectory trend information, it can ensure the macroscopic continuity of the generated planned driving trajectory. This information (such as path curvature and extension direction) allows the vehicle to anticipate the overall direction of a relatively long distance in the future. For example, on long straight sections of highways or ramps with large curvatures, the vehicle can clearly understand the macroscopic intention of "keeping straight" or "smoothly navigating the curve," thereby outputting a smooth, human-like trajectory and avoiding situations such as "dragging the dragon" (swaying left and right) or frequent micro-adjustments of the steering wheel. Furthermore, it can provide forward-looking path guidance because this information helps the system lock onto the correct macroscopic path in advance in complex road networks (such as interchanges, roundabouts, and intersections with multiple branches), preventing dangerous actions such as sudden braking or sharp turns due to insufficient perception of local information when approaching intersections.

[0039] If navigation semantic information only includes local driving constraint information, it can ensure the compliance and safety of driving behavior. This information (such as speed limits, no-entry restrictions, and turning instructions) is directly related to specific traffic rules and real-time road conditions. It can enforce constraints on the vehicle's instantaneous behavior, such as strictly enforcing the 30km / h speed limit at urban intersections or avoiding restricted areas in construction zones, fundamentally avoiding safety hazards such as speeding and violations. Furthermore, it can precisely guide specific driving actions. This information can guide the vehicle to complete refined operations. For example, when approaching an intersection, local information will clearly indicate "change lanes to the left" or "pull over," ensuring that the vehicle performs the specific turning, lane-changing, or parking actions in the correct location and does not miss key intersections.

[0040] If navigation semantic information contains both global trajectory trend information and local driving constraint information, then navigation semantic features can simultaneously convey the dual intent of "how to go globally" and "how to drive at the moment." This enables the system to understand both global long-distance planning and current traffic regulations, thereby ensuring that the subsequently generated planned driving trajectory can maintain macroscopic smoothness while strictly adhering to microscopic safety and compliance restrictions.

[0041] In addition, the system also needs to acquire the vehicle's BEV characteristics. The BEV characteristics described in this manual can be intuitively understood as an environmental map with real physical distances, viewed from above as a "bird's-eye view".

[0042] In one embodiment, the BEV features of the vehicle described in this invention can be raw BEV features directly generated based on surrounding environmental data. For example, the system can utilize BEV generation technology to generate BEV features in real time based on the aforementioned surrounding environmental data. Exemplarily, the system can input raw images captured by external cameras into a neural network (such as a ResNet backbone network) to extract 2D image features containing information such as color, texture, and edges. Then, a perspective transformation (converting 2D to 3D / BEV) is performed: using the camera's intrinsic and extrinsic parameters (position, angle, etc.) and depth estimation technology, the original 2D image features with a "nearer is larger, farther is smaller" perspective effect are projected and mapped into a unified 3D bird's-eye view space (BEV space). Finally, feature fusion and generation are performed, such as seamlessly stitching and fusing the BEV space features converted from different camera perspectives to form a complete BEV feature map covering the area around the vehicle (e.g., a 100m × 100m range). On this feature map, each grid (pixel) corresponds to a fixed physical distance in the real world (e.g., 0.5m × 0.5m).

[0043] In another embodiment, the BEV features of the vehicle described in this invention can also be enhanced BEV features obtained by fusing the vehicle's navigation semantic information and the aforementioned original BEV features. For example, the navigation semantic information generated in the previous embodiment can be encoded to obtain BEV spatiotemporal correlation features, and a navigation heatmap can be generated based on the vehicle's navigation semantic information. The BEV spatiotemporal correlation features and the navigation heatmap can then be fused to obtain enhanced BEV features. When generating a navigation heatmap based on the vehicle's navigation semantic information, a first type of heatmap can be generated based on the sample vehicle location information, and a second type of heatmap can be generated based on the sample navigation-related information. The spatial dimensions of both the first and second type of heatmaps are aligned with the sample BEV features. Then, the first type of heatmap and the second type of heatmap are concatenated by channels to obtain the navigation heatmap. When fusing the BEV spatiotemporal correlation features and the navigation heatmap, the navigation heatmap and the BEV spatiotemporal correlation features can be fused first to obtain fused features. Then, the fused features are weighted by channel dimensions using an attention mechanism to obtain the enhanced BEV features. Understandably, by transforming navigation semantic information into navigation heatmaps and fusing them with BEV spatiotemporal correlation features, the model can combine macroscopic global driving intentions with local environmental spatiotemporal information. The planned trajectory generated based on enhanced BEV features can accurately match the navigation intention (Query), overcoming the problem in related technologies where the planned trajectory is disconnected from the global driving intention due to relying solely on structured environmental information for deduction. This significantly improves the operational reliability of mid-to-high-level intelligent driving in complex road conditions.

[0044] Step 204: Generate navigation semantic features based on the navigation semantic information, and use the navigation semantic features to perform feature filtering and fusion processing on the BEV features, so as to suppress redundant information in the BEV features and enhance the target features related to the navigation semantic information.

[0045] In one embodiment, when generating navigation semantic features based on the navigation semantic information, a targeted generation method can be adopted according to the specific content of the navigation semantic information. For example, if the navigation semantic information only includes the global trajectory trend information, feature extraction and embedding mapping can be performed on the global trajectory trend information, and the resulting global trend feature vector can be used as the navigation semantic feature. If the navigation semantic information only includes the local driving constraint information, feature extraction and embedding mapping can be performed on the local driving constraint information, and the resulting local constraint feature vector can be used as the navigation semantic feature. If the navigation semantic information includes both the global trajectory trend information and the local driving constraint information, feature extraction and embedding mapping can be performed on the global trajectory trend information and the local driving constraint information respectively to obtain the corresponding global trajectory feature vector and local constraint feature vector; and the global trajectory feature vector and the local constraint feature vector can be fused to obtain the navigation semantic feature.

[0046] As mentioned above, the global trajectory trend information may include trajectory curvature information and / or trajectory slope information, and the local driving constraint information may include lane priority information, speed limit information, and / or traffic sign semantic information. Therefore, when performing feature extraction and embedding mapping processing on the global trajectory trend information, the path curvature information and / or path slope information can be obtained from the global trajectory trend information, and this information can be mapped into the global trend feature vector through a first embedding layer. Similarly, when performing feature extraction and embedding mapping processing on the local driving constraint information, the lane priority information, speed limit information, and / or traffic sign semantic information can be obtained from the local driving constraint information, and this information can be mapped into the local constraint feature vector through a second embedding layer.

