Intelligent driving method and device, and storage medium

Abstract

This application relates to the field of intelligent driving, and more specifically, to an intelligent driving method, apparatus, and computer storage medium based on an end-to-end model. The method includes: acquiring perception data from sensors and navigation information from a navigation system; encoding the perception data and navigation information to generate multiple tokens, wherein each token is represented by tensors of multiple types; concatenating tensors of the same type from the multiple tokens to generate a concatenated tensor of multiple types; and providing the concatenated tensor of multiple types to an end-to-end model to generate instructions for controlling vehicle driving.

Description

Technical Field

[0001] This application relates to the field of intelligent driving, and more specifically, to an intelligent driving method, apparatus, and computer storage medium based on an end-to-end model. Background Technology

[0002] In intelligent driving solutions, end-to-end intelligent driving models are receiving increasing attention. Compared to traditional modular pipelines (e.g., modules including perception, prediction, planning, and control), end-to-end models can directly map various raw data into commands for controlling vehicle movement, thus completing the entire process of decision-making within a single model. This approach avoids information loss and error accumulation between modules and possesses zero-shot learning capabilities, enabling it to better handle various driving scenarios.

[0003] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] This application provides an intelligent driving method, an intelligent driving device implementing the method, and a computer storage medium. The method can fully utilize navigation information from Standard Definition (SD) maps through a standardized token architecture without requiring excessive modifications to the end-to-end model architecture. Therefore, the method can reduce redundant feature learning by the model, reduce model computation, and shorten decision latency.

[0005] According to a first aspect of this application, an intelligent driving method is provided, the method comprising: acquiring perception data from sensors and navigation information from a navigation system; encoding the perception data and the navigation information to generate a plurality of tokens, wherein each token is represented by tensors of a plurality of types; concatenating tensors of the same type among the plurality of tokens to generate a concatenated tensor of a plurality of types; and providing the concatenated tensor of a plurality of types to an end-to-end model to generate instructions for controlling vehicle driving.

[0006] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, the plurality of tensors include: polygon tensors, polyline tensors, mask tensors, and property tensors.

[0007] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, the navigation information includes: road topology information, guidance information, and lane information.

[0008] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, the road topology information includes a main path and sub-paths associated with the main path; the guidance information includes a main action and an assistant action associated with the main action; the lane information includes lane type, lane location number, whether it is a recommended lane, the driving direction of the recommended lane, and the distance to the next lane segment.

[0009] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, encoding the road topology information to generate a road token includes: encoding the road topology information into a polyline tensor and an attribute tensor of the road token; and using a mask tensor of the road token to indicate that the polygon tensor of the road token is invalid while the polyline tensor is valid.

[0010] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, encoding the road topology information to generate a road token further includes: interpolating the shape points on the main path and branch paths in the road topology information to match the dimension of the polyline tensor.

[0011] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, encoding the inducement information to generate an action token includes: encoding the inducement information into the attribute tensor of the action token; and setting the values ​​in the polygon tensor, polyline tensor, and mask tensor of the action token as fill values.

[0012] As an alternative or supplement to the above solutions, in a method according to an embodiment of this application, encoding the lane information of each lane on the road to generate a corresponding lane token includes: encoding the lane information into the attribute tensor of the lane token; and setting the values ​​in the polygon tensor, polyline tensor, and mask tensor of the lane token as fill values.

[0013] According to a second aspect of this application, an intelligent driving device is provided, the device comprising: a memory; a processor; and a computer program stored in the memory and executable by the processor, the execution of the computer program causing the intelligent driving method according to the first aspect of this application to be performed.

[0014] According to a third aspect of this application, a computer storage medium is provided, the computer storage medium including instructions that, when executed, perform the intelligent driving method according to the first aspect of this application. Attached Figure Description

[0015] The above and / or other aspects and advantages of this application will become clearer and more readily understood from the following description taken in conjunction with the accompanying drawings, in which the same or similar elements are denoted by the same reference numerals. In the drawings: Figure 1 This is a flowchart of an intelligent driving method 100 according to an embodiment of this application; Figure 2 These are some exemplary illustrations of road topology information in illustrative navigation information according to embodiments of this application; Figure 3 These are some exemplary illustrations of lane information in illustrative navigation information according to embodiments of this application; Figure 4 This is a schematic block diagram of an intelligent driving data stream 400 according to an embodiment of this application; and Figure 5 This is a block diagram of an intelligent driving device 50 according to an embodiment of this application. Detailed Implementation

[0016] The following detailed description is merely exemplary in nature and is not intended to limit the disclosed technology or its application and use. Furthermore, it is not intended to be bound by any express or implied theory presented in the foregoing technical fields, background art, or the following detailed description.

