Model optimization method, device and equipment based on memory parking, medium and product

Abstract

The application provides a memory parking-based model optimization method, device, equipment, medium and product. According to an example of the application, the method can include: collecting task execution data of a vehicle end during execution of a memory parking task in an underground garage and perception data associated with the task execution data; the task execution data includes garage identification of the underground garage, and the task execution data is used to determine a task execution index of the underground garage; the perception data is used to optimize a memory parking model in a case where the task execution index meets a preset optimization condition, the memory parking model is used to execute a memory parking task in the underground garage, and the memory parking task includes a mapping task and / or a cruising task; and the task execution data and the perception data are sent to a cloud end to achieve targeted model optimization for each underground garage, thereby improving execution stability of the memory parking function in different underground garages.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to model optimization methods, devices, equipment, media and products based on memory parking. Background Technology

[0002] With the rapid development of intelligent driving technology, memory parking, as a core function of intelligent driving, enables vehicles to learn and remember specific driving routes, automatically completing driving operations in the same underground parking garage, thereby improving the convenience of driving for users.

[0003] However, the structural layout, perception conditions, traffic characteristics, and other environmental features of different underground parking garages vary greatly, resulting in poor stability of the memory parking function in different underground parking garages. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this application provides a model optimization method, apparatus, equipment, medium and product based on memory parking.

[0005] According to a first aspect of any embodiment of this application, a model optimization method based on memory parking is provided, the method being applied to a vehicle, the method comprising: The system collects task execution data and associated perception data during the vehicle's memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage and is used to determine the task execution indicators of the underground garage. The perception data is used to optimize the memory parking model when the task execution indicators meet preset optimization conditions. The memory parking model is used to perform the memory parking task in the underground garage, and the memory parking task includes mapping tasks and / or cruising tasks. The task execution data and the perception data are sent to the cloud.

[0006] According to a second aspect of any embodiment of this application, a model optimization method based on memory parking is provided, the method being applied in the cloud, the method comprising: The system receives task execution data and perception data associated with the task execution data sent by the vehicle terminal. The task execution data and the perception data are collected by the vehicle terminal during the execution of a memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage. Based on the task execution data, the task execution indicators of the underground parking garage are determined; When the task execution indicators meet the preset optimization conditions, the memory parking model is optimized using the perception data. The memory parking model is used to perform memory parking tasks in the underground garage. The memory parking tasks include mapping tasks and / or cruising tasks.

[0007] According to a third aspect of any embodiment of this application, a model optimization device based on memory parking is provided, the device being applied to a vehicle, the device comprising: The data acquisition module is used to collect task execution data and perception data associated with the task execution data during the vehicle's memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage and is used to determine the task execution index of the underground garage. The perception data is used to optimize the memory parking model when the task execution index meets the preset optimization conditions. The memory parking model is used to perform the memory parking task in the underground garage, and the memory parking task includes mapping task and / or cruise task. The data sending module is used to send the task execution data and the sensing data to the cloud.

[0008] According to a fourth aspect of any embodiment of this application, a model optimization device based on memory parking is provided, the device being applied in the cloud, the device comprising: The data receiving module is used to receive task execution data sent by the vehicle terminal and perception data associated with the task execution data. The task execution data and the perception data are collected by the vehicle terminal during the execution of the memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage. The indicator determination module is used to determine the task execution indicators of the underground parking garage based on the task execution data. The model optimization module is used to optimize the memory parking model using the perception data when the task execution indicators meet the preset optimization conditions. The memory parking model is used to perform memory parking tasks in the underground garage. The memory parking task includes mapping tasks and / or cruising tasks.

[0009] According to a fifth aspect of any embodiment of this application, an electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor executes the executable instructions to implement the method as described in any of the embodiments of the first or second aspect above.

[0010] According to a sixth aspect of any embodiment of this application, a computer-readable storage medium is provided having computer instructions stored thereon that, when executed by a processor, implement the method described in any of the embodiments of the first or second aspect described above.

[0011] According to a seventh aspect of any embodiment of this application, a computer program product is provided having a computer program / instructions stored thereon, which, when executed by a processor, implement the method described in any of the embodiments of the first or second aspect described above.

[0012] The technical solution provided in this application may include the following beneficial effects: As can be seen from the above embodiments, by collecting task execution data and associated perception data during the vehicle's memory parking task in the underground garage, the task execution data and perception data are sent to the cloud. By taking the underground garage as the dimension, the task execution indicators are determined based on the task execution data collected by the vehicle during the memory parking task, and the memory parking model is optimized based on the perception data. This allows the optimized memory parking model to adapt to the corresponding underground garage to perform mapping tasks and / or cruising tasks, achieving targeted model optimization for each underground garage, thereby improving the execution stability of the memory parking function in different underground garages.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0014] The accompanying drawings, which are incorporated in and form part of this application, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0015] Figure 1 This application illustrates an architecture diagram of a memory parking system according to an exemplary embodiment. Figure 2 This is a flowchart illustrating a model optimization method based on memory parking according to an exemplary embodiment of this application; Figure 3 This is a schematic diagram of an intersection according to an exemplary embodiment of this application; Figure 4 This is a flowchart illustrating an intersection mapping and intersection patrol method according to an exemplary embodiment of this application; Figure 5 This is a schematic diagram illustrating a cruise process according to an exemplary embodiment of this application; Figure 6 This is a schematic diagram of the architecture of another memory parking system according to an exemplary embodiment of this application; Figure 7 This is a schematic diagram of a vehicle-side architecture according to an exemplary embodiment of this application; Figure 8 This is an interactive flowchart illustrating a model optimization method based on memory parking according to an exemplary embodiment of this application; Figure 9 This is a flowchart illustrating another memory-based parking model optimization method according to an exemplary embodiment of this application; Figure 10 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of this application; Figure 11 This is a block diagram illustrating a memory-based parking model optimization device according to an exemplary embodiment of this application; Figure 12 This is a block diagram of another memory-based parking model optimization device illustrated in an exemplary embodiment of this application. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0017] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0018] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0019] In related technologies, memory parking systems typically use a unified end-to-end large model, using the same model for cruise control in different underground parking garages. This approach ignores the environmental differences between underground parking garages and makes it difficult to adapt and optimize for specific underground parking garages, resulting in poor stability of the memory parking function in different underground parking garages.

[0020] To address the aforementioned problems, this application proposes a model optimization method, apparatus, device, medium, and product based on memory parking. The following embodiments are provided to further illustrate this application: Please see Figure 1 , Figure 1 This is a schematic diagram of the architecture of a memory parking system according to an exemplary embodiment of this application. The memory parking system may include a vehicle-side device 10 and a cloud-side device 11, and realizes memory parking functions such as underground garage mapping, cruise route memory, trajectory reuse, and parking control based on the vehicle-cloud collaborative architecture of the vehicle-side device 10 and the cloud-side device 11.

[0021] Among them, the vehicle-side 10 is deployed in intelligent driving vehicles and is responsible for the perception, decision-making and execution of the parking function. It can collect environmental characteristics, vehicle position and driving trajectory data of underground garages, and has the capabilities of cruise path learning, scene memory, real-time trajectory planning and vehicle driving braking control.

[0022] Cloud 11 is deployed on the server and is responsible for the data management of the memory parking system. It has capabilities such as map storage, cruise route memory, map incremental update, remote vehicle status monitoring and memory parking model optimization. It is used to manage the cruise scene data and driving routes reported by each vehicle.

[0023] The vehicle terminal 10 and the cloud terminal 11 establish a two-way communication link through wireless communication methods such as vehicle networking, so that the vehicle terminal 10 can upload mapping data, cruise trajectory and vehicle operation status information to the cloud terminal 11, and the cloud terminal 11 can send mapping information of underground parking garage, dispatch control instructions and memory parking model to the vehicle terminal 10.

[0024] Please see Figure 2 , Figure 2 This is a flowchart illustrating a model optimization method based on memory parking according to an exemplary embodiment of this application. The method can be applied to the vehicle side and may include the following steps: Step S201: Collect task execution data and perception data associated with the task execution data during the vehicle's memory parking task in the underground garage; the task execution data includes the garage identifier of the underground garage and is used to determine the task execution indicators of the underground garage; the perception data is used to optimize the memory parking model when the task execution indicators meet the preset optimization conditions. The memory parking model is used to perform the memory parking task in the underground garage. The memory parking task includes mapping task and / or cruise task.

[0025] In this step, during the process of the vehicle entering the underground garage and executing the memory parking task, the vehicle obtains task execution data containing the garage identifier by reading the vehicle controller status signal, recording task execution node information, and marking the garage identifier. The collection of task execution data can cover the entire cycle of the memory parking task from start to finish.

[0026] During the memory parking task in the underground garage, perception data associated with the task execution data can also be collected through various means, such as camera to collect visual images, lidar to collect environmental point cloud data, millimeter-wave radar and ultrasonic radar to collect obstacle perception information, and positioning module to collect vehicle pose data.

[0027] Among them, the memory parking task is an intelligent driving task performed by the vehicle in an underground parking garage based on the characteristics of the site environment and the driving route. It can realize site path memory, route reuse, and autonomous driving functions. The memory parking task may include mapping tasks and / or cruising tasks.

[0028] The mapping task is the task of building a parking map and obtaining mapping data when the vehicle first enters the underground parking garage. This can include intersection mapping and parking space mapping. The cruise task is the task of driving within the parking area along a memorized route based on the mapping data. It can realize the automatic driving process from the garage entrance to the target parking space or from the target parking space to the garage exit. This can include parking entry cruise, parking exit cruise, and intersection cruise.

[0029] Parking cruise is the process of the vehicle traveling from the garage entrance to the target parking space along the parking map route; parking exit cruise is the process of the vehicle traveling from the target parking space to the garage exit along the parking exit map route; intersection cruise is the process of the vehicle traveling along the memorized route at the intersection position within the parking area.

[0030] Task execution data refers to the vehicle operation and status data generated by the vehicle during the execution of the memory parking task. It is used to evaluate the execution effect of the memory parking function in the current underground garage for the cloud. It may include the garage identifier of the underground garage, as well as data such as the start status, start time, execution progress, termination status, termination time, and vehicle control signals of the memory parking task.