[0047] The first embedding layer is a feature encoder that processes global trajectory trend information, mapping / converting this information into a global trend feature vector. It processes global trajectory trend information such as path curvature (describing whether the road is straight or curved) and path slope (describing the degree of uphill or downhill slope), which is typically continuous numerical data. This embedding layer can highly condense and represent the macroscopic geometry and orientation of the road, allowing the system to "understand" future road condition trends / movements over long distances. Correspondingly, the second embedding layer is a feature encoder that processes local driving constraint information, mapping this information into a local constraint feature vector. It processes local driving constraint information such as lane priority (e.g., bus lanes, main and auxiliary roads), speed limits (e.g., 60 km / h), and traffic sign semantics (e.g., no U-turns, stop and yield), which is typically discrete rules or specific numerical values. This embedding layer can accurately represent the microscopic driving rules and safety boundaries that must be followed on the current road segment. As can be seen, the first embedding layer processes "geometric and physical attributes" (what the road looks like), focusing on the macroscopic road morphology and emphasizing the extraction of continuous spatial morphological features (such as the smoothness of curves and the trend of slope changes); the second embedding layer processes "rules and logical attributes" (how driving is allowed on the road), focusing on the microscopic traffic regulations and vehicle status, emphasizing the extraction of discrete semantic category features or key numerical thresholds (such as the semantics of "prohibited" and the magnitude of "speed limit"). In a neural network implementation, the first and second embedding layers can each have their own independent set of learnable parameters (weights and biases), thus forming two completely independent and non-interfering feature extraction channels.

[0048] Understandably, this scheme, by setting independent first and second embedding layers, can establish dedicated heterogeneous data decoupling encoding channels for global geometric values ​​(such as curvature and slope) and local logical rules (such as speed limits and signage semantics) with drastically different physical meanings and dimensions, effectively avoiding information confusion and ensuring feature purity. This independent parameter design allows the feature selection module to perform optimized feature learning for continuous road morphological changes and discrete traffic rule semantic constraints, significantly improving the overall representation capability. Simultaneously, this loosely coupled modular architecture gives the system strong flexibility and scalability; when adding navigation information in subsequent iterations, only the corresponding embedding layer needs to be fine-tuned without reconstructing the entire network, helping to reduce system maintenance costs.

[0049] Specifically, feature extraction and embedding mapping can be performed on any of the above-mentioned types of information using one of the following methods: Vectorization encoding, such as mapping numerical information into feature vectors using a fully connected layer MLP (Multi-Layer Perceptron). This encoding method not only has strong nonlinear mapping capabilities but also a flexible and universal structure, suitable for numerical, tabular, or navigation semantic information without specific spatial / temporal structures. Sequence encoding, where the information is sequential data, recurrent neural networks (such as Long Short-Term Memory networks LSTM and Gated Recurrent Units GRU) or Transformer encoders can be used to extract temporal features. This encoding method not only has the ability to capture long-range dependencies but also supports high parallelism and global modeling (Transformer), suitable for navigation semantic information with sequential dependencies, such as text and time series. Graph structure encoding, where graph neural networks (such as graph convolutional networks GCN and graph attention networks GAT) can be used for feature aggregation when the information contains topological structures. This encoding method not only efficiently aggregates neighborhood information but also has dynamic weighting and strong adaptability, suitable for navigation semantic information with complex topological relationships. Convolutional feature extraction: When information is rasterized into a map format, CNN (Convolutional Neural Network) can be used to extract spatial features. This encoding method not only preserves spatial structure but also automatically extracts hierarchical features and possesses parameter efficiency and translation invariance, making it suitable for navigation semantic information in the form of images, videos, or rasterized / gridized maps. Embedding lookup: For discrete command information, the corresponding feature vector can be directly obtained through a pre-defined learnable embedding layer. This encoding method not only achieves semantic representation and dimensionality reduction but is also computationally efficient, making it suitable for discrete navigation semantic information.

[0050] Furthermore, when fusing the global trajectory feature vector and the local constraint feature vector in the aforementioned embodiments, a suitable fusion method can be selected based on the actual situation such as data type or data volume. For example, basic concatenation and linear transformation can be performed, such as directly concatenating the global trajectory feature vector and the local constraint feature vector in the channel dimension to form a longer joint vector, and then performing linear transformation and nonlinear activation through one or more fully connected layers (MLP) to finally map to the navigation semantic features in the target dimension. The implementation logic of this fusion method is simple, the computational overhead is minimal, and it can completely preserve the original information of the two vectors. As another example, if the global trajectory feature vector and the local constraint feature vector have the same dimension, they can also be directly added or multiplied element by element. The implementation logic of this fusion method is intuitive and efficient. The addition operation is equivalent to lossless superposition of global and local information in the feature space, while the multiplication operation has a certain implicit "gating" or "modulation" meaning, which can use one vector as a weight to dynamically adjust the feature response of the other vector. For example, assuming the global trajectory feature vector and the local constraint feature vector have the same dimension, they can be directly added or multiplied element-wise. This fusion method is intuitive and efficient. Addition is equivalent to lossless superposition of global and local information in the feature space, while multiplication implies implicit "gating" or "modulation," allowing one vector to be used as a weight to dynamically adjust the feature response of the other. Alternatively, a cross-attention mechanism can be used, treating the global trajectory feature vector as a query and the local constraint feature vector as a key and value. The attention weight matrix is ​​then used to dynamically extract the local constraint information most relevant to the global path (and vice versa). This fusion method automatically learns and captures the deep correlation between global trends and local constraints, achieving highly adaptive feature weighting, which is particularly suitable for complex and ever-changing driving scenarios. For example, gating concepts similar to those in Gated Recurrent Units (GRUs) or Long Short-Term Memory (LSTM) networks can be introduced. A gating network consisting of a sigmoid activation function and fully connected layers can be designed, taking global and local features as input to learn a dynamic weight vector with values ​​between 0 and 1. This weighted sum controls the retention ratio of the two feature vectors. This fusion method offers strong interpretability and scene adaptability. Alternatively, the global trajectory feature vector and the local constraint feature vector can be viewed as two independent nodes in a graph structure. By constructing a simple fully connected graph and utilizing the message passing and aggregation mechanisms of Graph Neural Networks (GNNs), the feature representations of these two nodes can be interactively updated, ultimately extracting the fused node features.This fusion method aggregates information through structured relationship modeling. If navigation semantic information is expanded to include more types of features in the future (such as adding nodes for traffic rules, weather information, etc.), the graph fusion method has excellent scalability and compatibility.