[0017] In the following detailed description of the embodiments, numerous specific details are set forth in order to provide a more thorough understanding of the disclosed technology. However, it will be apparent to those skilled in the art that the disclosed technology can be practiced without these specific details. In other instances, well-known features have not been described in detail to avoid unnecessarily complicating the description.

[0018] Terms such as "possessing" and "comprising" indicate that, in addition to the units (modules) and steps directly and explicitly stated in the specification and claims, the technical solution of this application does not exclude the presence of other units (modules) and steps not directly or explicitly stated. Terms such as "first" and "second" do not indicate the order of the units in terms of time, space, size, etc., but are merely used to distinguish the units. Furthermore, the steps in this document are not limited to being performed in the order they are written; a step written later may be performed simultaneously with or before a step written earlier.

[0019] In intelligent driving technology, navigation information is an indispensable core support. Just as humans rely on navigation guidance when driving in unfamiliar environments, intelligent driving vehicles cannot obtain global path cognition and lane-level driving guidance by relying solely on real-time sensor perception in unfamiliar road scenarios. Furthermore, solutions that rely solely on sensors to determine vehicle control commands are easily affected by factors such as adverse weather (e.g., rain, snow, fog), road obstructions, and sensor noise, and lack global path constraints and prior road knowledge. This leads to problems such as decision-making bias, control lag, lane selection errors, or path deviations in complex scenarios (such as turning at intersections, changing lanes in multiple lanes, merging and exiting at high speeds, switching between complex road networks, and driving on unfamiliar roads).

[0020] However, the presentation or output format of existing navigation information differs significantly from that of perceived data. Therefore, in order for an end-to-end model (e.g., a deep learning neural network model) to utilize both types of input to generate vehicle control commands, it may be necessary to make it significantly modified and / or enable it to understand the features in the navigation information. This may further increase the computational load and / or decision delay.

[0021] The embodiments of this application will now be described in detail with reference to the accompanying drawings. Figure 1 , Figure 1 This is a flowchart of an intelligent driving method 100 according to an embodiment of this application. In step 102, perception data from sensors and navigation information from a navigation system are acquired. As an example, the sensors may include multiple cameras on the vehicle, LiDAR, speed sensors, temperature sensors, etc. Accordingly, the perception data may be visual data from cameras, point cloud data from radar, vehicle speed, ambient temperature, etc. This perception data can help the vehicle's intelligent driving model to identify the vehicle's current location, the surrounding environment, and objects in the environment (e.g., dynamic and static objects).

[0022] In some embodiments, the navigation system may be a lane-level navigation system (instead of a high-precision navigation system), and the navigation information it can provide may include road topology information, guidance information, and lane information. Therefore, the information provided by the navigation system will not consume excessive communication and computing resources.

[0023] Furthermore, in some embodiments, the road topology information may include the main path and branch paths associated with the main path. (See reference...) Figure 2 , Figure 2 These are some exemplary illustrations of road topology information in illustrative navigation information according to embodiments of this application. Specifically, Figure 2 The diagrams show the intersections ( Figure 2(left side), two-way intersection ( Figure 2 The middle part) and the three-way intersection ( Figure 2 The right side of the diagram shows the main path and branch paths, where the main path is represented by a solid line and the branch paths by dashed lines.