[0031] Garage identification is used to distinguish different underground garages and differentiate different scene characteristics. It can provide a basis for evaluating the subsequent task execution indicators of the underground garage and can include location information, underground garage number, etc. Perception data is the underground garage environment information collected in real time by various on-board sensors. It can be used as sample data for cloud-based optimization of the memory parking model. Perception data can include visual image data, point cloud data, vehicle pose data, etc.

[0032] The memory parking model is a decision-making model used to realize environmental perception, mapping, path planning, and driving control in memory parking scenarios. It can be an end-to-end intelligent driving model that can output corresponding mapping strategies and / or cruise strategies based on input perception data. The optimized memory parking model can generate adapted mapping strategies and / or cruise strategies for the current underground parking garage to execute mapping and / or cruise tasks.

[0033] The mapping strategy is the control logic for environmental perception acquisition, road feature extraction, intersection detection, and topological map construction when the vehicle performs mapping tasks in an underground parking garage. It can adapt to the site characteristics of underground parking garages, such as low lighting and simple features, and achieve accurate and reliable underground parking garage map construction.

[0034] The cruise strategy is the decision-making logic of the vehicle when performing cruise tasks in an underground parking garage. It is based on the mapping data of the underground parking garage and includes path tracking, intersection detection, obstacle avoidance strategy, and speed control strategy. It is used to control the vehicle to perform parking cruise, parking exit cruise, and intersection cruise.

[0035] Task execution metrics are quantitative indicators reflecting the effectiveness of a vehicle's memory parking task in an underground parking garage. They reflect the reliability and stability of the memory parking task in the underground garage. Task execution metrics can include mapping execution metrics and / or cruise execution metrics. Mapping execution metrics quantify the quality of the vehicle's mapping task in the underground garage, reflecting the reliability of the underground garage's site topology map and semantic feature construction. Mapping execution metrics can include mapping success rate, mapping failure rate, etc.

[0036] Mapping success rate is a metric used to quantify the effectiveness of vehicles in completing mapping tasks in underground parking garages. It reflects a vehicle's mapping capability within the current underground parking garage. Mapping success rate can include parking entry mapping success rate, parking exit mapping success rate, intersection mapping success rate, etc.

[0037] The parking entry mapping success rate represents the proportion of vehicles successfully creating a parking route from the garage entrance to the target parking space within an underground parking garage. It is calculated as the ratio of the total number of successful parking entry mappings to the total number of parking entry mappings multiplied by 100%. The parking exit mapping success rate represents the proportion of vehicles successfully creating a parking route from the target parking space to the garage exit within an underground parking garage. It is calculated as the ratio of the total number of successful parking exit mappings to the total number of parking exit mappings multiplied by 100%.

[0038] Cruise execution metrics are used to quantify the effectiveness of vehicles autonomously navigating within underground parking garages along routes memorized from map data. Cruise execution metrics can include cruise success rate, cruise failure rate, etc. Cruise success rate can include parking entry cruise success rate, parking exit cruise success rate, intersection cruise success rate, etc.

[0039] The parking cruise success rate represents the proportion of a vehicle that successfully completes cruise driving along the parking map route. It is calculated as the ratio of the total number of successful parking cruises to the total number of parking cruises, multiplied by 100%. The parking exit cruise success rate represents the proportion of a vehicle that successfully completes cruise driving along the parking exit map route. It is calculated as the ratio of the total number of successful parking exits to the total number of parking exits, multiplied by 100%.

[0040] Task execution metrics can also include parking quality indicators such as parking success rate and parking failure rate. Parking success rate is an indicator used to quantify the effectiveness of vehicles completing parking and exit operations in a mapped underground parking garage, reflecting the vehicle's parking ability within the garage. Parking success rate can include both parking entry success rate and parking exit success rate.

[0041] The parking success rate represents the proportion of times a vehicle successfully completes an automatic parking maneuver based on a memorized map route. It is calculated as the ratio of total successful parking attempts to the total number of parking attempts multiplied by 100%. The parking exit success rate represents the proportion of times a vehicle successfully exits a parking space based on a memorized map route. It is calculated as the ratio of total successful parking exits to the total number of parking exits multiplied by 100%.

[0042] Step S202: Send the task execution data and perception data to the cloud.

[0043] In this step, after the vehicle completes the collection of task execution data and perception data, it uploads the collected task execution data and perception data to the cloud through the communication link using methods such as real-time transmission and timed transmission.

[0044] Based on the received task execution data, the cloud platform categorizes and organizes the data according to dimensions such as underground parking garages and their entrances, using garage identifiers within the task execution data. For example, it can determine the corresponding underground parking garage based on the garage identifier and identify the task execution metrics for that garage. These metrics are then compared with preset optimization conditions. If the metrics meet the preset conditions, the memory parking model is optimized based on the received perception data, resulting in an optimized memory parking model for that specific underground parking garage.

[0045] The cloud sends the optimized memory parking model to the vehicle. The optimized memory parking model can provide the vehicle with a control strategy adapted to the current underground garage, ensuring the stable execution of memory parking tasks in different underground garages.

[0046] Among them, the preset optimization conditions are used to determine whether the memory parking function needs to be optimized in the corresponding underground garage. They can reflect whether the performance of the memory parking function in the underground garage meets the standards. These conditions can include task execution indicators being lower than the corresponding preset threshold (such as a cruise success rate of less than 90%), N consecutive task execution failures, and the proportion of a certain type of termination reason exceeding the preset proportion threshold.

[0047] For example, if mapping success rate and other mapping performance indicators are lower than the corresponding threshold, it can be determined that the preset optimization conditions are met. Based on the perception data, the memory parking model can be optimized so that the optimized memory parking model can generate a mapping strategy that is adapted to the current underground parking garage, thereby improving the mapping success rate of the vehicle in performing mapping tasks in the underground parking garage.

[0048] For example, if the cruise success rate and other cruise execution indicators are lower than the corresponding threshold, it can be determined that the preset optimization conditions are met. Based on the perception data, the memory parking model is optimized so that the optimized memory parking model generates a cruise strategy that is adapted to the current underground garage, thereby improving the cruise success rate of the vehicle performing cruise tasks in the underground garage.

[0049] In one embodiment, to accurately quantify the performance of the memory parking function in the underground parking garage, the cloud can determine the number of successful executions and the total number of executions of the memory parking task for the underground parking garage based on task execution data. For example, it can count the number of successful executions and the total number of executions of the cruise task within the current memory parking task for the underground parking garage.

[0050] The number of successful executions is the cumulative number of times a memory parking task is successfully executed in the underground parking garage within a preset time period (e.g., one year). This includes successful executions of memory parking tasks such as intersection mapping, parking entry navigation, parking exit navigation, and intersection navigation. The total number of executions is the cumulative number of memory parking tasks executed in underground parking spaces within the preset time period.

[0051] The task execution indicators for the underground parking garage can be determined based on the ratio of the number of successful executions to the total number of executions, using methods such as ratio calculation and classification statistics. For example, the current patrol success rate of the underground parking garage can be calculated based on the ratio of the number of successful executions of the patrol task to the total number of executions; similarly, the current mapping success rate of the underground parking garage can be calculated based on the ratio of the number of successful executions of the mapping task to the total number of executions.

[0052] As mentioned above, by determining the number of successful executions and the total number of executions of the underground parking garage memory parking task based on task execution data, and then determining the task execution indicators of the underground parking garage based on the number of successful executions and the total number of executions, the execution effect of the memory parking function of the underground parking garage can be accurately reflected, providing a more accurate judgment basis for model optimization and improving the reliability of model optimization.

[0053] The memory parking-based model optimization method in this embodiment collects task execution data and associated perception data from the vehicle during the memory parking task in the underground garage. This data is then sent to the cloud. By using the underground garage as a dimension, task execution indicators are determined based on the task execution data collected from the vehicle during the memory parking task. The memory parking model is then optimized based on the perception data, enabling the optimized model to adapt to mapping and / or cruising tasks in the corresponding underground garage. This achieves targeted model optimization for each underground garage, thereby improving the stability of the memory parking function in different underground garages.

[0054] In the foregoing embodiments, it was described that the vehicle-side collected task execution data and perception data with the current underground parking garage identifier and uploaded them to the cloud. The cloud determined the task execution indicators of the underground parking garage and completed the model iteration optimization, enabling the memory parking model to adapt to the scene characteristics of different underground parking garages. In the following embodiments, the model optimization process based on memory parking will be described in more detail, and it can be applied to any of the above embodiments.

[0055] In one embodiment, to achieve targeted model optimization for intersection scenarios such as crossroads, straight-left intersections, straight-right intersections, and left-right intersections in underground parking garages, the memory parking model may include an intersection detection model. The intersection detection model is a sub-model responsible for intersection recognition in the intersection scenarios of the underground parking garage, used to control the vehicle to complete the memory parking task in various complex intersection scenarios. Task execution data may include the vehicle's intersection position in the underground parking garage when the memory parking task terminates.

[0056] After sending the task execution data and perception data to the cloud, the cloud uses the perception data of the intersection location to iteratively optimize the original intersection detection threshold in the intersection detection model. For example, the cloud first analyzes the visual feature changes of ordinary driving sections and intersection locations in underground parking garages, and combines this with environmental features such as lighting characteristics and lane line distribution in the underground parking garage to determine a baseline threshold suitable for the intersection location. Then, through multiple rounds of calibration using actual operating samples, the baseline threshold is fine-tuned to obtain the optimized intersection detection threshold corresponding to the intersection location.

[0057] The cloud sends the optimized intersection detection model to the vehicle. The vehicle receives the optimized intersection detection model sent by the cloud. The intersection detection model includes the intersection detection threshold corresponding to the optimized intersection location. The intersection detection threshold is a visual feature index used to determine the intersection location in the underground parking garage. It is calculated by the cloud based on the perception data of the intersection location and adjusted for scene adaptation. It can distinguish the feature differences between ordinary driving sections and intersection locations.

