A visual memory parking method based on a deep learning network, medium, device and vehicle

Abstract

The application discloses a visual memory parking method based on a deep learning network, a medium, equipment and a vehicle, and comprises the following steps: acquiring a preloaded map of a parking lot; acquiring a feature map of a target vehicle at a current time and a real vehicle pose of the target vehicle on the preloaded map at a previous time; predicting a virtual vehicle pose of the target vehicle on the preloaded map at the current time based on the real vehicle pose of the target vehicle on the preloaded map at the previous time; cutting out a local map of the target vehicle on the preloaded map based on the predicted virtual vehicle pose of the target vehicle on the preloaded map at the current time; aligning the feature map and the local map under the deep learning network, and outputting a real vehicle pose of the target vehicle in the preloaded map at the current time; and parking based on a parking path corresponding to a target parking space in the cloud memory and the selected target parking space. Through the above method, the cruise process of memory parking can be realized only by visual observation and the vehicle's own odometer, the complex road topological information and semantic lane are avoided, only the memorized route and the light semantic map need to be maintained, the storage space is reduced, the maintenance cost is reduced, and the data between platforms is easier to migrate.

Description

Technical Field

[0001] This application relates to the field of intelligent driving, and in particular to a visual memory parking method, medium, device and vehicle based on deep learning networks. Background Technology

[0002] In autonomous driving technology, memory-based parking technology primarily relies on the vehicle's sensors, cameras, and advanced algorithms. These sensors and cameras accurately capture the vehicle's position and surrounding environment. By processing this data through algorithms, the vehicle can learn and remember specific parking paths and maneuvers. When the vehicle re-enters the same or similar parking environment, the system can automatically guide it to park along the memorized path.

[0003] Most existing memory parking methods rely on LiDAR data to construct complex parking lot point cloud maps and lane semantic elements to achieve relocation and tracking during the memory parking cruise process. The drawbacks of this method are that the vehicle's historical memory data is inconvenient to store and share across different platforms, and it is limited by the type of sensor, meaning it cannot rely solely on vision for vehicle positioning and navigation. Summary of the Invention

[0004] This application provides a visual memory parking method, medium, device, and vehicle based on deep learning networks to solve the problems of existing memory parking methods, which require memorizing complex road topology information and semantic lane information, have a large amount of data to process, high storage space requirements, and high difficulty in data sharing between different platforms.

[0005] To address the aforementioned technical problems, this application provides a visual memory parking method based on a deep learning network. The visual memory parking method includes: acquiring a preloaded map of a parking lot; acquiring a feature map of a target vehicle at the current moment and the actual vehicle pose of the target vehicle on the preloaded map at the previous moment; predicting a virtual vehicle pose of the target vehicle on the preloaded map at the current moment based on the actual vehicle pose of the target vehicle on the preloaded map at the previous moment; cropping a partial map of the target vehicle on the preloaded map based on the predicted virtual vehicle pose of the target vehicle on the preloaded map at the current moment; aligning the feature map and the partial map using a deep learning network to output the actual vehicle pose of the target vehicle on the preloaded map at the current moment; and acquiring a parking path corresponding to the target parking space from cloud memory based on the actual vehicle pose in the preloaded map and a selected target parking space for parking.

[0006] By using cropped local maps in subsequent matching and calculations, the accuracy of vehicle localization is improved. Furthermore, by aligning the current feature map and local map using a deep network, the vehicle's true pose at the current moment is obtained, further enhancing its localization accuracy and facilitating subsequent automated parking based on this location. Feature map acquisition relies solely on visual sensing, maintaining a memorized route and a lightweight semantic map, resulting in small storage space and low maintenance costs. It can also be applied to various vision-based platforms.

[0007] In some embodiments, predicting the virtual vehicle pose of the target vehicle on the preloaded map at the current moment based on the actual vehicle pose of the target vehicle on the preloaded map at the previous moment includes: predicting the virtual vehicle pose of the target vehicle on the preloaded map at the current moment based on the actual vehicle pose of the target vehicle on the preloaded map at the previous moment and the odometer pose of the target vehicle at the previous moment.

[0008] The trajectory parameters of the target vehicle are collected by devices such as wheel speed or inertial measurement units. The relative motion data generated by the vehicle during driving is used to correct the current position of the vehicle. Combined with the vehicle pose of the previous frame, the current position of the vehicle is predicted, further improving the accuracy of the vehicle's position and attitude on the map.

[0009] In some embodiments, the visual memory parking method further includes: collecting historical driving data of the target vehicle and data output by the inertial measurement unit at the current moment, and generating the odometer pose of the target vehicle at the current moment.