[0024] For example, a navigation system can instruct a vehicle to turn right at the next intersection. Therefore, as... Figure 2 As shown in the left-hand section, the path to turn right at the next intersection is the main path (marked as "mainpath"), represented by solid lines, while the paths to go straight through the intersection and then turn left at that intersection are two branching paths (marked as "subpath"), represented by dashed lines. As another example, a navigation system could instruct a vehicle to follow the main road (or the road to the left) at the next two-way intersection, thus... Figure 2 As shown in the middle section, the path along the main road (or the road to the left) at the two-way intersection is represented by solid lines, while the path to the auxiliary road (or the road to the right) at the same intersection is represented by dashed lines. In another example, the navigation system could instruct the vehicle to follow the middle road at the next three-way intersection, thus... Figure 2 As shown in the right-hand section, the path along the middle road at this three-way intersection is the main path, represented by solid lines, while the two paths along the left and right roads at the same intersection are the branching paths, represented by dashed lines. It should be understood that... Figure 2 This is for illustrative purposes only, and not for limiting purposes, regarding the main path and branch paths. Figure 2 Based on this, those skilled in the art should be able to understand the corresponding distribution of main paths and branch paths in other scenarios (such as roundabouts, merging ramps, auxiliary roads, etc.).

[0025] In some embodiments, the guidance information in the navigation information includes a primary action and an auxiliary action associated with the primary action. Specifically, the guidance information provided by the navigation system may consist of two parts: a primary action and an auxiliary action. The primary action provides an overall direction in a concise form, while the auxiliary action associated with the primary action can supplement or explain the primary action.

[0026] In some embodiments, the active action may include: going straight, driving forward to the left, driving forward to the right, turning left, turning right, making a U-turn, etc., while the auxiliary action may include: maintaining the current road, entering a ramp, entering an auxiliary road, using the middle / left / right lane, entering a dedicated left / right turn lane, etc. For example, if the guidance information provided by the navigation system is "drive forward to the right and enter the ramp," then the active action is "drive forward to the right," and the auxiliary action is "enter the ramp." As another example, if the guidance information is "keep to the left and drive along the main road," then the active action is "keep to the left," and the auxiliary action is "drive along the main road." In this way, the active action can serve as an unwavering core directional guide to provide macroscopic driving direction constraints; the auxiliary action can serve as detailed guidance instructions to precisely lock the desired driving path, thereby avoiding lane selection errors or path deviations in unfamiliar road networks.

[0027] In some embodiments, lane information in the navigation information includes lane type, lane location number, whether it is a recommended lane, the direction of travel of the recommended lane, and the distance to the next lane segment. (See reference) Figure 3 , Figure 3 These are some exemplary illustrations of lane information in illustrative navigation information according to embodiments of this application. Figure 3 The diagram illustrates a portion of the lane information provided by the navigation system in two typical scenarios.

[0028] Figure 3 The left side of the image illustrates a portion of the lane information provided by the navigation system when the vehicle is traveling on a three-lane road. Specifically, as shown... Figure 3 The left-hand diagram illustrates that the current three lanes are all ordinary lanes (not bus lanes, high-occupancy vehicle lanes, or similar special lanes); the first and second lanes from the left (numbered 1 and 2, not shown in the diagram) are for straight travel, and the third lane from the left (numbered 3, not shown in the diagram) is for either straight travel or a right turn; all three lanes are recommended lanes (as shown, displayed on a white base with another color overlaid, for example, gray); and the recommended lanes are for straight travel. However, it should be understood that in other embodiments, lane numbering may also proceed from right to left.

[0029] Figure 3 The right side of the diagram illustrates a portion of the lane information provided by the navigation system when the vehicle is traveling on another three-lane road. For example... Figure 3The right-hand side diagram shows the lane information provided, indicating the following: all three lanes are currently ordinary lanes; the first lane (numbered 1) from the left is for left turns, the second lane (numbered 2) is for straight travel, and the third lane (numbered 3) is for both straight travel and right turns; the second and third lanes from the left are recommended lanes, while the first lane is not; and the recommended lane is for straight travel.

[0030] In some embodiments, the lane information provided may also include an indication of lane changes ahead (e.g., "no change" as shown in the figure) and the distance to the next lane segment.

[0031] As mentioned above Figure 2 and Figure 3 As described, this navigation information (e.g., road topology, navigation guidance information, and lane information) provides intelligent driving vehicles with the road-related prior information they need, thus improving the accuracy and speed of driving decisions. Specifically, navigation guidance information can provide clear destination orientation, adapting to the need for path guidance in unfamiliar environments; lane information can provide precise lane positioning and lane selection constraints, avoiding lane misselection on unfamiliar roads; and road topology information can construct global path cognition and local driving constraints to ensure that the final output control commands conform to navigation intent, adapt to road conditions, and guarantee driving safety.