[0058] When the vehicle subsequently performs a memory parking task at the intersection of the underground parking garage, it can acquire perception data of the intersection location. This perception data can be image data collected by cameras, reflecting environmental changes at the intersection of the underground parking garage.

[0059] For example, in the ordinary driving section of an underground parking garage, the visual features of the image data collected during the memory parking task are continuous and straight lane line outlines; while at the intersection, there may be situations where the intersection view is wide, there are more lane lines, and pedestrians and vehicles suddenly appear, and the visual features in the image data will show obvious changes such as feature point replacement, displacement, and sharp increase.

[0060] The intersection detection model outputs the detection results of the intersection location based on the perception data. For example, the perception data is input into the optimized intersection detection model, and the intersection detection model determines the comparison result between the perception data and the intersection detection threshold, and determines the detection result based on the comparison result between the perception data and the intersection detection threshold.

[0061] For example, the intersection detection threshold can be a mutation intensity threshold, which is used to determine whether the degree of visual feature mutation at the intersection location in the underground parking garage meets the intersection recognition standard, such as 10-30 pixels / frame. The intersection detection model can determine the visual feature mutation intensity at the intersection location based on the perception data according to the following formula 1: Formula 1 in, It represents the intensity of visual feature abrupt change at the intersection location. The intensity of visual feature abrupt change can quantify the degree of change in perceived data between the current moment and the previous moment. The greater the intensity of visual feature abrupt change, the more obvious the abrupt change in the visual features of the scene, and it can reflect whether the scene where the vehicle is located has changed from a normal driving section to an intersection location. This indicates the number of successfully matched feature points, such as corner points, in the two frames of perception data. Indicates the current time The state of features such as coordinates or descriptors in the perceived data; Indicates the previous moment The state of features such as coordinates or descriptors in the perceived data; The L2 norm is used to calculate the degree of difference in the feature state of the same feature point in two consecutive frames of perceived data.

[0062] The intersection detection model compares the calculated visual feature mutation intensity with the optimized mutation intensity threshold. When the visual feature mutation intensity is greater than the mutation intensity threshold, the detection result is determined that the current intersection location is an intersection; when the visual feature mutation intensity is less than or equal to the mutation intensity threshold, the detection result is determined that the current intersection location is a normal driving section.

[0063] The intersection detection threshold can also be an intersection distance threshold. The intersection distance threshold is used to verify whether the spatial scale of the visually abrupt change region meets the intersection's passage conditions. It is a threshold for determining the intersection's location from a spatial distance perspective, such as 8 meters. The intersection detection model determines the distance information corresponding to the visually abrupt change region based on the perceived data of the intersection location, compares this distance information with the intersection distance threshold, and outputs the intersection location detection result based on the comparison result.

[0064] Figure 3 A schematic diagram of a crossroads is shown. Figure 3 During the intersection detection process shown, the vehicle can continuously perform cross-scanning using a forward-facing camera to obtain perception data and locate the intersection entrance points AB. If a visual feature abrupt change is detected at points CD, the size of the visual abrupt change area is calculated, and the distance between points BC and BC is measured. When the distance between points BC and BC is greater than or equal to the intersection distance threshold, the current intersection location is determined to be an intersection. If the distance between points BC and BC is less than the intersection distance threshold, the area between points BC and BC is determined to be impassable, and the current intersection location is determined not to be an intersection.

[0065] Based on the detection results, the vehicle can determine the vehicle control strategy for the intersection location. For example, based on the intersection type and environmental characteristics in the detection results, differentiated vehicle control strategies can be generated for mapping and cruising tasks. During mapping, the intersection feature acquisition logic can be optimized to improve the quality of intersection modeling; during cruising, vehicle speed, path tracking, and vehicle steering strategies can be adjusted. Furthermore, the vehicle speed characteristics when the driver passes through the intersection can be referenced to ensure smooth passage through various intersection areas, improving the reliability of task execution in underground parking garage intersection scenarios.

[0066] As described above, by collecting the intersection location at the time of task termination and outputting the intersection location detection result based on the optimized intersection detection threshold and the perception data of the intersection location, it is possible to achieve adaptation and optimization of different underground parking garage intersection scenarios with the intersection as the dimension. This enables the intersection detection model to accurately identify sudden changes in the visual features of the intersection, avoid problems such as sudden braking, wrong stopping, or driver takeover caused by inaccurate intersection detection, and improve the mapping accuracy and traffic safety of vehicles at underground parking garage intersections.

[0067] In one embodiment, when outputting the intersection location detection result through the intersection detection model, the underground parking garage features can be determined first based on perception data. For example, based on perception data such as image information collected by cameras and point cloud information collected by radar, coordinate transformation can be completed by combining camera intrinsic and extrinsic parameters and vehicle pose information, and the underground parking garage features can be generated after fusing multi-sensor features.

[0068] Underground parking garage features are a set of features used to characterize the semantic and geometric information of the environment inside an underground parking garage. They can be bird's-eye view (BEV) feature maps, including elements such as lane lines, parking spaces, and intersections.

[0069] For example, the basement features in the underground parking garage can be determined by fusing the BEV features corresponding to all cameras and radars according to the following Formula 2: Formula 2 in, Coordinates in the BEV feature map eigenvalues ​​at that location ; Indicates the first One camera sensor, Indicates depth dimension; For the first Each camera at pixel coordinates ,depth Three-dimensional features at the location; This is a camera pixel-world coordinate projection function used to project pixels in the camera coordinate system to the BEV coordinate system. As an indicator function, when the projected coordinates are in the BEV coordinate system The value is 1 if a match is found, and 0 otherwise.

[0070] After determining the features of the parking garage, a query vector is generated using the parking garage features as keys and values, and the intersection detection threshold as a reference. This process identifies candidate features among the parking garage features that belong to the intersection type. For example, the parking garage features are used as keys (K) and values ​​(V) for attention calculation, and the optimized intersection detection threshold is used as a constraint to generate the intersection query vector (Q).

[0071] By initially matching the query vector with the features of the parking garage, a cross-attention mechanism is used to filter out candidate features of the parking garage that belong to the intersection type, such as lane line intersection areas and visual abrupt change areas, so as to achieve the initial positioning of the intersection candidate area and reduce the complexity of subsequent calculations.

[0072] For example, candidate features belonging to the intersection type in the basement features can be determined by cross-attention calculation according to the following formula 3: Formula 3 in, This indicates that after cross-attention processing, the first... The group query vector in the corresponding region The enhanced feature vectors output at each position are candidate features after feature interaction and information aggregation are completed; This is a cross-attention operation function used to fuse semantic and geometric information between the query vector and the features of the geodatabase; For the first A query vector is generated by combining the optimized intersection detection threshold with the features of the underground parking lot, and is used to retrieve the intersection feature area. The BEV feature map can be used as the key and value for cross-attention computation.

[0073] By using a cross-attention mechanism, the initial query vector is fully associated with scene information in the global BEV features, which enhances the expressive power of intersection features and further accurately distinguishes between intersection areas and ordinary driving areas. This enables precise screening of candidate features for intersection types, which can both reduce the scope of subsequent data processing and ensure the effectiveness of candidate features.

[0074] After identifying candidate features, self-attention calculation can be performed on them to obtain enhanced candidate features. For example, by using a self-attention mechanism and combining it with an intersection detection threshold, self-attention calculation can be performed on the selected candidate features to capture local feature associations and global contextual dependencies within the candidate region, resulting in enhanced candidate features after intersection feature enhancement. For example, self-attention calculation can be performed according to the following formula 4, based on the intersection detection threshold, to obtain enhanced candidate features: Formula 4 in, For the enhanced candidate features corresponding to the candidate features, For query vector, , For the key vector and value vector of the candidate features, The dimension of the query vector; This represents the dot product operation between the query vector and the key vector, used to quantify the feature similarity between different positions within a candidate feature. This is a normalization function used to convert the scaled similarity results into a weight distribution in the 0-1 range, where the weights represent the degree of correlation between different feature locations.

[0075] By constraining the self-attention calculation process by combining the intersection detection threshold, environmental noise and interference information in the candidate features can be filtered out, and the feature expression of intersection areas such as lane line intersections, abrupt changes in the field of view, and changes in the roadside contour can be enhanced, providing highly recognizable enhanced candidate features for subsequent result judgment.

[0076] After obtaining the enhanced candidate features, they can be compared and normalized with the intersection detection threshold to determine the detection result of the intersection location. For example, the enhanced candidate features can be represented as a vector, where each dimension corresponds to the semantic or geometric attribute of the candidate feature. The difference or ratio between each element in the enhanced candidate feature and the intersection detection threshold can be calculated to obtain the original comparison value. Subsequently, functions such as Sigmoi and Softmax are used to normalize the original comparison value to obtain the normalized intersection confidence score. If the intersection confidence score is greater than a preset probability threshold, the candidate feature is determined to be an intersection and the intersection type is determined; otherwise, the candidate feature is determined not to be an intersection.

[0077] By employing a self-attention mechanism to enhance candidate features, intersection features can be strengthened and environmental interference can be suppressed. Furthermore, by combining intersection detection thresholds and normalization processing for intersection determination, the accuracy and robustness of intersection recognition can be further improved.

[0078] To accurately quantify the mapping quality and cruise performance of the vehicle at intersections in underground parking garages, task execution metrics can include the intersection mapping success rate and / or intersection cruise success rate. The intersection mapping success rate characterizes the proportion of times the vehicle successfully constructs a map of the intersection area at the intersection location in the underground parking garage. The intersection mapping success rate is calculated as the ratio of the total number of successful intersection mapping attempts to the total number of intersection mapping attempts multiplied by 100%.

[0079] The intersection cruise success rate is used to characterize the proportion of vehicles that successfully complete cruise driving along the memory-mapped route at the intersection position in the underground parking garage. The intersection cruise success rate is calculated by multiplying the ratio of the total number of successful intersection cruises to the total number of intersection cruises by 100%.