[0010] By collecting relative motion data generated by the vehicle during driving, an odometer pose based on odometer coordinates is formed, making the vehicle's positioning more accurate.

[0011] In some embodiments, aligning the feature map and the local map under a deep learning network and outputting the vehicle pose of the target vehicle in the preloaded map at the current time includes: aligning the feature map and the local map under a deep learning network to obtain the vehicle pose of the target vehicle in the local map; and converting the vehicle pose in the local map into the vehicle pose in the preloaded map based on the clipping position of the local map in the preloaded map.

[0012] By aligning feature maps and local maps using deep learning networks, the accuracy of matching is improved; vehicle poses in the local map are converted into the preloaded map, which facilitates subsequent cropping of the local map based on the vehicle poses in the preloaded map.

[0013] In some embodiments, the deep learning network includes a first spatial transformation network (STN), a second spatial transformation network (STN), and a target ground truth. Aligning the feature map and the local map under the deep learning network to obtain the vehicle pose of the target vehicle in the local map includes: inputting the feature map and the local map to the first spatial transformation network (STN) to obtain a first feature map warp and a first transformation parameter estimate; inputting the first feature map warp and the feature map to the second spatial transformation network (STN) to obtain a second feature map warp and a second transformation parameter estimate; forming the total residual of the deep learning network based on the second feature map warp, the first transformation parameter estimate, the second transformation parameter estimate, and the target ground truth; and calculating the regional pose of the target vehicle in the local map based on the target ground truth and the total residual.

[0014] By aligning feature maps and local maps using a deep learning network, the error in predicting the vehicle's current position is corrected, improving the accuracy of vehicle localization. A first spatial transformation network (STN) performs coarse image matching to identify potential image regions. A second spatial transformation network (STN) further processes the image output from the first STN, using feature maps to correct the predicted vehicle position and outputting a distorted feature map. The network residual is then calculated using ground truth data to train the image matching model, further improving image matching accuracy.

[0015] In some embodiments, the ground truth data includes a ground truth map and a ground truth pose; the step of forming the total residual of the deep learning network based on the second feature map distortion, the first transformation parameter estimate, the second transformation parameter estimate, and the target ground truth includes: forming a first residual based on the ground truth map and the second feature map distortion; forming a second residual based on the product of the ground truth pose and the first transformation parameter estimate and the second transformation parameter estimate; and summing the first residual and the second residual to obtain the total residual of the deep learning network.

[0016] A deep learning network is used to align real-time acquired feature maps with cropped local maps to obtain more accurate vehicle positions and poses in the map. The ground truth maps and ground truth poses from the deep learning network are used to correct the outputs of the first and second spatial transformation networks (STNs), improving the matching accuracy between feature maps and the map.

[0017] In some embodiments, the preloaded map includes: a global map stitched together by the feature maps during each learning process of the deep learning network, parking lot entrance information, and memorized parking routes.

[0018] By stitching together semantic information that can be identified from image information to form a global map, the location of the current vehicle in the preloaded map can be determined by matching only the environmental semantic information in the preloaded map, thus reducing the storage space occupied by the memory data.

[0019] In some embodiments, obtaining the preloaded map of the parking lot includes: obtaining the preloaded map of the parking lot closest to the target vehicle from the cloud; matching the feature map of the target vehicle at the entrance of the parking lot with the parking lot entrance information to determine the parking lot that the target vehicle entered and loading the corresponding preloaded map.

[0020] By matching the entrance information of the parking lot, it can be determined whether the current vehicle has entered the parking lot, and whether the parking lot the vehicle has entered matches the parking lots in the preloaded map.

[0021] In some embodiments, after obtaining the preloaded map of the parking lot, the method further includes: determining the initial real vehicle pose of the target vehicle on the preloaded map based on the feature map of the target vehicle and the parking lot entrance.

[0022] By matching the parking lot entrance, the accurate time and location of the vehicle when it enters the parking lot can be obtained, providing the initial location and time point for subsequent updates to the vehicle's location.

[0023] In some embodiments, the step of obtaining the parking path corresponding to the target parking space from cloud memory for parking includes: obtaining the parking path corresponding to the target parking space from cloud memory; converting the parking path into a parking path in odometer coordinates based on the current vehicle pose and odometer pose of the target vehicle; and guiding the target vehicle to park based on the parking path in odometer coordinates.

[0024] By obtaining parking routes from the cloud and converting the route information into odometer coordinates, parking data that was originally in the global map coordinate system can be applied to downstream parking operations.

[0025] In some embodiments, after obtaining the parking path corresponding to the target parking space from the cloud memory and parking, the method further includes: converting the parking path in the odometer coordinates memorized during the parking process of the target vehicle into a parking path in the global map coordinate system; uploading the parking path in the global map coordinate system to the cloud and updating the preloaded map of the corresponding parking lot.