[0051] For example, when fusing the global trajectory feature vector and the local constraint feature vector, the global trajectory feature vector and the local constraint feature vector can be concatenated along the channel dimension, and a pre-defined fully connected layer can be used to perform linear mapping and nonlinear transformation on the concatenated vector. Specifically, the concatenation process can include channel concatenation and linear mapping, such as concatenating the global trend feature vector and the local constraint feature vector end-to-end along the channel dimension, and then performing linear mapping and nonlinear transformation through a fully connected layer or convolutional layer to output the final navigation semantic features. Alternatively, if the dimensions of the global trajectory feature vector and the local constraint feature vector are the same, element-wise addition or subtraction can be performed on them. This method has a lower computational cost and can achieve direct information superposition. Furthermore, a self-attention mechanism or a pre-defined gating network can be used to calculate the dynamic weight coefficients of the global trajectory feature vector and the local constraint feature vector respectively, and then a weighted sum can be performed based on these dynamic weight coefficients. One approach, based on attention mechanisms to calculate dynamic weight coefficients, is essentially attention pooling fusion. Here, the two vectors can be treated as a sequence, their importance weights are calculated using self-attention, and then weighted aggregation is performed to obtain the final feature containing key information from both. The other approach, based on gating networks to calculate dynamic weights, is essentially dynamic fusion based on gating mechanisms. For example, a learnable gating network (e.g., using a sigmoid function to generate weight coefficients between 0 and 1) is introduced to dynamically adjust the contribution ratio of global trend information and local constraint information before weighted summation. This allows the feature selection module to adaptively determine whether global trajectory trend information or local driving constraint information is more relevant to the actual scenario.

[0052] In one embodiment, after generating navigation semantic features in the aforementioned manner, post-processing can be performed. For example, positional encoding can be added to the navigation semantic features to represent the relative positional relationship of the navigation semantic information in time or space. The added navigation semantic features can then be used as the object of subsequent feature filtering and fusion processing. The positional encoding can employ absolute positional encoding (APE) to represent specific distances or timestamps, or relative positional encoding (RPE) to represent the relative sequence of navigation commands. Specifically, it can be implemented using sine / cosine encoding or learnable embedding vectors. By adding positional encoding, a clear spatiotemporal context can be provided for the navigation intent, enabling the system to accurately perceive "when and where" to do something while clearly understanding "what to do," thereby contributing to the generation of a smoother and more coherent driving trajectory. Simultaneously, as a key spatial anchor point, positional encoding can achieve precise spatial alignment between navigation semantics and BEV environmental features, significantly improving the accuracy of feature filtering and fusion. Furthermore, this approach achieves feature decoupling between semantic content and location information, which not only reduces the representation burden of the model but also effectively improves the system's generalization ability and inference efficiency when handling driving tasks with different ranges.

[0053] Furthermore, when using the navigation semantic features to perform feature filtering and fusion processing on the BEV features, appropriate processing methods can be adopted according to the actual situation.

[0054] In one embodiment, the navigation semantic features can be used as query terms, and the BEV features as key-value pairs. A cross-attention mechanism is used to filter and fuse the BEV features. Essentially, this process treats the navigation semantic features as navigation intent, actively scanning the "visual scene (BEV features)" to weaken or even filter out irrelevant clutter in the BEV features. This approach achieves intent-driven feature filtering through the cross-attention mechanism. By actively querying the BEV features using navigation semantic features as query terms, it can accurately locate and highlight spatial areas strongly related to the current driving intent from a complex panoramic environment, while effectively suppressing interference from irrelevant background noise. Furthermore, this dynamic attention weighting mechanism breaks the limitations of traditional fixed-weight fusion, enabling the model to adaptively adjust its attention to different spatial locations in the environment according to different navigation commands. This achieves deep alignment and efficient fusion of high-level abstract semantics and low-level spatial geometric features, helping to improve the accuracy and safety of trajectory planning in complex road conditions.

[0055] Specifically, the aforementioned cross-attention mechanism can be a global cross-attention mechanism, or it can employ various advanced or variant implementations. For example, a bidirectional cross-attention mechanism can be used, such as sharing an attention matrix to allow two feature sequences (i.e., the navigation semantic features and BEV features) to update each other simultaneously; in other words, not only is navigation information used to filter BEV features, but the environmental context in the BEV is also used to correct and enrich the navigation semantic features. This approach enables deeper dual-sequence interaction, enhances feature complementarity, and has higher computational efficiency (fewer parameters, faster inference speed), while retaining richer feature information. Another example is a deformable / sparse cross-attention mechanism, where the mechanism no longer focuses on the global context but dynamically samples a small number of key reference points on the BEV feature map for attention calculation based on the query's guidance, greatly reducing computational complexity (from a quadratic relationship with the feature map size to a linear relationship) and improving fusion efficiency. This is particularly suitable for processing high-resolution BEV feature maps or achieving high real-time inference speeds in remote perception scenarios. For example, a cross-modal adaptive attention mechanism (CMAA) can be employed. This mechanism introduces a learnable, context-aware modality gating mechanism, going beyond simple Q / K / V calculations to dynamically weight the contributions of different modalities based on the current scene context (such as changes in sensor reliability due to weather, lighting, etc.). This mechanism addresses the limitation of fixed-weight fusion strategies in adapting to changing environments. For instance, in heavy rain, the model can automatically reduce the weight of visual features and rely more heavily on features from LiDAR or millimeter-wave radar, thereby improving the system's robustness under extreme conditions. Another example is multi-head cross-attention with talking heads, where features are divided into multiple "attention heads" for parallel processing, each focusing on a different feature subspace (e.g., some heads focus on lane geometry, while others focus on traffic sign semantics). Furthermore, a "talking heads" mechanism can be used to facilitate information exchange and collaboration between different attention heads. This mechanism enhances the diversity of feature representations, avoids the model focusing only on the correlation of a single dimension, and ensures that the processed semantic fusion features are more comprehensive and accurate.