[0080] As described above, by optimizing the memory parking model based on the intersection mapping success rate and / or intersection cruise success rate at the intersection location, the model can be precisely optimized for individual intersections that do not meet the performance standards. This enables targeted iteration for various intersection scenarios in underground parking garages without redundant modifications to the overall model. It adapts to complex intersection scenarios in underground parking garages, such as low light and obstructed views, improving the mapping accuracy of each intersection location and reducing misjudgment of intersection features and path deviations during cruise.

[0081] For further details on intersection mapping and intersection patrol scenarios in underground parking garages, please refer to [link / reference]. Figure 4 , Figure 4 A flowchart of an intersection mapping and intersection patrol method is shown. The method may include the following steps: Step S401: Confirm that the vehicle has arrived at the underground parking garage.

[0082] In this step, the vehicle can determine that it has arrived at the underground parking garage through positioning signals, underground parking garage entrance and exit recognition, etc., and trigger the memory parking mapping or cruise process.

[0083] Step S402: Start memory parking map creation.

[0084] In this step, when the vehicle first enters the underground parking garage, it initiates the memory parking mapping process and performs mapping tasks within the garage. The vehicle can collect environmental data along the driving route to construct mapping data including features such as lane lines, parking spaces, and intersections, providing path references for subsequent cruising and parking tasks.

[0085] Step S403: Determine whether the intersection map was successfully created.

[0086] In this step, the vehicle first acquires the perception data of the current intersection location, inputs the perception data into the intersection detection model, and then performs the following processing flow through the intersection detection model to output the detection result of the intersection location: Based on the perceived data, the characteristics of the underground parking garage are determined; using the parking garage characteristics as keys and values, and the initial or optimized intersection detection threshold as a reference, a query vector is generated. By matching and filtering the query vector with the parking garage characteristics, candidate features belonging to the intersection type are determined; self-attention calculation is performed on the candidate features to obtain enhanced candidate features; the enhanced candidate features are compared with the intersection detection threshold and normalized to output the detection results of the intersection location.

[0087] Based on the detection results at the intersection location, the vehicle determines the vehicle control strategy for the current intersection location, such as adjusting the vehicle speed, path tracking accuracy, and perception data collection frequency during the mapping process, in order to adapt to the mapping requirements of the intersection scenario and ensure the smooth execution of the mapping process.

[0088] Subsequently, the vehicle-side system combines the detection results of the intersection location, the mapping data of the intersection location, and whether there were any reasons for task termination such as sudden braking or driver takeover during the mapping process to determine whether the intersection mapping was successful: if the intersection detection results accurately identify the intersection location and type, the mapping data is complete, and no abnormal task termination events occur, the intersection mapping is determined to be successful; if there are cases of missed or false detection of intersections, incomplete mapping data, or mapping interruption caused by events such as sudden braking or driver takeover, the intersection mapping is determined to be unsuccessful.

[0089] If the intersection mapping is successful, proceed to step S404; if the intersection mapping fails, send the mapping data, garage identification, intersection location, and perception data to the cloud and proceed to step S410. The cloud uses the mapping data and perception data to train the mapping model in the memory parking model and optimizes the mapping control strategy output by the memory parking model.

[0090] Step S404: Store the mapping data and the garage identifiers for the underground parking garage.

[0091] In this step, the vehicle-side system associates and stores the successfully constructed mapping data with the garage identifier of the underground parking garage, thus binding the mapping data with the corresponding site.

[0092] Step S405: In response to the vehicle arriving at the underground parking garage, obtain mapping data.

[0093] In this step, when the vehicle arrives at the mapped underground parking garage again, it reads the locally stored mapping data.

[0094] Step S406: Execute the memory parking task and enter cruise control mode.

[0095] In this step, the vehicle initiates the memory parking cruise process based on the acquired mapping data, performs the cruise task in the underground garage, controls the vehicle to drive towards the target parking space along the memorized route, and continuously collects task execution data and perception data during the cruise.

[0096] Please see Figure 5 , Figure 5 A schematic diagram of a cruise process is shown. For example, after entering the underground parking garage from the entrance / exit passage, the vehicle can travel along the mapped route indicated by the dashed arrow, proceeding through the garage passage to the target parking space. During the cruise, perception data and task execution data are collected to provide a data foundation for subsequent model optimization.

[0097] Step S407: Determine whether to proceed through the intersection by cruise control.

[0098] In this step, during the cruise, the vehicle determines whether it has successfully passed through intersections such as crossroads, straight-left intersections, and straight-right intersections in the underground parking garage.

[0099] If a sudden braking or driver takeover occurs, the intersection cruise is deemed to have failed, and step S409 is executed; if the intersection cruise is successfully passed, step S408 is executed.

[0100] Step S408: Collect intersection location and sensing data.

[0101] In this step, for intersection scenarios that the vehicle passes through, the vehicle can collect the intersection location and corresponding perception data, which will be used for subsequent cloud-based model optimization and effect evaluation of the intersection scenario.

[0102] Step S409: Collect intersection location, sensing data, and driver takeover data.

[0103] In this step, for intersections that have not been cruised through, the vehicle can collect driver takeover data such as steering and braking control signals, and record data such as intersection location and perception data to form sample data of intersection malfunction scenarios.

[0104] Step S410: Optimize the memory parking model.

[0105] In this step, the cloud receives data uploaded from the vehicle, including intersection location, perception data, driver takeover data, and garage signage. Through model pruning and other methods, if the intersection mapping success rate and intersection cruise success rate are below preset thresholds, the cloud iteratively optimizes the memory parking model based on the collected perception data, adjusting intersection detection thresholds, mapping strategies, and cruise strategies. The cloud can then send the optimized memory parking model to the vehicle via Over-The-Air (OTA) technology, enabling the vehicle to perform memory parking tasks such as mapping and cruise based on the optimized model.

[0106] In one embodiment, to achieve precise optimization of the memory parking model, the task execution data may include the termination reason of the memory parking task. The termination reason is used to indicate the perception data used to optimize the memory parking model and / or the optimization direction of the memory parking model. The optimization direction can reflect the functional modules, algorithm logic, or model parameters that the memory parking model needs to iterate, and can clarify the scope and direction of model optimization.

[0107] For example, the cloud can analyze the termination reasons, select perception data that matches the termination reasons as training samples, and use the perception data associated with the termination reasons to optimize the memory parking model. Taking the termination reason as failure to avoid a stationary obstacle as an example, the cloud can select perception data such as environmental point cloud data of the underground parking garage and surrounding environment images as training samples to train the obstacle perception algorithm of the memory parking model.

[0108] Furthermore, by leveraging sensor data and the garage environment characteristics reflected in it, and based on the optimization direction indicated by the termination cause, the algorithm logic and model parameters to be adjusted in the memory parking model can be determined. The memory parking model can then be optimized along the optimization direction indicated by the termination cause. For example, taking incorrect parking position as the termination cause, the cloud-based system combines the parking space line characteristics and obstacle distribution from the sensor data to optimize the trajectory planning algorithm and parking space control parameters in the memory parking model, adjusting the model's parking path generation logic and parking space strategy.

[0109] As mentioned above, by identifying the termination reason of the memory parking task, we can determine the perception data required for model optimization and the optimization direction of model iteration, thereby reducing the computational overhead of model optimization, improving data filtering efficiency, and enhancing the optimization accuracy of the memory parking model.

[0110] In one embodiment, in order to accurately obtain perception data for optimizing the memory parking model, task execution data of the vehicle during the memory parking task can be collected, and in response to the termination of the memory parking task indicated by the task execution data, perception data within the collection period can be obtained.

[0111] For example, during the execution of a memory parking task in an underground parking garage, the vehicle can continuously collect task execution data, which can provide real-time feedback on the running status of the memory parking task. When the task execution data indicates that the memory parking task has terminated, perception data within a preset collection period can be acquired. The collection period includes the termination time of the memory parking task, for example, 10 seconds before the termination time to 10 seconds after the termination time. The collection period can completely cover the scene process before and after the task termination, ensuring that the perception data can reconstruct the task termination scene.

[0112] The termination scenarios for memory parking tasks can include the termination of mapping tasks and the termination of cruise tasks. The vehicle can determine whether various memory parking tasks have entered the termination state by monitoring the function trigger commands, vehicle operating status and user operation commands in the task execution data in real time.

[0113] The termination of mapping tasks can include the termination of berthing inbound mapping tasks and the termination of berthing outbound mapping tasks. The berthing inbound mapping task is initiated by the user triggering the creation of an inbound route. It is terminated when conditions such as successful inbound route generation, failed inbound route generation, or the user actively clicks to exit are met. Similarly, the berthing outbound mapping task is initiated by the user triggering the creation of an outbound route. It is terminated when conditions such as successful outbound route generation, failed outbound route generation, or the user actively clicks to exit are met.

[0114] Cruise mission termination can include termination of cruise on the parking entry route and termination of cruise on the parking exit route. The cruise mission starts after the user activates memory parking and selects the corresponding cruise route on the vehicle's device. The cruise mission is terminated when the cruise process completes normally, cruise operation fails, or the driver takes over the vehicle. The vehicle's device can also identify and mark termination reasons such as failed oncoming traffic, failed obstacle avoidance, and failed yielding to vulnerable road users.

[0115] In response to mission termination scenarios such as abnormal termination of mapping tasks or failure of cruise tasks, the vehicle can acquire perception data from 10 seconds before to 10 seconds after the mission termination time, such as video image data from various cameras including front-view, rear-view, side-view, and surround-view cameras, as well as environmental detection data from lidar, millimeter-wave radar, and ultrasonic radar.

[0116] The vehicle will upload the task execution data, including the start time, end time, underground garage signage, intersection signage, task execution results, number of executions, number of successes, termination reasons, mapping data, vehicle positioning data, vehicle brake and steering wheel control signals, and perception data collected during the parking task execution period to the cloud.

[0117] Based on the task execution data, determine the mapping success rate, cruise success rate, and the proportion of various termination reasons for the underground parking garage. When the task execution indicators are lower than the preset threshold, determine the optimization direction of the memory parking model based on the proportion of various termination reasons. Use perception data to reconstruct the task termination scenario and optimize the memory parking model.