[0026] By recording the parking process of vehicles and generating and uploading memory data, the memory paths recorded in the corresponding parking lot map information can be updated, providing updated parking memory data for the target vehicle and other vehicles. Converting the memory data to a global map coordinate system simplifies the way data is stored in the cloud and improves the applicability of the memory data.

[0027] To solve the above-mentioned technical problems, another technical solution adopted in this application is to provide a storage medium on which program data is stored, wherein the program data is executed by a processor to implement the steps of the visual memory parking method described above.

[0028] The beneficial effects of this storage medium are similar to those of the visual memory parking method described above, and will not be repeated here.

[0029] This application also provides an in-vehicle computer device, including a processor and a memory interconnected thereto, the memory storing a computer program, and the processor executing the computer program to implement the steps of the visual memory parking method described above.

[0030] The beneficial effects of this onboard computer device are similar to those of the visual memory parking method described above, and will not be repeated here.

[0031] This application also provides a vehicle including the on-board computer equipment described above. During operation, the vehicle can complete a visual memory automatic parking process under the control of the on-board computer equipment.

[0032] The beneficial effects of this application are as follows: Unlike existing technologies, this application discloses a visual memory parking method, medium, device, and vehicle based on a deep learning network. By collecting image information around the vehicle and matching and cropping it with a map, precise positioning of the target vehicle in the map is achieved. The sensing method, relying solely on visual observation, reduces the types of sensors required for the memory parking method, improving the vehicle's positioning accuracy in the map. By uploading the visually memorized parking path to the cloud and converting the acquired visual memory path for downstream use, the sharing of trajectory and parking space information between different vehicles is realized. The memory parking cruise data achieved solely through visual observation can be applied to other platforms with visual observation devices, expanding the applicability of historical data. This eliminates the need for the system to memorize complex road topology information and semantic lanes; it only needs to maintain the memorized route and a lightweight semantic map, resulting in small storage space and low maintenance costs. This application also uses a deep learning network to correct and update the vehicle's positioning and attitude, further improving the accuracy of the vehicle's attitude and position in the map. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0034] Figure 1 This is a flowchart illustrating an embodiment of the visual memory parking method based on deep learning networks provided in this application.

[0035] Figure 2 This is a flowchart illustrating Embodiment 2 of the visual memory parking method based on deep learning networks provided in this application;

[0036] Figure 3 This is a flowchart illustrating Embodiment 3 of the visual memory parking method based on deep learning networks provided in this application;

[0037] Figure 4 This is a flowchart illustrating Embodiment 4 of the visual memory parking method based on deep learning networks provided in this application;

[0038] Figure 5 This is a flowchart illustrating Embodiment 5 of the visual memory parking method based on deep learning networks provided in this application;

[0039] Figure 6 This is a flowchart illustrating Embodiment Six of the visual memory parking method based on deep learning networks provided in this application;

[0040] Figure 7 This is a flowchart of Embodiment 7 of the visual memory parking method based on deep learning networks provided in this application;

[0041] Figure 8 This is a schematic diagram of the structure of an embodiment of the storage medium provided in this application;

[0042] Figure 9 This is a schematic diagram of the structure of an embodiment of the vehicle-mounted computer device provided in this application;

[0043] Figure 10 This is a structural schematic diagram of an embodiment of the vehicle provided in this application. Detailed Implementation

[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0045] The terms "first," "second," and "third" used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0046] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0047] Autonomous driving technology is one of the most revolutionary advancements in the automotive industry in recent years. With continuous technological development, autonomous driving is no longer limited to vehicle driving control but has also expanded to vehicle parking systems. Memory parking, as a practical automatic parking technology, enables the vehicle to automatically remember and reproduce the driver's parking maneuvers, greatly simplifying the parking process.

[0048] participate Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the visual memory parking method based on deep learning networks provided in this application. The visual memory parking method includes:

[0049] Step 10: Obtain the preloaded map of the parking lot.

[0050] A pre-loaded map that fits a preset range is retrieved from the cloud for image matching and vehicle positioning during subsequent driving and parking processes.

[0051] Specifically, the preloaded maps include: a global map pieced together from feature maps during each learning process of the deep learning network, parking lot entrance information, and memorized parking routes.

[0052] Feature map concatenation is the operation of directly connecting different feature maps along a certain dimension to form a wider feature map. This method does not change the depth information of the feature maps, but increases the expressive power of the model by expanding the width of the feature maps.