[0032] return Figure 1 In step 104, the perceived data and navigation information are encoded to generate multiple tokens, each represented by tensors of multiple types. Specifically, to enable the end-to-end model to easily utilize the perceived data and navigation information, these data need to be encoded into tokens. In the field of intelligent driving, a token is the smallest semantic / feature unit of raw information and serves as the fundamental data carrier for the intelligent driving model to understand and compute. To balance the semantic expression and computational efficiency of tokens, a suitable tensor structure needs to be designed so that tokens can express input information from different sources, with different content and functions (e.g., perceived data from sensors and navigation information from navigation systems) in a unified form.

[0033] In some embodiments, tensors representing multiple types of tokens may include: polygon tensors, polyline tensors, mask tensors, and attribute tensors. As an example, each polygon / polyline tensor may be represented by multiple points (e.g., four points), and each point may consist of coordinates (e.g., two-dimensional coordinates). A mask tensor may consist of a two-dimensional vector to indicate the validity of the polygon and polyline tensors of the token. Attribute tensors may include attribute values ​​across multiple dimensions, such as a primary attribute, one or more sub-attributes, an identifier, a timestamp, etc. Based on the combination of these tensors, the environment and objects within it, indicated by perception data and navigation information, can be understood by the end-to-end model. The process of encoding perception data and navigation information into these tensors is described in detail below.

[0034] In some embodiments, the sensing data can indicate the size and location of objects in the environment, which may include dynamic objects (e.g., surrounding motor vehicles, pedestrians, bicycles, and the vehicle itself) and static objects (e.g., medians, buildings, etc.). The number of tokens representing the object can be determined depending on the duration of the object's presence in the sensing data (including historical and current sensing data) and the sampling period of the sensing data. Information related to these objects (e.g., location, size) can be encoded into polygon tensors and attribute tensors of the object tokens. Specifically, the size and location of the object can be determined based on the coordinates of multiple corner points, which can then be encoded into multiple points in the polygon tensor; the object's type, identifier, and time of appearance can correspondingly correspond to relevant attributes in the attribute tensor (e.g., a primary attribute for pedestrian, a sub-attribute for fill value, an identifier for the pedestrian, and a timestamp for the acquisition time of the corresponding sensing data). In addition, the polyline tensor of the object token is set to a padding value (e.g., 0), and the mask tensor of the object token is encoded to indicate that the polygon tensor of the object token is valid while the polyline tensor is invalid (e.g., the mask tensor is encoded in the form of [1,0]).

[0035] In some embodiments, road topology information in the navigation information can be encoded into a polyline tensor and an attribute tensor of the road token. Specifically, the main path and branch paths in the road topology information can be expressed as polylines and thus correspond to the polyline tensor. The attributes of the path (e.g., whether it is a main path or a branch path) can correspond to the main attributes in the attribute tensor. Furthermore, the mask tensor of the road token needs to be encoded to indicate that the polygon tensor of the road token is invalid while the polyline tensor is valid (e.g., encoded as [0,1]).

[0036] In some embodiments, it is also necessary to interpolate the shape points on the main path and branch paths in the road topology information to match the dimension defined by the polyline tensor. As an example, the main path used to maintain straight-line travel in the road topology information may only be represented by two shape points (e.g., two endpoints), while the polyline tensor is defined by four points. In this case, two additional points can be generated at 1 / 3 and 2 / 3 distances between the two shape points to match the dimension of the polyline tensor.

[0037] In some embodiments, the guiding information in the navigation information can be encoded into the attribute tensor of the action token. Specifically, one or more subsequent guiding information provided in the navigation information can correspond to attributes in the attribute tensor of one or more action tokens.

[0038] As an example, for a vehicle at a given moment, the next two navigation guidance messages are "Drive right ahead, enter the ramp" and "Keep left, follow the main road." These two navigation guidance messages can be encoded into two action tokens (e.g., encoded into their attribute tensors). For the first action token, in its attribute tensor, the primary attribute is encoded as the value corresponding to action number 1, the first sub-attribute indicating the primary action is encoded as the value corresponding to "Drive right," the second sub-attribute indicating the auxiliary action is encoded as the value corresponding to "Enter the ramp," and other sub-attributes can be encoded as filler values. For the second action token, in its attribute tensor, the primary attribute is encoded as the value corresponding to action number 2, the first sub-attribute indicating the primary action is encoded as the value corresponding to "Keep left," the second sub-attribute indicating the auxiliary action is encoded as the value corresponding to "Follow the main road," and other sub-attributes can be encoded as filler values. Furthermore, the values ​​in the polygon tensor, polyline tensor, and mask tensor of these two action tokens are all set to filler values.