[0056] Furthermore, when performing feature filtering and fusion processing on the BEV features using a cross-attention mechanism, the navigation semantic features can first be linearly transformed to obtain a query vector, and the BEV features can be linearly transformed to obtain key and value vectors respectively. Then, the similarity score between the query vector and the key vector is calculated, and the similarity score is normalized to obtain an attention weight matrix. Finally, the value vectors are weighted and summed based on the attention weight matrix to obtain the semantic fusion feature. This method maps heterogeneous modalities to a unified high-dimensional space for deep interaction by performing a linear transformation of query (Q) / key (K) / value (V) vectors on the navigation semantic features and BEV features. Utilizing the similarity calculation and normalization of the query and key, the model can dynamically generate attention weights based on content matching, thereby accurately quantifying the contribution of environmental features at different spatial locations to the current driving intent. Finally, weighted summation achieves targeted aggregation of information, efficiently injecting navigation intent into spatial features while fully preserving BEV environmental details, achieving a high degree of alignment and fusion of deep semantic and geometric information. This method significantly improves the accuracy of the final generated trajectory. Figure 1 Consistency and Environmental Adaptability: By efficiently injecting navigation intent into spatial features, the generated trajectory can accurately respond to macro-path guidance, pre-planning smooth and coherent lane change or turning paths, and avoiding trajectory swaying near intersections. At the same time, the aggregation of environmental features based on dynamic attention weights ensures that the trajectory strictly follows the traffic rules and obstacle avoidance constraints of key areas during the generation process, achieving high-quality trajectory output that combines safety, compliance, and a human-like driving experience.

[0057] Besides employing cross-attention mechanisms, basic feature fusion methods or advanced adaptive / dynamic fusion methods can also be used to achieve feature selection and fusion processing. A basic feature fusion method can be element-wise addition, where navigation semantic features and BEV features are directly added element-wise. This method has extremely low computational complexity and is suitable for systems with extremely high real-time requirements. Alternatively, channel concatenation and convolution operations can be used, where navigation semantic features and BEV features are concatenated along the channel dimension, and then the merged feature map is integrated through one or more convolutional layers (such as 1×1 or 3×3 convolutions), reducing the channel dimension. This method balances computational efficiency and effective feature preservation. Advanced adaptive / dynamic fusion methods can include cross-modal adaptive attention mechanisms, such as introducing context-aware modality gating mechanisms. These mechanisms explicitly and dynamically adjust the contribution weights of different modalities based on scene-related reliability (such as changes in sensor quality due to weather or lighting), rather than simply performing fixed attention calculations. Alternatively, a dedicated DDF (Dual Dynamic Fusion) module can be designed to adaptively select valuable information from different modalities (such as cameras and LiDAR), thereby fully utilizing the inherent complementary information between different modalities and solving the information loss problem caused by simple fusion strategies. Another approach is the CIT (Cross-modal Interaction Transform) algorithm, which establishes a bidirectional interaction mechanism between two feature spaces, enhances their respective feature representations through a self-attention mechanism, and then fuses them to solve the semantic misalignment problem between modalities. Alternatively, a Max Sigmoid attention mechanism can be used, employing features from auxiliary modalities (such as text features) to update and focus the features of the primary modality (such as the BEV region), etc., which will not be elaborated further.

[0058] It should be noted that this solution utilizes the navigation semantic features to perform feature filtering and fusion processing on the BEV features to suppress redundant information in the BEV features and enhance target features related to the navigation semantic information. This means that the processing amplifies the response values ​​of environmental features (such as target lane lines and intersection areas) that are highly matched with the navigation semantic information (such as lane changes, turns, or speed limits) in the fused features, while attenuating or even eliminating the response values ​​of background environmental features (such as stationary objects on the roadside or non-target areas) that are unrelated to the navigation semantic information. Thus, while retaining key environmental constraints, interference noise is filtered out, generating high signal-to-noise ratio semantic fusion features focused on the current driving task.

[0059] Step 206: Input the processed semantic fusion features into the trajectory generation module to generate the planned driving trajectory of the vehicle.

[0060] As can be seen from steps 202-206 and related embodiments above, this solution creatively introduces navigation semantic information, including global trajectory trends and / or local driving constraints, as a guidance signal. It utilizes navigation semantic features generated based on this semantic information to "query" and "filter" BEV features, and generates a planned driving trajectory based on the filtered semantic fusion features. This solution can achieve numerous technical effects: On the one hand, this solution enables accurate representation of navigation intent across all dimensions, enhancing the logic and safety of trajectory generation. By acquiring navigation semantic information containing "global trajectory trend information and / or local driving constraint information" and generating navigation semantic features accordingly, it overcomes the limitations of existing technologies that rely solely on single commands or location points. This fusion mechanism allows the generated navigation semantic features to convey high-order semantic information related to navigation, such as global driving trends (e.g., the "straight ahead" intent on a long, straight highway, the curvature characteristics of a roundabout) and / or local driving constraints (e.g., speed limits at urban intersections, turning commands). This high-order semantic information ensures that the system can accurately understand the driving task when facing complex road conditions, thereby effectively avoiding problems such as trajectory swaying or exceeding planned speed limits due to missing intent.

[0061] On the other hand, this solution enables task-driven adaptive feature focusing, significantly reducing interference from redundant information. Utilizing navigation semantic features carrying complete driving intent, it performs deep feature filtering and fusion processing on BEV features containing comprehensive environmental information. This process automatically identifies and enhances target features strongly relevant to the current navigation task (such as the outline of the target lane, the vehicle ahead in the current lane, and the passable area at intersections), while effectively suppressing redundant background information unrelated to the task (such as distant buildings not on the navigation trajectory, irrelevant vehicles in the oncoming lane, and stationary walls outside parking spaces). Through this proactive feature filtering, the feature weights of irrelevant areas are significantly diluted, ensuring that key driving environment information dominates the feature space, thereby significantly improving the accuracy and robustness of trajectory planning in complex scenarios.

[0062] Furthermore, this solution optimizes the integration performance of downstream tasks, improving trajectory generation efficiency and system operating efficiency. The semantic fusion features obtained after filtering and fusion processing suppress (e.g., remove or weaken) a large amount of noisy data that does not contribute to the current driving decision, resulting in extremely high information density. Inputting these high-quality features into the trajectory generation module not only avoids interference from high-dimensional redundant features in the downstream prediction and planning stages but also effectively reduces data transmission bandwidth pressure and the computational load of subsequent modules. This enables the system to achieve more efficient real-time inference with lower resource consumption while ensuring comprehensive perception, thereby improving the overall operating efficiency of the system.

[0063] In one embodiment, when inputting the processed semantic fusion features into the trajectory generation module, the semantic fusion features can be directly input into the trajectory generation module. Alternatively, they can be dimensionality-reduced before input, such as by calling a preset standardization compression module to perform channel dimension compression and format standardization on the semantic fusion features, and then inputting the processed task adaptation features into the trajectory generation module. This method significantly reduces the computational load and parameter count of the subsequent trajectory generation module by using a standardization compression module to perform channel compression and format standardization on the high-dimensional semantic fusion features, greatly improving the inference efficiency and real-time response capability of the assisted driving system. At the same time, this module can effectively filter feature redundancy and noise, extract highly discriminative core task adaptation features, and achieve precise alignment of the front-end and back-end networks in terms of feature dimensions, enhancing the robustness, flexibility, and scalability of the overall model architecture.