[0118] The cloud platform can statistically analyze the number of successful parking maneuvers and the total number of executions for different memory parking task dimensions, including parking entry cruise, parking exit cruise, and intersection cruise scenarios. This allows for the determination of task execution metrics such as the success rate of parking entry cruise, parking exit cruise, and intersection cruise in underground parking garages. Furthermore, it can statistically analyze various reasons for the termination of memory parking tasks and calculate the percentage of each type of termination reason.

[0119] The cloud platform compares task execution metrics with preset thresholds. When a task execution metric falls below the preset threshold, it determines the optimization direction for the memory parking model based on the proportion of various termination reasons. Simultaneously, the cloud platform uses perception data uploaded from the vehicle to reconstruct the task termination scenario and performs algorithm optimization and model parameter adjustment on the memory parking model to update the vehicle-side control strategies, such as mapping and cruise strategies, for underground parking garages.

[0120] As described above, by collecting task execution data from the vehicle during the memory parking task, and responding to the task execution data indicating the termination of the memory parking task, the system can obtain perception data within the collection period, including the termination time. This can accurately cover the complete scene before and after the task termination, reduce the transmission of invalid and redundant data, and reduce the storage pressure and communication resource consumption of the vehicle.

[0121] In one embodiment, in order to reduce the resource consumption of data transmission and achieve accurate data collection, the vehicle can receive data collection rules sent by the cloud, and collect task execution data and perception data during the memory parking task performed by the vehicle in the underground garage based on the data collection rules.

[0122] For example, the vehicle can adjust the data collection logic based on the data collection rules issued by the cloud, and collect task execution data and perception data during the memory parking task performed by the vehicle in the underground garage according to the collection period and / or preset collection frequency indicated by the data collection rules.

[0123] The data collection rules are determined by the cloud based on the task execution indicators of the underground parking garage. These rules include at least one of the following: collection period, collection frequency, garage identification in the underground parking garage, and the location of the entrance to the underground parking garage. The collection period is the time range for data collection determined for the memory parking task. It can be limited to the complete cycle from the start to the end of the memory parking task, or it can be limited to a time window of preset duration before and after the task's termination. The collection frequency is the sampling interval frequency at which the vehicle collects task execution data and / or sensing data, and it can control the amount of data collected per unit time.

[0124] For example, by analyzing the task execution indicators of an underground parking garage, if the cloud determines that the task execution indicators of the underground parking garage are lower than a preset threshold, then the underground parking garage is identified as a high-risk location for failure. The data collection period for the underground parking garage can be set to a time window of 10 seconds before and after the task termination time, and the frequency of sensing data collection can be increased. If the task execution indicators of the underground parking garage are higher than or equal to the preset threshold, then a shorter data collection period can be configured for the underground parking garage, and the frequency of sensing data collection can be reduced.

[0125] If the success rate of intersection patrol at a certain intersection is lower than a preset threshold, then the garage identifier and intersection location of the underground garage will be configured in the data collection rules so that the vehicle can collect perception data at that intersection location; if the success rate of intersection patrol at a certain intersection is higher than or equal to the preset threshold, then there is no need to configure that intersection location, or a shorter data collection period and a lower data collection frequency will be configured for that intersection location.

[0126] As described above, by using differentiated data collection rules issued by the cloud based on the task execution indicators of underground parking garages in various locations, adaptive collection of perception data and task execution data can be achieved. This enables the data collection rules to adapt to the execution effect of memory parking tasks in underground parking garages, ensuring the availability of scenario data required for model optimization while achieving reasonable resource allocation.

[0127] For further details on the vehicle-side and cloud-side architecture, please refer to [link / reference]. Figure 6 , Figure 6 A schematic diagram of another memory parking system architecture is shown. Exemplarily, the memory parking system may include a vehicle-side unit 60 and a cloud-side unit 61. The vehicle-side unit 60 may include a driver-side area controller 601, a passenger-side area controller 602, a central computing unit 603, a smart cockpit area controller 604, a smart driving area controller 605, and a powertrain area controller 606.

[0128] The driver-side area controller 601 is a component of the vehicle's area control architecture, responsible for managing the vehicle's functions on the driver's side, including driver's side window adjustment, zoned air conditioning control, body lighting control, and door lock control. It establishes a communication link with the central computing unit 603 via high-speed Ethernet to achieve low-latency transmission of control commands and status signals, and collaborates with other area controllers to complete the vehicle's overall control logic. It also meets the real-time scheduling and operation requirements of the vehicle software, ensuring the stable and reliable operation of various vehicle functions on the driver's side, thereby enhancing the vehicle's intelligence and driving experience.

[0129] The passenger-side area controller 602 is responsible for managing the vehicle's onboard functions on the passenger side, enabling functions such as passenger-side window adjustment and passenger-side air conditioning control. The passenger-side area controller 602 can also establish a communication link with the central computing unit 603 via high-speed Ethernet to achieve data exchange. It can cooperate with the driver-side area controller 601 to jointly realize the vehicle's body control functions. The passenger-side area controller 602 can adopt a hardware abstraction layer architecture to achieve software and hardware decoupling, supporting independent version iteration and function upgrades, shortening the development cycle of onboard functions and reducing later maintenance costs.

[0130] The central computing unit 603 serves as the computing hub for the vehicle-mounted unit 50, responsible for integrating the computing resources and data interaction functions of various regional controllers. It can adopt a heterogeneous chip architecture and be configured with high-speed Ethernet and other communication interfaces to meet the hardware requirements for high-performance computing and safe vehicle operation. It employs a decoupled hardware and software architecture and a service-oriented design, supporting iterative upgrades of vehicle functions and programs via over-the-air (OTA) downloads.

[0131] The central computing unit 603 can perform intelligent driving perception and decision-making calculations, intelligent cockpit multimodal interaction processing, and coordinate task and vehicle data resource scheduling, which complies with the safety specifications of vehicle functions and provides computing power support for the execution of memory parking tasks.

[0132] The Intelligent Cockpit Area Controller 604 is an in-vehicle cockpit interaction control unit responsible for cockpit service functions such as multi-screen display, voice interaction, and navigation prompts. The Intelligent Cockpit Area Controller 604 is equipped with a processing chip, integrating a multi-screen transmission interface and an audio processing module. It adopts a layered architecture of application service layer, intermediate adaptation layer, and system bottom layer, supporting online version iteration upgrades. The Intelligent Cockpit Area Controller 604 can receive user operation commands through multimodal interaction methods, realizing human-computer interaction functions such as activating memory parking function, driving route selection, and task running status display.

[0133] The Intelligent Driving Area Controller 605 serves as the business processing unit for intelligent driving, responsible for handling tasks such as memory parking, highway navigation, and city navigation. The 605 is equipped with a high-performance heterogeneous processing chip, integrating multi-sensor communication interfaces and hardware redundancy to meet automotive-grade safety and heat dissipation requirements. It adopts a layered architecture design, enabling collaborative processing of functional modules such as perception, prediction, path planning, and vehicle control, and supports remote online program upgrades. The 605 utilizes a memory parking model to process and identify perceived data, enabling environmental perception and vehicle-side control strategy output for underground parking garages.

[0134] The cloud 61 may include the Infrastructure as a Service (IaaS) layer 611, the Platform as a Service (PaaS) layer 612, and the Software as a Service (SaaS) layer 613.

[0135] The Infrastructure as a Service (IaaS) layer 611 provides the foundational resource services for the cloud, which can include both public and private clouds. Public clouds can rapidly allocate computing, storage, and network resources to meet business needs. For example, during large-scale model optimization, appropriate computing resources can be automatically allocated based on the scale of the optimization task, ensuring efficient task execution and reducing infrastructure construction and maintenance costs.

[0136] Private clouds are used to manage task execution data, perception data, mapping data, and other data, ensuring data security and privacy protection for intelligent driving functions. They also allow for customized configuration and management of intelligent driving data processing and analysis workflows to meet individual business needs.

[0137] Platform as a Service (PAS) layer 612 possesses data storage, data processing, and vehicle-cloud collaborative scheduling capabilities. In terms of data storage, it can store multimodal perception data such as camera images, LiDAR point cloud data, and vehicle sensor data. It provides a cloud database service that supports multimodal data storage, ensuring unified storage and management of different data types. For example, for image data and point cloud data, corresponding data storage formats and index structures are adopted for rapid retrieval and access.

[0138] The Platform as a Service (PAS) layer 612 employs a distributed storage architecture, distributing data across multiple nodes to improve storage capacity and performance. Simultaneously, data redundancy and backup mechanisms ensure data reliability and availability. Even if some nodes fail, normal data access will not be affected.

[0139] The Platform as a Service (PAS) layer 612 can also manage the lifecycle of intelligent driving data, which includes stages such as data collection, storage, processing, analysis, and deletion. Data can be tiered based on usage frequency and importance; for example, frequently used data can be stored on high-performance storage media, while historical data can be stored on lower-cost storage media to optimize storage costs.

[0140] The Platform as a Service (PAS) layer 612 can clean and preprocess the data uploaded from the vehicle, removing noise, outliers, and duplicate data to improve data quality. For example, it can filter noise in images and remove abnormal reflection points from radar data. It can also perform preprocessing operations such as format conversion and normalization to facilitate subsequent analysis and processing.

[0141] The Platform as a Service (PAS) layer 612 can leverage the computing power of cloud computing, employing parallel computing and distributed processing technologies to rapidly process large-scale intelligent driving data. For example, data processing tasks can be broken down into multiple sub-tasks and distributed across multiple computing nodes for simultaneous processing, thereby improving data processing efficiency.

[0142] The Platform as a Service (PAS) layer 612 can utilize data analytics and mining techniques to extract information and knowledge from intelligent driving data. For example, by analyzing vehicle driving data within task execution data, it can uncover driving patterns and behavioral characteristics under different road conditions; by analyzing perception data, it can identify potential faults and anomalies. These analytical results can provide a data foundation for optimizing the memory parking model.