[0039] In some embodiments, lane information in the navigation information can also be encoded into the attribute tensor of the lane token. Specifically, a lane token can be generated for each lane on the current road, and then the lane information related to that lane can be encoded into the attribute tensor of the lane token. Simultaneously, the values ​​in the polygon tensor, polyline tensor, and mask tensor of the lane token are all set to padding values. In some embodiments, to obtain prior information about the road ahead, the lane information of the currently traveling lane and the lane information of the next future road segment ahead can be obtained at each time step.

[0040] For a lane token, in the attribute tensor, the main attribute can be used to indicate the lane type, the first sub-attribute can be used to indicate the direction of travel of the lane, the second sub-attribute can be used to indicate whether the lane is a recommended lane, the third sub-attribute can be used to indicate the distance to the next lane segment, and the identifier can indicate the lane number (e.g., which lane from the left).

[0041] In some embodiments, lane information, along with perception data, can assist intelligent driving models in quickly identifying the current lane position and the distribution of adjacent lanes, enabling real-time adjustments to steering angle and driving posture to achieve lane-level yaw optimization. As an example, when driving on a long straight road, the center trajectory of the current recommended lane (e.g., lane 3) is locked based on the lane number in the lane information and perception data, preventing the vehicle from veering off to lane 2 on the left or lane 4 on the right. Furthermore, lane information can be combined with guidance information to facilitate the resolution of route and lane selection problems. For example, in situations such as... Figure 3 In the three-lane road illustrated on the left, when a vehicle is traveling in the left lane and the next navigation guidance information is "turn right at the next intersection," the intelligent driving model can first determine that only the rightmost lane, number 3, is the recommended lane. Then, it determines that the vehicle needs to complete two lane changes before the next intersection to move from lane 1 to lane 3. This helps the intelligent driving model determine the timing of each lane change more accurately and safely.

[0042] Continue to refer to Figure 1 In step 106, tensors of the same type from the multiple tokens are concatenated to generate concatenated tensors of multiple types. As described above, multiple tokens can be generated through step 104, including perception tokens, road tokens, action tokens, and lane tokens related to perception data (e.g., images, radar point clouds). Although these tokens have different data sources and indicate different types of information, they all have the same formal representation (e.g., polygon tensors, polyline tensors, mask tensors, and attribute tensors). Therefore, tensors of each type from all tokens can be concatenated along dimension 0 to obtain concatenated tensors of multiple types. This concatenation operation can facilitate adaptation to end-to-end model inputs and thus help reduce the complexity of the model architecture.

[0043] In step 108, multiple types of concatenated tensors are provided to the end-to-end model to generate commands for controlling vehicle movement. Specifically, based on the multiple types of concatenated tensors, the end-to-end model can understand the surrounding environment and the distribution of objects in the environment as indicated by perception data, and make appropriate judgments and generate control commands by combining the requirements in the navigation information. Thus, the intelligent driving method 100 can fully utilize perception data and navigation information without increasing structural complexity, thereby improving the robustness and environmental adaptability of the end-to-end model.

[0044] Next reference Figure 4 , Figure 4 This is a schematic block diagram of an intelligent driving data stream 400 according to an embodiment of this application. Figure 4 As illustrated in the diagram, perception data 402 from sensors and navigation information 404 from the navigation system are encoded into multiple tokens 406. The navigation information 404 may include road topology information 4041, guidance information 4042, and lane information 4043. Each token 406 can be represented by a polygon tensor 4061, a polyline tensor 4062, a mask tensor 4063, and an attribute tensor 4064. After concatenating tensors of the same type from all tokens, the concatenated tensors of multiple types are provided as input to the end-to-end model 408. The end-to-end model 408 generates instructions 410 for controlling the vehicle based on these inputs, which conform to the navigation intent, adapt to road conditions, and ensure driving safety.