[0064] The standardized compression module only needs to achieve dimensional uniformity of the fused features and compression of redundant information to adapt to the input requirements of the downstream trajectory generation module. This embodiment of the invention does not limit the specific algorithm or mechanism within this module. For example, this module can be a low-dimensional convolutional layer (i.e., the channel dimension of the convolutional layer is smaller than the channel dimension of the semantic fusion features, where the channel dimension of the processed task-adaptive features is equal to the channel dimension of the convolutional layer; specifically, the convolutional layer can be a 1×1 convolutional layer), a pooling layer (such as global average pooling (GAP), max pooling, etc.), a feature selection algorithm module, or a lightweight neural network layer, etc.

[0065] In one embodiment, after generating the planned driving trajectory of the vehicle in the aforementioned manner, the driver assistance system can control the vehicle to drive along the trajectory. Specifically, if there is only one planned driving trajectory, the vehicle can be controlled to drive along that trajectory. If there are multiple planned driving trajectories, safety and / or comfort scores can be applied to each trajectory, and the vehicle can then be controlled to drive along the trajectory with the highest score. This solution improves the robustness and safety of decision-making: Neural networks (especially large end-to-end models), while trained, may still output trajectories with potential collision risks or that do not comply with traffic rules when facing extreme long-tail scenarios or perceptual noise. By adding a "safety score" step, the system no longer blindly trusts the model's direct output but instead performs a secondary screening of candidate trajectories based on explicit safety rules (such as collision detection, compliance with drivable areas, etc.). This is equivalent to adding a hard constraint to the model's decision-making, ensuring that the final executed trajectory is absolutely safe and greatly reducing the risk of accidents. Furthermore, this solution balances human-like design with passenger comfort. While ensuring basic safety across multiple routes, and considering that different routes may correspond to different driving styles (such as rapid acceleration, sharp turns, and smooth acceleration / deceleration), the system uses a comfort score (typically considering acceleration, jerk, and other metrics) to accurately select the smoothest route from multiple safe options that best aligns with human driving habits. This not only enhances the passenger experience but also makes the vehicle's driving behavior more human-like and easier for other road users to predict.

[0066] In one embodiment, the trajectory generation module can be specifically trained using a neural network. For example, it can be trained according to the following steps: First, obtain sample navigation semantic information of the sample vehicle, sample bird's-eye view BEV features, and the corresponding real driving trajectory generated by the human driver during driving. The sample navigation semantic information includes global trajectory trend information and / or local driving constraint information of the sample navigation trajectory. The sample BEV features are generated based on the sample vehicle's perceived surrounding environment data. Second, generate sample navigation semantic features based on the sample navigation semantic information, and use these features to perform feature filtering and fusion processing on the sample BEV features to suppress redundant information and enhance target features related to the sample navigation semantic information. Then, input the processed sample semantic fusion features into the neural network model to be trained, and obtain the sample planned driving trajectory of the sample vehicle output by the model. Adjust the internal parameters of the neural network model based on the deviation between the real driving trajectory and the sample planned driving trajectory. Repeat the above steps until a preset stopping condition is met, at which point training stops, resulting in a trained trajectory generation module.

[0067] Therefore, the entire training process is a supervised training loop of continuous trial and error and optimization, and each round of training can be divided into the following four stages: ① Data preparation stage Collect a large amount of historical data on real vehicles driving on actual roads. This data can include three core components: sample navigation semantic information, i.e., navigation instructions received by the sample vehicles (such as global curvature / slope trends, local speed limits / lane priorities, etc.); sample BEV features, i.e., environmental features perceived by the sample vehicles through sensors such as cameras and LiDAR (such as lane lines, obstacle positions, etc.); and real driving trajectories, i.e., the actual driving trajectories generated by human drivers in the current driving scenario (characterized by sample navigation semantic information and sample BEV features), serving as the "standard answer" for model learning.

[0068] ② Feature fusion processing stage The collected sample navigation semantic information is transformed into "sample navigation semantic features" that the system can understand, and this is used as a guide to filter and fuse sample BEV features. The purpose of this step is to enable the model to learn to focus on truly important information in the environment based on navigation intent, just like a human driver (e.g., focusing on the left-hand intersection when turning left, and focusing on the target lane when changing lanes), while automatically filtering out irrelevant background noise.

[0069] ③ Trajectory prediction and deviation calculation stage The processed "sample semantic fusion features" are input into the neural network model to be trained, and the model will output a "sample planned driving trajectory" (i.e., the current prediction result of the model). Subsequently, the system will calculate the deviation between this predicted trajectory and the "real driving trajectory" of the human driver (in deep learning, this is usually reflected as a loss function, such as Manhattan distance L1 or Euclidean distance L2).

[0070] ④ Parameter Iterative Optimization Stage Based on the calculated deviation, the backpropagation algorithm is used to adjust the internal neural network parameters (weights and biases) of the neural network model. The core logic is: if the trajectory predicted by the model deviates from the actual trajectory, the internal parameters are fine-tuned using mathematical methods so that the model can predict the actual trajectory more closely when it encounters a similar situation in the future.

[0071] Repeat the above process until the model converges, that is, continuously repeat the above steps, using massive sample data to iteratively train the trajectory generation module until the stopping condition is met (such as the deviation between the model's predicted trajectory and the real trajectory is reduced to a preset range, or the model's performance parameters reach the expected standard, etc.). At this point, the training can be ended—the neural network model at this time can be used as the trajectory generation module in the driver assistance system.

[0072] This training method endows the model with strong anthropomorphic driving capabilities. By directly using the real driving trajectories of human drivers as supervision signals (standard answers), the model learns not only simple geometric paths, but also high-order features and implicit knowledge containing human driving experience, comfort preferences, and the ability to cope with complex road conditions. This makes the final generated trajectory smoother, more natural, and highly similar to the driving style of a skilled human driver. Furthermore, it achieves end-to-end deep alignment between intent and environment: during this training process, if the model cannot accurately align navigation semantics (e.g., "speed limit ahead 60") with BEV features (e.g., "speed limit sign detected"), its output trajectory will have a significant deviation, thus being "penalized" during subsequent parameter adjustments. Therefore, the model can learn to "filter environmental information based on navigation intent" during the feature fusion stage. This mechanism greatly improves the model's logical reasoning and feature extraction capabilities in complex scenarios. In addition, this method is highly adaptive and can avoid hard-coded rules. This data-driven training method allows the model to automatically mine the processing patterns of various long-tail scenarios (corner cases) from massive amounts of data. Whether dealing with sudden changes in traffic signs or complex intersection games, the model can continuously evolve through training, possessing stronger generalization ability and robustness.