[0143] The Platform as a Service (PAS) layer 612 establishes a real-time communication channel between the vehicle and the cloud, ensuring that the vehicle can upload data to the cloud in a timely manner, and that the cloud can quickly send instructions and information to the vehicle. The vehicle-cloud collaborative cloud engine within the PAS layer 612 reduces data transmission latency and bandwidth consumption, improving communication efficiency by employing efficient communication protocols and data compression technologies. For example, during vehicle operation, collected task execution data and perception data can be transmitted to the cloud in real time. The cloud then uses a memorized parking model to make decisions based on this data and sends the vehicle control strategy back to the vehicle.

[0144] The Platform as a Service (PAS) layer 612 is responsible for task scheduling and resource management on both the vehicle and cloud sides. It allocates computing, storage, and network resources rationally based on task priority and resource availability. For example, when multiple vehicles simultaneously upload data for model optimization, the cloud can dynamically allocate computing resources according to the task requirements and resource usage of each vehicle, ensuring the smooth execution of the task.

[0145] The Platform as a Service (PAS) layer 612 also supports the updating and deployment of memory parking models. Once the memory parking model is optimized in the cloud, it can be quickly deployed to the vehicle, upgrading the vehicle's memory parking function. Simultaneously, it can monitor and manage the model's operational status on the vehicle, promptly identifying and resolving issues during model operation.

[0146] The Software as a Service (SaaS) layer 613 can handle functions such as accident analysis, memory parking data operation, and user behavior data mining. For example, it can collect and store task execution data and perception data when accidents occur during memory parking tasks performed by the vehicle in an underground garage, in order to analyze the cause of the accident and determine liability. Through cloud storage services, accident data can be stored long-term and an accident database can be established for subsequent querying and analysis.

[0147] The Software as a Service (SaaS) layer 613 can utilize data analytics to conduct in-depth analysis of accident data during the memory parking process, determining the causes and patterns of accidents. For example, by analyzing sensor data and driving trajectories before a vehicle collision, it can determine whether there are issues such as improper driving behavior, sensor malfunctions, or algorithm defects. Simultaneously, by combining environmental information at the time of the accident, such as weather and road conditions, it can comprehensively analyze the causes of the accident and optimize the memory parking model based on these causes.

[0148] The Software as a Service (SaaS) layer 613 can provide safety improvement suggestions and measures based on the results of accident analysis. For example, it can generate suggestions for algorithm optimization based on discovered algorithm defects; generate suggestions for strengthening sensor detection and maintenance based on sensor malfunctions; and remind users to pay attention to safe driving in cases of improper driving behavior, and guide users to drive correctly through the assistance functions of the intelligent driving system.

[0149] The Software as a Service (SaaS) layer 613 can determine the task execution indicators of the underground parking garage based on task execution data, and formulate appropriate vehicle control strategies to improve vehicle driving safety and comfort. Simultaneously, it can optimize the memory parking model based on the analysis results by analyzing driver takeover data.

[0150] The Software as a Service (SaaS) layer 613 can collect user behavior data and usage preferences, such as driving habits, operation frequency, and driving routes, during the memory parking task. By analyzing the behavioral data, the user's driving behavior characteristics can be determined, and the memory parking model can be optimized based on these characteristics. For example, for users who prefer aggressive driving, the memory parking model can be trained to learn more sensitive acceleration and braking responses; for users who prioritize comfort, the memory parking model can be trained to adjust the suspension system and seat position to improve driving comfort. The SaaS layer 613 can also optimize the human-computer interaction logic and user interface layout of the memory parking function to adapt to the personalized needs of different users and improve user satisfaction with the memory parking function.

[0151] Please continue reading. Figure 7 , Figure 7A schematic diagram of a vehicle-side architecture is shown. For example, the vehicle-side unit 60 may also include a data acquisition unit 607. The data acquisition unit 607 can provide multimodal perception data for the memory parking function, serving as the information foundation for realizing underground parking garage environment recognition, vehicle positioning, and memory path determination. It may include a camera, LiDAR, millimeter-wave radar, ultrasonic radar, a high-precision map unit, and an inertial measurement unit.

[0152] The camera system collects visual image data from multiple perspectives, including forward and surround views, enabling it to identify parking lot features such as parking space lines, ground markings, and charging station locations in underground garages. The lidar system generates 3D point cloud data, accurately measuring parking space dimensions, obstacle distances, and outlines to construct local 3D maps of the underground garage, achieving centimeter-level positioning and improving the reliability of perception in low-light and harsh environments.

[0153] Millimeter-wave radar is used for dynamic target detection at medium to long ranges. It can monitor and remember vehicles and pedestrians approaching from behind in the road, outputting target speed and distance information to help determine driving safety conditions. Ultrasonic radar is used for precise short-range perception. It can detect the distance between the vehicle body and the edge of the parking space, as well as the distance between vehicles in front and behind, controlling the accuracy of cruise control, steering, and braking. It is also used to identify low obstacles to avoid wheel scrapes or chassis collisions.

[0154] Inertial measurement units (IMUs) and wheel speed sensors can provide motion state information of the vehicle. IMUs can measure data such as vehicle acceleration and angular velocity to correct positioning deviations and ensure the continuity of vehicle positioning.

[0155] The map unit can provide local map information for the underground parking garage, including parking space coordinates, lane width, speed limits, etc. The local map can compare multimodal perceived data with memory data to calibrate the vehicle's current position and ensure the accuracy of the remembered path.

[0156] The intelligent driving area controller 605 is equipped with a memory parking model, which can include a mapping model, a perception model, and a predictive control model. Among them, the mapping model is responsible for constructing mapping data of the underground parking garage, providing a data foundation for route planning, environmental positioning, and scene matching for the cruise mission. The intersection detection model in the mapping model can perform intersection detection during the mapping mission.

[0157] The perception model can employ a multi-source feature fusion architecture to convert perception data from cameras, LiDAR, millimeter-wave radar, and other sources into bird's-eye view features, enabling the identification of environmental features such as parking spaces, obstacles, and road markings. The intersection detection model within the perception model can perform intersection detection during patrol missions. Furthermore, the perception model can capture the correlations between perceived data through feature association mechanisms, enhancing its environmental perception capabilities in complex scenarios such as underground parking garages.

[0158] The predictive planning and control model can realize cruise path planning, obstacle movement state prediction and vehicle-side control strategy generation based on the mapping data of the underground parking garage, the vehicle's current position and the environmental information output by the perception model. It has the capabilities of path planning, obstacle avoidance and personalized parking adaptation. The predictive planning and control model can adjust the parking operation according to the user's driving habits, improving the safety and comfort of the parking process.

[0159] The intelligent driving area controller 605 is responsible for processing perception data, route planning, and decision-making for memory parking. It synchronizes key data such as mapping data to the central computing unit 603 for storage. The central computing unit 603 issues task commands to the intelligent driving area controller 605 based on the vehicle's operating status. The intelligent driving area controller 605 executes the commands and provides feedback on its operating status, thus achieving collaborative processing of the memory parking task. This vehicle-side architecture, involving the central computing unit 603 and the intelligent driving area controller 605, enhances the richness of intelligent driving data.

[0160] The intelligent driving area controller 605 can send data such as environmental information, parking space information, and task execution progress during the parking process to the intelligent cockpit area controller 604, which then displays the data to the user via the in-vehicle display screen. The intelligent cockpit area controller 604 receives user commands and forwards them to the intelligent driving area controller 605, enabling human-machine interaction.

[0161] The intelligent driving area controller 605 can send control commands such as torque requests and shift requests to the powertrain area controller 606 according to the vehicle-side control strategy, adjusting the vehicle's speed and power output. The powertrain area controller 606 executes the commands and provides feedback on the powertrain system status. Based on the feedback information, the intelligent driving area controller 605 adjusts subsequent planning and control strategies to ensure the stable execution of the memory cruise mission.

[0162] Please see Figure 8 , Figure 8 An interactive flowchart of a model optimization method based on memory parking is shown. This method can be applied to memory parking systems. This method provides a brief description of the complete model optimization process; for detailed implementation, please refer to the aforementioned embodiments. The method may include the following steps: Step S801: During the execution of the memory parking task in the underground garage, the vehicle collects task execution data.

[0163] In this step, the vehicle can determine the collection frequency of task execution data and / or perception data based on the data collection rules issued by the cloud. During memory parking tasks such as mapping and cruising in the underground garage, task execution data such as the memory parking task status, task start and end time, and garage signage in the underground garage are collected. At the same time, multimodal perception data from cameras, LiDAR, millimeter-wave radar, ultrasonic radar, map unit, and inertial measurement unit are collected through the data acquisition unit.

[0164] Step S802: The vehicle responds to the task execution data indication that the parking task has been terminated and acquires the perception data within the collection period.

[0165] In this step, the vehicle analyzes the monitored task execution data in real time. When it is detected that the memory parking task is terminated due to mapping success / failure, cruise success / failure, etc., based on the collection period in the data collection rules, it obtains the perception data corresponding to the collection period, including the task termination time, such as multi-sensor data, mapping data, positioning data, and vehicle control signal data.

[0166] Step S803: The vehicle sends task execution data and perception data to the cloud.

[0167] In this step, the vehicle uploads task execution data, such as the underground parking garage signage, intersection location, task termination reason, number of task executions, and start / stop times, as well as perception data within the collection period, to the cloud for storage and processing via the vehicle-cloud communication link.

[0168] Step S804: Based on the task execution data, the cloud determines the number of successful executions and the total number of executions for the memory parking task in the underground garage.

[0169] In this step, the cloud receives and parses the task execution data and perception data reported by the vehicle. It then calculates task execution indicators such as parking entry mapping success rate, parking exit mapping success rate, parking entry cruise success rate, and parking exit cruise success rate at the underground parking garage level, or at the intersection location level of the underground parking garage, such as intersection mapping success rate and intersection cruise success rate. Simultaneously, it categorizes and summarizes various termination reasons, including failed oncoming traffic, failed stationary obstacle avoidance, failed vulnerable road user avoidance, and driver takeover, to determine the percentage distribution of each type of termination reason.