[0045] Figure 5 This is a block diagram of an intelligent driving device 50 according to an embodiment of this application. The device 50 includes a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. The execution of the computer program 530 causes... Figure 1 The intelligent driving method 100 shown is executed.

[0046] Furthermore, as described above, this application can also be implemented as a computer storage medium storing a program for causing a computer to execute the methods described in any of the above embodiments. Here, various types of computer storage media can be used as computer storage media, such as disks (e.g., magnetic disks, optical disks, etc.), cards (e.g., memory cards, optical cards, etc.), semiconductor memory (e.g., ROM, non-volatile memory, etc.), and tapes (e.g., magnetic tape, cassette tape, etc.).

[0047] Where applicable, the various embodiments provided in this application may be implemented using hardware, software, or a combination of hardware and software. Furthermore, where applicable, without departing from the scope of this application, the various hardware and / or software described herein may be combined into composite components comprising software, hardware, and / or both. Where applicable and within the scope of this application, the various hardware and / or software components described herein may be divided into sub-components comprising software, hardware, or both. Additionally, where applicable, it is contemplated that software components may be implemented as hardware components, and vice versa.

[0048] The software (such as program code or data) according to this application can be stored on one or more computer storage media. It is also contemplated that the software identified herein can be implemented using one or more networked and / or otherwise general-purpose or special-purpose computers and / or computer systems. Where applicable, the order of the various steps described herein can be changed, combined into compound steps, and / or divided into sub-steps to provide the features described herein.

[0049] The embodiments and examples presented herein are provided to best illustrate embodiments of this application and its particular applications, thereby enabling those skilled in the art to implement and use this application. However, those skilled in the art will understand that the above description and examples are provided for ease of illustration and example only. The descriptions presented are not intended to cover all aspects of this application or to limit this application to the precise forms disclosed.

Claims

1. An intelligent driving method, characterized in that, The method includes: Acquire perception data from sensors and navigation information from navigation systems; The perception data and the navigation information are encoded to generate perception tokens and navigation tokens, each including a polygon tensor, a polyline tensor, a mask tensor, and an attribute tensor. The polygon tensor, polyline tensor, mask tensor, and attribute tensor in the perception token and the navigation token are concatenated along the 0th dimension to generate an input token, which includes the concatenated polygon tensor, polyline tensor, mask tensor, and attribute tensor; and The input token is provided to the end-to-end model to generate instructions for controlling vehicle movement.

2. The method as described in claim 1, wherein, The navigation information includes: road topology information, guidance information, and lane information, and the navigation token correspondingly includes: road token, action token, and lane token.

3. The method as described in claim 2, wherein, The road topology information includes the main path and the branch paths associated with the main path; the guidance information includes the main action and the auxiliary actions associated with the main action; the lane information includes lane type, lane location number, whether it is a recommended lane, the driving direction of the recommended lane, and the distance to the next lane segment.

4. The method of claim 2, wherein, Encoding the road topology information to generate a road token includes: The road topology information is encoded into the polyline tensor and attribute tensor of the road token; and Use the mask tensor of the road token to indicate that the polygon tensor of the road token is invalid while the polyline tensor is valid.

5. The method of claim 4, wherein, Encoding the road topology information to generate a road token further includes: interpolating the shape points on the main path and branch paths in the road topology information to match the dimension of the polyline tensor.

6. The method of claim 2, wherein, Encoding the inducement information to generate an action token includes: Encode the inducement information into the attribute tensor of the action token; and Set the values ​​in the polygon tensor, polyline tensor, and mask tensor of the action token to the fill value.

7. The method of claim 2, wherein, For each lane on the road, the lane information is encoded to generate a corresponding lane token, including: The lane information is encoded into the attribute tensor of the lane token; and Set the values ​​in the polygon tensor, polyline tensor, and mask tensor of the lane token to the fill value.

8. An intelligent driving device, characterized in that, The apparatus includes: a memory; a processor; and a computer program stored in the memory and executable by the processor, the execution of which causes the method as described in any one of claims 1-7 to be performed.

9. A computer storage medium, characterized in that, The computer storage medium includes instructions that, when executed, perform the method as described in any one of claims 1-7.

Images