[0073] Regarding the trajectory generation method proposed in the aforementioned embodiments, the following will be combined with... Figure 3 The plan will be explained in detail. Figure 3 This is a flowchart illustrating another trajectory generation method for assisted driving, as shown in an embodiment of the present invention. Figure 3 The complete process shown can be divided into three core stages: navigation information preprocessing, feature processing, trajectory generation, and final planning output, which will be explained in turn below.

[0074] Phase 1: Navigation Information Preprocessing As the starting point of the entire process, the system needs to obtain the vehicle's real-time navigation information, such as obtaining the original navigation instructions from the navigation system (e.g., "turn left 500 meters ahead", "keep going straight").

[0075] Next, the navigation information is preprocessed. For example, the original navigation instructions can be broken down into two dimensions: global trajectory trend information (focusing on the macro path over a long distance, such as the overall direction of the road, curvature trend or slope change, etc.) and local driving constraint information (focusing on specific rules within a short distance, such as the speed limit of the current lane, traffic light status, intersection type, etc.).

[0076] For example, the extracted global trajectory trend information is denoted as Pnavi, with dimensions B×Nnavi×4; Nnavi is the number of path points in a single batch, and 4 corresponds to the (x,y,z,k) of the path points, where x and y represent the two-dimensional coordinates on the horizontal plane (e.g., the longitudinal and lateral positions in the map coordinate system or the vehicle's local coordinate system), z represents the height or elevation information (the z-axis is introduced to adapt to sloping road surfaces, such as uphill, downhill, and overpasses, in order to achieve 3D path planning, not just 2D planar path planning), and k represents the heading angle. The extracted local driving constraint information is denoted as Gconstraint, with dimensions B×Dconstraint, where Dconstraint is the local constraint dimension, including steering commands, speed limit encoding vectors, etc.

[0077] Preprocessing operations can include global path normalization and local constraint normalization. Global path normalization can resample each batch of Pnavi to retain the N valid path points; local constraint normalization can map continuous variables such as speed limits in Gconstraint to the [0,1] interval, and convert steering commands into one-hot vectors, such as straight = [1,0,0,0,0], left turn = [0,1,0,0,0].

[0078] Phase Two: Feature Processing Process of the Feature Processing Module The feature filtering module is mainly responsible for deeply integrating navigation intent with environmental perception, that is, generating navigation semantic features based on navigation semantic information, and using these features to filter the vehicle's BEV features.

[0079] First, feature extraction is performed: on the one hand, the "global trajectory trend information" is encoded and transformed into a high-dimensional feature vector that the system can understand (i.e., the aforementioned global trajectory feature vector); on the other hand, the "local driving constraint information" is encoded and key constraint features are extracted (i.e., the aforementioned local constraint feature vector). Then, feature fusion processing (plus sign icon) is performed, such as concatenating or adding the above two types of features to form a preliminary navigation feature representation.

[0080] Then, a linear transformation is performed on the fused navigation feature representation (such as a fully connected layer) to map it to a specific feature space, generating standardized navigation semantic features.

[0081] For example, following the foregoing embodiments, when extracting the global trajectory feature vector, global pooling is performed on Pnavi to extract the global position and curvature feature vector Fnavi-global for each batch:

[0082] The global average pooling operation is as follows: for each batch of resampled path points {P1,P2,...,PN} (where Pi=(xi,yi,zi,ki)), the arithmetic mean is calculated according to the four feature dimensions to obtain four basic feature values:

[0083] The above results are concatenated and then mapped and expanded to the Dglobal dimension through a fully connected layer, finally obtaining the global trajectory feature vector Fnavi-global for each batch, with a dimension of B×Dglobal.

[0084] During local constraint feature vector extraction, the preprocessed local navigation constraint vector Gconstraint is input into the fully connected layer, and high-dimensional embedding is completed through nonlinear transformation to obtain the local constraint feature vector Fcons-local with a fixed Dlocal dimension.

[0085] The Fcons-local dimension is B×Dlocal, where Dlocal is the dimension of the local constraint features.

[0086] During feature concatenation and linear transformation, the high-dimensional mapped Fcons-local is concatenated with the global trajectory feature Fnavi-global, and then linearly transformed to generate Lq navigation semantics Qnavi:

[0087] The Qnavi dimension is B×Lq×Dq, where Lq is the number of queries and Dq is the single query dimension.

[0088] Next, a cross-attention mechanism is used for feature selection and fusion: First, the vehicle's BEV features are acquired and attention weights are calculated. For example, the "navigation semantic features" generated in the previous step are used as a query vector to "ask" the input BEV features (i.e., the bird's-eye view features of the surrounding environment perceived by the vehicle). By calculating similarity, it can be determined which areas in the environment (such as intersections and lane lines) are relevant to the current navigation intent, and which areas are irrelevant (or have weak relevance). Then, feature weighted fusion is performed, that is, the BEV features are weighted and summed according to the calculated weights. This step can suppress irrelevant redundant information in the environment (such as roadside trees and irrelevant buildings) and strengthen target features related to the driving task, ultimately obtaining semantic fusion features.

[0089] For example, following the foregoing embodiments, when acquiring BEV features, BEV features directly generated based on surrounding environment data or BEV features Xanchor of multi-navigation-autonomous vehicle anchoring can be acquired, with dimensions of B×T×(C'+3)×H'×W', where T is the number of time frames, C'+3 is the number of BEV channels, and H'×W' is the spatial size.

[0090] Then it is flattened into a key-value vector, that is, Xanchor is flattened into a key (K) and value (V) vector, retaining the batch and time-series dimensions:

[0091] Its dimension is B×(T×H'×W')×(C'+3), where (T×H'×W') represents the total number of spatiotemporal locations of this feature.

[0092] When calculating attention weights, scaled dot product attention is used, with Qnavi as the query vector, and weights are calculated with K:

[0093] The dimension is B×Lq×(T×H'×W'), and the larger the weight value, the stronger the correlation between the corresponding spatial location and the navigation task.

[0094] When performing feature weighted fusion, weights are applied to V to obtain the fused feature Xweighted=Wattn×V corresponding to multiple queries, with dimensions B×Lq×(C'+3).

[0095] Finally, the features are subjected to dimensionality reduction and standardization, such as compressing the fused high-dimensional features (e.g., using 1×1 convolution or pooling), removing redundant dimensions, and standardizing their format. This step is to adapt to the input requirements of downstream modules and output task-adapted features.