[0170] Step S805: The cloud uses the perception data associated with the termination reason to optimize the memory parking model based on the optimization direction indicated by the termination reason.

[0171] In this step, the cloud compares the task execution metrics of the underground parking garage with preset thresholds. When the task execution metrics fall below the preset thresholds, it is determined that the task execution metrics meet preset optimization conditions. The optimization direction of the memory parking model is determined by combining the weighted proportions of various termination reasons. The parking garage environment at the time of task termination is fully reconstructed using perception data uploaded from the vehicle. The intersection detection thresholds and algorithm logic in the memory parking model are then optimized in a targeted manner to achieve scenario-based adaptation and optimization of the memory parking model.

[0172] For example, in scenarios where feature misjudgment and abnormal task termination frequently occur at intersections in underground parking garages, the cloud can retrieve the perception data corresponding to the intersection location, statistically analyze the visual feature change patterns between ordinary driving sections and intersection locations, and adjust and correct the original intersection detection threshold so that the optimized intersection detection threshold can accurately distinguish the visual feature differences between ordinary driving sections and intersection locations in the underground parking garage.

[0173] If the task execution metrics remain below the preset threshold or the percentage of abnormal task terminations is too high within the preset time window, the cloud can extend the data collection period for underground parking garages and set a higher collection frequency. If the task execution metrics are consistently higher than or equal to the preset threshold and the task terminates normally, the time range of the collection period can be reduced and the data collection frequency lowered to differentiate the data collection rules for different underground parking garages.

[0174] The cloud sends the adjusted data collection rules to the vehicle. The vehicle receives the data collection rules from the cloud and can adjust the collection strategy for task execution data and perception data according to the collection period and collection frequency in the data collection rules.

[0175] Step S806: The cloud sends the optimized memory parking model to the vehicle.

[0176] In this step, after the cloud completes the intersection detection threshold optimization of the memory parking model, the optimized memory parking model is sent to the vehicle through the vehicle-cloud communication link.

[0177] Understandably, the cloud can also distribute the optimized memory parking model to other vehicles, improving the memory parking capability of multiple vehicles in underground garages, increasing the success rate of mapping and cruising in underground garages, and improving the overall performance of memory parking tasks.

[0178] Step S807: The vehicle-side generates a vehicle-side control strategy based on the optimized memory parking model.

[0179] In this step, when the vehicle subsequently performs a memory parking task in the current underground garage, it can use the vehicle control strategy output by the optimized memory parking model to perform mapping or cruise.

[0180] For example, when the vehicle re-enters the same intersection in the underground parking garage to perform mapping or cruise tasks, it can collect perception data of the intersection location, output the detection results of the intersection location through the optimized intersection detection model, determine the intersection location as a crossroads based on the detection results, and determine vehicle control strategies such as deceleration and adjustment of driving direction at the intersection to improve scene adaptability and task operation stability.

[0181] Please see Figure 9 , Figure 9 This is a flowchart illustrating another memory-based parking model optimization method according to an exemplary embodiment of this application. The method can be applied in the cloud and may include the following steps: Step S901: Receive task execution data and perception data associated with the task execution data sent by the vehicle terminal. The task execution data and perception data are collected by the vehicle terminal during the memory parking task performed in the underground garage. The task execution data includes the garage identifier of the underground garage.

[0182] In this step, the cloud receives task execution data and perception data collected by the vehicle during the memory parking tasks such as mapping memory and cruise memory in the underground garage through the vehicle-cloud communication link.

[0183] Step S902: Based on the task execution data, determine the task execution indicators for the underground parking garage.

[0184] In this step, the cloud platform categorizes and statistically analyzes the task execution data by garage based on the garage identifiers in the task execution data, and calculates the corresponding task execution indicators for the underground garages. It also categorizes and summarizes various termination reasons, such as failed oncoming traffic, failed obstacle avoidance, failed vulnerable road user avoidance, and driver takeover, calculates the frequency of each termination reason, and determines the percentage distribution, thus obtaining the evaluation results of the task execution quality for the underground garages.

[0185] Step S903: If the task execution indicators meet the preset optimization conditions, optimize the memory parking model using perception data. The memory parking model is used to perform memory parking tasks in the underground garage. The memory parking tasks include mapping tasks and / or cruising tasks.

[0186] In this step, the cloud compares the task execution metrics corresponding to each underground parking garage with preset thresholds. When a task execution metric falls below a preset threshold, it can be determined that the task execution metric meets the preset optimization conditions. The optimization direction of the memory parking model is determined by combining the weighted proportions of various termination reasons. Then, the parking garage environment scenario where the task terminated is reconstructed using perception data. The algorithm logic or model parameters of the memory parking model are optimized accordingly to achieve parking garage adaptability optimization. The optimized memory parking model can adapt to the environmental characteristics of the corresponding underground parking garage, so that the vehicle can stably execute memory parking tasks in that underground parking garage subsequently.

[0187] The memory parking-based model optimization method in this embodiment receives task execution data sent by the vehicle and perception data associated with the task execution data. Based on the task execution data, it determines the task execution indicators for the underground parking garage. When the task execution indicators meet preset optimization conditions, it optimizes the memory parking model using the perception data. The memory parking model is used to perform memory parking tasks in the underground parking garage. By taking the underground parking garage as the dimension, it determines the task execution indicators based on the task execution data collected by the vehicle when performing memory parking tasks, and optimizes the memory parking model based on the perception data. This allows the optimized memory parking model to adapt to the corresponding underground parking garage to perform mapping tasks and / or cruise tasks, achieving targeted model optimization for each underground parking garage, thereby improving the execution stability of the memory parking function in different underground parking garages.

[0188] In one embodiment, the memory parking model may include an intersection detection model, and the task execution data may include the intersection position in the underground parking garage when the memory parking task terminates. After optimizing the memory parking model using perception data, the cloud can send the optimized intersection detection model to the vehicle. The intersection detection model includes an intersection detection threshold corresponding to the optimized intersection position. The intersection detection model is used to output the intersection position detection result based on the perception data of the intersection position, and the detection result is used to determine the vehicle control strategy for the intersection position. The threshold optimization, model distribution, and strategy determination process of this embodiment can be found in the corresponding aforementioned embodiments, and will not be repeated here.

[0189] In one embodiment, the task execution data may include the termination reason of the memory parking task. When the cloud uses the perception data to optimize the memory parking model, it may include at least one of the following: optimizing the memory parking model using the perception data associated with the termination reason; optimizing the memory parking model using the perception data based on the optimization direction indicated by the termination reason. The execution logic, data filtering method, and model optimization process of this embodiment can be found in the corresponding foregoing embodiments, and will not be repeated here.

[0190] In one embodiment, the cloud can determine the data collection rules for the vehicle in the underground parking garage based on the task execution indicators of the underground parking garage; the data collection rules are sent to the vehicle, and the data collection rules are used to collect task execution data and perception data of the vehicle during the memory parking task in the underground parking garage. The rule configuration logic and data collection process for different parking garages in this embodiment can be found in the corresponding embodiments described above, and will not be repeated here.

[0191] In one embodiment, when determining the task execution indicators of the underground parking garage based on task execution data, the cloud can determine the number of successful executions and the total number of executions of the memory parking task for the underground parking garage based on the task execution data; and determine the task execution indicators of the underground parking garage based on the number of successful executions and the total number of executions. The indicator statistics rules, data classification methods and quantification process of this embodiment can be found in the corresponding embodiments described above, and will not be repeated here.

[0192] Figure 10 This is a schematic diagram illustrating the structure of an electronic device according to an exemplary embodiment of this application. The electronic device may be, for example, a mobile phone, computer, digital broadcasting terminal, messaging device, game console, tablet device, personal digital assistant, server, smart home appliance, in-vehicle system, etc. (Reference) Figure 10 At the hardware level, the electronic device includes a processor 1001, an internal bus 1002, a network interface 1003, memory 904, and non-volatile memory 1005, and may also include other hardware required for business operations. The processor 1001 reads the corresponding computer program from the non-volatile memory 1005 into the memory 1004 and then runs it, forming a memory-based parking model optimization device at the logical level. Of course, in addition to the software implementation, this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0193] Figure 11 This is a block diagram illustrating a memory-based parking model optimization device according to an exemplary embodiment of this application. (Refer to...) Figure 11 The device may include: The data acquisition module 1101 is used to collect task execution data and perception data associated with the task execution data during the vehicle's memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage and is used to determine the task execution index of the underground garage. The perception data is used to optimize the memory parking model when the task execution index meets the preset optimization conditions. The memory parking model is used to perform the memory parking task in the underground garage. The memory parking task includes mapping task and / or cruise task. The data sending module 1102 is used to send the task execution data and the sensing data to the cloud.

[0194] Optionally, the memory parking model includes an intersection detection model, and the task execution data includes: the intersection position in the underground garage when the memory parking task terminates; The device further includes: The threshold receiving module is used to receive the optimized intersection detection model sent by the cloud, wherein the intersection detection model includes the optimized intersection detection threshold corresponding to the intersection location; The data acquisition module is used to acquire the sensing data of the intersection location; The intersection detection module is used to output the detection result of the intersection position based on the perception data of the intersection position through the intersection detection model; The strategy determination module is used to determine the vehicle-side control strategy for the intersection location based on the detection results.

[0195] Optionally, the intersection detection module, when used to output the detection result of the intersection position based on the perception data of the intersection position through the intersection detection model, includes: Based on the perceived data, the characteristics of the underground parking garage are determined; Using the basement features as keys and values, and the intersection detection threshold as a reference, a query vector is generated to determine candidate features belonging to the intersection type in the basement features. Self-attention calculation is performed on the candidate features to obtain enhanced candidate features; The enhanced candidate features are compared with the intersection detection threshold and normalized to determine the detection result of the intersection location.

[0196] Optionally, the task execution metrics include: the intersection mapping success rate and / or intersection patrol success rate at the intersection location.

[0197] Optionally, the task execution data includes: the termination reason of the memory parking task, wherein the termination reason is used to indicate the perception data used to optimize the memory parking model and / or the optimization direction of the memory parking model.

[0198] Optionally, the data acquisition module 1101, when collecting task execution data during the vehicle's memory parking task in the underground garage and perception data associated with the task execution data, includes: Collect the task execution data of the vehicle during the memory parking task; In response to the task execution data indicating the termination of the memory parking task, perception data within the collection period is acquired, the collection period including the termination time of the memory parking task.

[0199] Optionally, the device further includes: The rule receiving module is used to receive data collection rules sent by the cloud. The data collection rules are determined by the cloud based on the task execution indicators of the underground parking garage. The collection rules include at least one of the following: collection period, collection frequency, parking garage identifier of the underground parking garage, and intersection identifier of the intersection location in the underground parking garage. The data acquisition module 1101 is used to collect task execution data and perception data of the vehicle during the memory parking task performed by the vehicle in the underground garage, based on the data acquisition rules.

[0200] Figure 12 This is a block diagram illustrating another memory-based parking model optimization device according to an exemplary embodiment of this application. (Refer to...) Figure 12 The device may include: The data receiving module 1201 is used to receive task execution data sent by the vehicle terminal and perception data associated with the task execution data. The task execution data and the perception data are collected by the vehicle terminal during the execution of the memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage. The indicator determination module 1202 is used to determine the task execution indicators of the underground parking garage based on the task execution data. The model optimization module 1203 is used to optimize the memory parking model using the perception data when the task execution indicators meet the preset optimization conditions. The memory parking model is used to perform memory parking tasks in the underground garage. The memory parking tasks include mapping tasks and / or cruising tasks.

[0201] Optionally, the memory parking model includes an intersection detection model, and the task execution data includes: the intersection position in the underground parking garage when the memory parking task terminates; the device further includes: The model sending module is used to send the optimized intersection detection model to the vehicle terminal. The intersection detection model includes the optimized intersection detection threshold corresponding to the intersection location. The intersection detection model is used to output the detection result of the intersection location based on the perception data of the intersection location. The detection result is used to determine the vehicle terminal control strategy for the intersection location.

[0202] Optionally, the task execution data includes: the reason for termination of the memory parking task; The model optimization module 1203, when used to optimize the memory parking model using the perception data, includes at least one of the following: The memory parking model is optimized using perception data associated with the termination reason; Using the perceived data, the memory parking model is optimized based on the optimization direction indicated by the termination reason.

[0203] Optionally, the device further includes: The rule determination module is used to determine the data collection rules of the vehicle terminal in the underground parking garage based on the task execution indicators of the underground parking garage. The collection rules include at least one of the following: collection period, collection frequency, parking garage identifier of the underground parking garage, and intersection identifier of the intersection location in the underground parking garage. The rule sending module is used to send the data collection rules to the vehicle terminal. The data collection rules are used to collect task execution data and perception data of the vehicle terminal during the memory parking task performed in the underground garage.

[0204] Optionally, the indicator determination module 1202, when determining the task execution indicators of the underground parking garage based on the task execution data, includes: Based on the task execution data, determine the number of successful executions and the total number of executions for the memory parking task in the underground garage; The task execution indicators for the underground parking garage are determined based on the number of successful executions and the total number of executions.

[0205] 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.

[0206] 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 modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0207] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions, such as a memory including instructions, is also provided, which can be executed by a processor of a memory-based parking model optimization device to implement the method as described in any of the above embodiments.

[0208] The non-transitory computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk and optical data storage device, etc., and this application does not limit it.

[0209] In an exemplary embodiment, a computer program product including a computer program / instruction is also provided, which can be executed by a processor of a memory-based parking model optimization device to implement the method described in any of the above embodiments.

[0210] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0211] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention filed herein. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and alterations can be made without departing from its scope. The scope of this application is limited only by the appended claims.

[0212] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A model optimization method based on memory parking, characterized in that, The method is applied to the vehicle end, and the method includes: The system collects task execution data and associated perception data during the vehicle's memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage and is used to determine the task execution indicators of the underground garage. The perception data is used to optimize the memory parking model when the task execution indicators meet preset optimization conditions. The memory parking model is used to perform the memory parking task in the underground garage, and the memory parking task includes mapping tasks and / or cruising tasks. The task execution data and the perception data are sent to the cloud.

2. The method according to claim 1, characterized in that, The memory parking model includes an intersection detection model, and the task execution data includes: the intersection position in the underground garage when the memory parking task terminates; After sending the task execution data and the sensing data to the cloud, the method further includes: Receive the optimized intersection detection model sent by the cloud, wherein the intersection detection model includes the optimized intersection detection threshold corresponding to the intersection location; Acquire sensing data of the intersection location; The intersection detection model outputs the detection result of the intersection location based on the perception data of the intersection location; Based on the detection results, a vehicle-side control strategy is determined for the intersection location.

3. The method according to claim 2, characterized in that, The step of outputting the detection result of the intersection position through the intersection detection model based on the perception data of the intersection position includes: Based on the perceived data, the characteristics of the underground parking garage are determined; Using the basement features as keys and values, and the intersection detection threshold as a reference, a query vector is generated to determine candidate features belonging to the intersection type in the basement features. Self-attention calculation is performed on the candidate features to obtain enhanced candidate features; The enhanced candidate features are compared with the intersection detection threshold and normalized to determine the detection result of the intersection location.

4. The method according to claim 2, characterized in that, The task execution metrics include: the success rate of intersection mapping and / or the success rate of intersection patrol at the intersection location.

5. The method according to claim 1, characterized in that, The task execution data includes: The termination reason for the memory parking task is used to indicate the perception data used to optimize the memory parking model and / or the optimization direction of the memory parking model.

6. The method according to claim 1, characterized in that, The collection of task execution data during the vehicle's memory parking task in the underground garage, and the perception data associated with the task execution data, includes: Collect the task execution data of the vehicle during the memory parking task; In response to the task execution data indicating the termination of the memory parking task, perception data within the collection period is acquired, the collection period including the termination time of the memory parking task.

7. The method according to claim 1, characterized in that, The method further includes: The system receives data collection rules sent from the cloud, which are determined by the cloud based on the task execution indicators of the underground parking garage. The data collection rules include at least one of the following: collection period, collection frequency, parking garage identifier of the underground parking garage, and intersection identifier of the intersection location in the underground parking garage. The collection of task execution data during the vehicle's memory parking task in the underground garage, and the perception data associated with the task execution data, includes: Based on the data collection rules, the vehicle-mounted device collects task execution data and perception data during the memory parking task performed in the underground garage.

8. A model optimization method based on memory parking, characterized in that, The method is applied in the cloud and includes: The system receives task execution data and perception data associated with the task execution data sent by the vehicle terminal. The task execution data and the perception data are collected by the vehicle terminal during the execution of a memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage. Based on the task execution data, the task execution indicators of the underground parking garage are determined; When the task execution indicators meet the preset optimization conditions, the memory parking model is optimized using the perception data. The memory parking model is used to perform memory parking tasks in the underground garage. The memory parking tasks include mapping tasks and / or cruising tasks.

9. The method according to claim 8, characterized in that, The memory parking model includes an intersection detection model, and the task execution data includes: the intersection position in the underground garage when the memory parking task terminates; After optimizing the memory parking model using the perceived data, the method further includes: The optimized intersection detection model is sent to the vehicle. The intersection detection model includes an optimized intersection detection threshold corresponding to the intersection location. The intersection detection model is used to output the detection result of the intersection location based on the perception data of the intersection location. The detection result is used to determine the vehicle control strategy for the intersection location.

10. The method according to claim 8, characterized in that, The task execution data includes: the reason for the termination of the memory parking task; The optimization of the memory parking model using the perceived data includes at least one of the following: The memory parking model is optimized using perception data associated with the termination reason; Using the perceived data, the memory parking model is optimized based on the optimization direction indicated by the termination reason.

11. The method according to claim 8, characterized in that, The method further includes: Based on the task execution indicators of the underground parking garage, the data collection rules of the vehicle terminal in the underground parking garage are determined. The collection rules include at least one of the following: collection period, collection frequency, parking garage identifier of the underground parking garage, and intersection identifier of the intersection location in the underground parking garage. The data acquisition rules are sent to the vehicle terminal, and the data acquisition rules are used to collect task execution data and perception data of the vehicle terminal during the memory parking task in the underground garage.

12. The method according to claim 8, characterized in that, The process of determining the task execution indicators for the underground parking garage based on the task execution data includes: Based on the task execution data, determine the number of successful executions and the total number of executions for the memory parking task in the underground garage; The task execution indicators for the underground parking garage are determined based on the number of successful executions and the total number of executions.

13. A model optimization device based on memory parking, characterized in that, The device is applied to a vehicle end, and the device includes: The data acquisition module is used to collect task execution data and perception data associated with the task execution data during the vehicle's memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage and is used to determine the task execution index of the underground garage. The perception data is used to optimize the memory parking model when the task execution index meets the preset optimization conditions. The memory parking model is used to perform the memory parking task in the underground garage, and the memory parking task includes mapping task and / or cruise task. The data sending module is used to send the task execution data and the sensing data to the cloud.

14. A model optimization device based on memory parking, characterized in that, The device is used in the cloud and includes: The data receiving module is used to receive task execution data sent by the vehicle terminal and perception data associated with the task execution data. The task execution data and the perception data are collected by the vehicle terminal during the execution of the memory parking task in the underground garage. The task execution data includes the garage identifier of the underground garage. The indicator determination module is used to determine the task execution indicators of the underground parking garage based on the task execution data. The model optimization module is used to optimize the memory parking model using the perception data when the task execution indicators meet the preset optimization conditions. The memory parking model is used to perform memory parking tasks in the underground garage. The memory parking task includes mapping tasks and / or cruising tasks.

15. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor implements the method as described in any one of claims 1-7 or the method as described in any one of claims 8-12 by executing the executable instructions.

16. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the method as described in any one of claims 1-7 or the method as described in any one of claims 8-12.

17. A computer program product having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the method as described in any one of claims 1-7 or the method as described in any one of claims 8-12.

Images