[0096] During dimensionality reduction, the number of feature channels is mapped to the Dfused dimension using 1×1 convolution:

[0097] Its dimensions are B×Lq×Dfused; When standardizing the output, the Xfused layer is normalized to obtain the final simplified fusion feature Xfinal=LayerNorm(Xfused), which has the dimension B×Lq×Dfused.

[0098] Phase 3: Trajectory Generation and Output The trajectory generation module receives "task adaptation features" as input. As a trained neural network model, this module can plan the vehicle's future driving path based on environmental features that incorporate navigation intent, and output the planned driving trajectory downstream. This trajectory can actually be a sequence of trajectory points arranged in a temporal order, where each trajectory point can specifically be (x, y, z, k).

[0099] Subsequently, downstream modules of the trajectory generation module (such as the driving control module, steering control module, power module, etc. located downstream of the trajectory generation module in the driver assistance system) can control the vehicle to drive along the planned driving trajectory as described above.

[0100] visible, Figure 3 The process describes a closed-loop logic of "intention guiding perception, perception assisting decision-making": the system first parses the navigation intention, then uses the intention to filter out key information from complex environmental perception data, and finally generates a driving trajectory that conforms to human driving habits based on this concise and high-value information.

[0101] Corresponding to the above-described embodiments of the assisted driving trajectory generation method, the present invention also provides an embodiment of an assisted driving trajectory generation device.

[0102] Please see Figure 4 , Figure 4 This is an exemplary embodiment illustrating the hardware structure of an electronic device. At the hardware level, the device includes a processor 402, an internal bus 404, a network interface 406, memory 408, and non-volatile memory 410, and may also include other necessary hardware. One or more embodiments of the present invention can be implemented in software, for example, the processor 402 reads the corresponding computer program from the non-volatile memory 410 into memory 408 and then runs it. Of course, besides software implementation, one or more embodiments of the present invention do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0103] Please see Figure 5 , Figure 5This is a block diagram illustrating an assisted driving trajectory generation device as an exemplary embodiment. The device 500 can be applied to, for example... Figure 4 The illustrated electronic device is used to implement the technical solution of the present invention. The device includes: The acquisition unit 501 is used to acquire the navigation semantic information and bird's-eye view BEV features of the vehicle. The navigation semantic information includes global trajectory trend information and / or local driving constraint information of the navigation trajectory. The BEV features are generated based on the surrounding environment data perceived by the vehicle. The filtering unit 502 is used to generate navigation semantic features based on the navigation semantic information, and to use the navigation semantic features to perform feature filtering and fusion processing on the BEV features, so as to suppress redundant information in the BEV features and enhance the target features related to the navigation semantic information. The planning unit 503 is used to input the processed semantic fusion features into the trajectory generation module to generate the planned driving trajectory of the vehicle.

[0104] Optionally, the acquisition unit 501 is specifically used for: When the navigation semantic information only includes the global trajectory trend information, feature extraction and embedding mapping are performed on the global trajectory trend information, and the resulting global trend feature vector is used as the navigation semantic feature. When the navigation semantic information only includes the local driving constraint information, feature extraction and embedding mapping are performed on the local driving constraint information, and the resulting local constraint feature vector is used as the navigation semantic feature. When the navigation semantic information includes the global trajectory trend information and the local driving constraint information, feature extraction and embedding mapping are performed on the global trajectory trend information and the local driving constraint information respectively to obtain the corresponding global trajectory feature vector and local constraint feature vector; the global trajectory feature vector and the local constraint feature vector are fused to obtain the navigation semantic features.

[0105] Optionally, the acquisition unit 501 is specifically used for: Obtain path curvature information and / or path slope information from the global trajectory trend information, and map this information into the global trend feature vector through the first embedding layer; The lane priority information, speed limit information, and / or traffic sign semantic information in the local driving constraint information are obtained, and the information is mapped into the local constraint feature vector through the second embedding layer.

[0106] Optionally, the filtering unit 502 is specifically used for one of the following: The global trajectory feature vector and the local constraint feature vector are concatenated in the channel dimension, and the concatenated vector is linearly mapped and nonlinearly transformed through a preset fully connected layer. When the dimensions of the global trajectory feature vector and the local constraint feature vector are the same, element-wise addition or element-wise subtraction is performed on the two. The dynamic weight coefficients of the global trajectory feature vector and the local constraint feature vector are calculated using a self-attention mechanism or a pre-defined gating network, and the two are then weighted and summed based on the dynamic weight coefficients.

[0107] Optionally, the filtering unit 502 is specifically used for: Using the navigation semantic features as query terms and the BEV features as key-value pairs, the BEV features are subjected to feature filtering and fusion processing through a cross-attention mechanism.

[0108] Optionally, the filtering unit 502 is specifically used for: A query vector is obtained by performing a linear transformation on the navigation semantic features, and a key vector and a value vector are obtained by performing a linear transformation on the BEV features, respectively. Calculate the similarity score between the query vector and the key vector, and normalize the similarity score to obtain the attention weight matrix; The semantic fusion features are obtained by weighting and summing the value vectors based on the attention weight matrix.

[0109] Optionally, the planning unit 503 is specifically used for: The preset standardization compression module is invoked to perform channel dimension compression and format standardization on the semantic fusion features, and the processed task adaptation features are input into the trajectory generation module. The standardized compression module is a low-dimensional convolutional layer, pooling layer, feature selection algorithm module, or lightweight neural network layer.

[0110] Optionally, a training unit 504 is also included for training the trajectory generation module through the following steps: The sample navigation semantic information, sample bird's-eye view BEV features, and the actual driving trajectory generated by the human driver during the driving process of the sample vehicle are obtained. The sample navigation semantic information includes the global trajectory trend information and / or the local driving constraint information of the sample navigation trajectory. The sample BEV features are generated based on the sample surrounding environment data perceived by the sample vehicle. Based on the sample navigation semantic information, sample navigation semantic features are generated, and the sample BEV features are used to perform feature filtering and fusion processing on the sample BEV features in order to suppress redundant information in the sample BEV features and enhance the target features related to the sample navigation semantic information. The processed sample semantic fusion features are input into the neural network model to be trained, and the sample planned driving trajectory of the sample vehicle is obtained from the output of the model; the internal parameters of the neural network model are adjusted based on the deviation between the actual driving trajectory and the sample planned driving trajectory. Repeat the above steps until the preset stopping condition is met, at which point training stops and the trained trajectory generation module is obtained.

[0111] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0112] Accordingly, the present invention also provides a vehicle equipped with an environmental perception device and a driver assistance system, which is used to implement the driver assistance trajectory generation method as described in any of the above embodiments.

[0113] Accordingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the auxiliary driving trajectory generation method as described in any of the above embodiments.

[0114] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0115] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which can take the form of a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email sending and receiving device, game console, tablet computer, wearable device, or any combination of these devices.

[0116] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0117] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0118] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A trajectory generation method for assisted driving, characterized in that, A feature fusion module applied to a vehicle-deployed driver assistance system, the driver assistance system further comprising a trajectory generation module, the method comprising: The navigation semantic information and bird's-eye view BEV features of the vehicle are obtained. The navigation semantic information includes global trajectory trend information and / or local driving constraint information of the navigation trajectory. The BEV features are generated based on the surrounding environment data perceived by the vehicle. Based on the navigation semantic information, navigation semantic features are generated. Using the navigation semantic features as query terms and the BEV features as key-value pairs, the BEV features are filtered and fused through a cross-attention mechanism to suppress redundant information in the BEV features and enhance target features related to the navigation semantic information. The processed semantic fusion features are input into the trajectory generation module to generate the planned driving trajectory of the vehicle.

2. The method according to claim 1, characterized in that, The step of generating navigation semantic features based on the navigation semantic information includes: When the navigation semantic information only includes the global trajectory trend information, feature extraction and embedding mapping are performed on the global trajectory trend information, and the resulting global trend feature vector is used as the navigation semantic feature. When the navigation semantic information only includes the local driving constraint information, feature extraction and embedding mapping are performed on the local driving constraint information, and the resulting local constraint feature vector is used as the navigation semantic feature. When the navigation semantic information includes the global trajectory trend information and the local driving constraint information, feature extraction and embedding mapping are performed on the global trajectory trend information and the local driving constraint information respectively to obtain the corresponding global trajectory feature vector and local constraint feature vector; the global trajectory feature vector and the local constraint feature vector are fused to obtain the navigation semantic features.

3. The method according to claim 2, characterized in that, The step of performing feature extraction and embedding mapping on the global trajectory trend information includes: obtaining path curvature information and / or path slope information in the global trajectory trend information, and mapping the information into the global trend feature vector through the first embedding layer; The feature extraction and embedding mapping process for the local driving constraint information includes: obtaining lane priority information, speed limit information and / or traffic sign semantic information from the local driving constraint information, and mapping the information into the local constraint feature vector through a second embedding layer.

4. The method according to claim 2, characterized in that, The fusion processing of the global trajectory feature vector and the local constraint feature vector includes one of the following: The global trajectory feature vector and the local constraint feature vector are concatenated in the channel dimension, and the concatenated vector is linearly mapped and nonlinearly transformed through a preset fully connected layer. When the dimensions of the global trajectory feature vector and the local constraint feature vector are the same, element-wise addition or element-wise subtraction is performed on the two. The dynamic weight coefficients of the global trajectory feature vector and the local constraint feature vector are calculated using a self-attention mechanism or a pre-defined gating network, and the two are then weighted and summed based on the dynamic weight coefficients.

5. The method according to claim 1, characterized in that, The feature selection and fusion process for the BEV features using a cross-attention mechanism includes: A query vector is obtained by performing a linear transformation on the navigation semantic features, and a key vector and a value vector are obtained by performing a linear transformation on the BEV features, respectively. Calculate the similarity score between the query vector and the key vector, and normalize the similarity score to obtain the attention weight matrix; The semantic fusion features are obtained by weighting and summing the value vectors based on the attention weight matrix.

6. The method according to claim 1, characterized in that, The step of inputting the processed semantic fusion features into the trajectory generation module includes: The preset standardization compression module is invoked to perform channel dimension compression and format standardization on the semantic fusion features, and the processed task adaptation features are input into the trajectory generation module. The standardized compression module is a low-dimensional convolutional layer, pooling layer, feature selection algorithm module, or lightweight neural network layer.

7. The method according to claim 1, characterized in that, Also includes: When there is only one planned driving trajectory, the driver assistance system is used to control the vehicle to drive according to that planned driving trajectory; When there are multiple planned driving trajectories, the safety and / or comfort scores are calculated for each of the multiple planned driving trajectories, and the driver assistance system is used to control the vehicle to drive according to the planned driving trajectory with the highest score.

8. The method according to any one of claims 1-7, characterized in that, The trajectory generation module is trained through the following steps: The sample navigation semantic information, sample bird's-eye view BEV features, and the actual driving trajectory generated by the human driver during the driving process of the sample vehicle are obtained. The sample navigation semantic information includes the global trajectory trend information and / or the local driving constraint information of the sample navigation trajectory. The sample BEV features are generated based on the sample surrounding environment data perceived by the sample vehicle. Based on the sample navigation semantic information, sample navigation semantic features are generated. Using the navigation semantic features as query terms and the BEV features as key-value pairs, the BEV features are filtered and fused through a cross-attention mechanism to suppress redundant information in the sample BEV features and enhance target features related to the sample navigation semantic information. The processed sample semantic fusion features are input into the neural network model to be trained, and the sample planned driving trajectory of the sample vehicle is obtained from the output of the model; the internal parameters of the neural network model are adjusted based on the deviation between the actual driving trajectory and the sample planned driving trajectory. Repeat the above steps until the preset stopping condition is met, at which point training stops and the trained trajectory generation module is obtained.

9. A driver assistance system, characterized in that, The driver assistance system includes a feature fusion module and a trajectory generation module, wherein, The feature fusion module is used to: acquire vehicle navigation semantic information and bird's-eye view BEV features, wherein the navigation semantic information includes global trajectory trend information and / or local driving constraint information of the navigation trajectory, and the BEV features are generated based on the surrounding environment data perceived by the vehicle; generate navigation semantic features based on the navigation semantic information, and use the navigation semantic features as query terms and the BEV features as key-value pairs to perform feature filtering and fusion processing on the BEV features through a cross-attention mechanism to suppress redundant information in the BEV features and enhance target features related to the navigation semantic information; and input the processed semantic fusion features into the trajectory generation module. The trajectory generation module is used to generate the planned driving trajectory of the vehicle based on the semantic fusion module.

10. A vehicle, characterized in that, It is equipped with an assisted driving system for implementing the method as described in any one of claims 1-8.

11. An electronic device, characterized in that, It includes a communication interface, a processor, a memory, and a bus, wherein the communication interface, the processor, and the memory are interconnected via the bus; The memory stores machine-readable instructions, and the processor implements the method as described in any one of claims 1-8 by executing the machine-readable instructions.

12. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-readable instructions, which, when invoked and executed by a processor, implement the method of any one of claims 1-8.

Images