[0053] For example, consider two different convolutional layer feature maps, each with the same spatial dimensions (height and width) but different numbers of channels (i.e., depth). Feature map A: dimensions are H x W x C1, where H is the height, W is the width, and C1 is the number of channels. Feature map B: dimensions are H x W x C2, where H and W are the same as feature map A, but the number of channels C2 may differ from C1. When concatenating row feature maps, the feature map is not expanded along the height or width dimension, but along the channel dimension. Specifically, feature map A and feature map B are connected along the channel dimension to form a new feature map. The dimensions of the concatenated feature map are H x W x (C1 + C2). In this process, the height H and width W remain unchanged, but the number of channels increases from C1 and C2 to C1 + C2. By concatenating different feature maps, information from different sources or different convolutional layers is combined, increasing the diversity of the input data and enabling the model to learn more complex patterns and structures.

[0054] In one or more stages of a deep learning network, feature maps from different levels or sources are concatenated at depth. This combines semantic and detailed information from different levels, enriching the feature representation. Through feature map concatenation, the network can fuse multi-scale and multi-level feature information to construct a global map. The global map not only contains local details from each feature map but also the overall semantic information of the multiple feature maps. By forming a global map through feature map concatenation, the global map can be better matched directly with the feature maps captured by the vehicle.

[0055] Parking lot entrance information is used to identify the parking lot a vehicle enters and its location upon entry. Specifically, parking lot entrance information may include: parking lot signage, entrance location information, number of entrances, entrance feature images, and entrance width and height, etc.

[0056] The remembered parking path is based on parking lot map information and is used to guide the vehicle through the automatic parking process.

[0057] Further, see Figure 2 Step 10 also includes:

[0058] Step 11: Obtain a preloaded map of the parking lot closest to the target vehicle from the cloud.

[0059] The system platform uses GPS or mobile device location data to locate the nearest parking lot while the vehicle is in motion and retrieves a pre-loaded map of that parking lot.

[0060] Optionally, the cloud database stores a large amount of parking lot information, including location, capacity, and fees. The system filters and sorts based on search criteria such as distance, price, or reviews. During this process, the system can prioritize matching parking lots that already have high-precision map data, based on the completeness of the map data recorded by each parking lot in the cloud, to ensure the accuracy and usability of the pre-loaded maps.

[0061] Once downloaded, the system will parse the map data and display it in the vehicle's navigation or parking assistance system.

[0062] Specifically, the preloaded map data for the parking lot includes information such as the overall layout, capacity, cost, current vehicle location, or recommended parking routes.

[0063] Optionally, the conditions for obtaining the preloaded map of the parking lot can also be: parking lots within a certain range of the target, such as within 100 meters, 150 meters, 200 meters, 300 meters or 500 meters, parking lots in the current vehicle's direction of travel or parking lots with a memory path stored in the cloud, etc.

[0064] Preloading technology involves loading tiled images outside the map view in advance. The map view is the area currently visible to the user on the map; it's the intuitive interface for user interaction and doesn't directly store map data. Instead, it retrieves data from the cloud and renders it based on user requests. Tiled images are map data fragmented into smaller pieces and stored remotely. When a user requests map data, the cloud first sends the corresponding tiled images, then combines them into a complete map view. Preloading technology reduces loading time when users pan or zoom on the map.

[0065] In addition, the system can intelligently recommend the most suitable parking lots and parking routes based on users' historical behavior and preferences, improving parking efficiency and convenience.

[0066] Step 12: Match the feature map of the target vehicle at the parking lot entrance with the parking lot entrance information to determine the parking lot the target vehicle entered and load the corresponding pre-loaded map.

[0067] When a vehicle arrives at the parking lot entrance, a feature map of the area surrounding the vehicle at that moment is captured. Based on the captured feature map, the landmarks of the target parking lot are identified. The identified landmarks are matched with parking lot entrance information in the database to confirm the parking lot the vehicle has entered, and the corresponding pre-loaded map of that parking lot is loaded.

[0068] Optionally, the signage for the target parking lot may be a parking lot QR code, parking lot sign, parking lot lane lines, landmark buildings, or other signage with identifiable features.

[0069] For example, if a vehicle detects that the parking lot sign reads "People's Park Parking Lot Entrance," then the parking lot entrance currently detected by the vehicle is the entrance to the People's Park parking lot. As another example, if a vehicle detects that a company's logo is posted outside the security booth at the parking lot entrance, then the parking lot entrance currently detected by the vehicle may be the parking lot of that company's building.

[0070] Meanwhile, when matching a vehicle to a parking lot entrance, the exact time the vehicle arrived at the entrance can be obtained based on the acquisition time of the feature map that matches the parking lot entrance; the vehicle's pose at the current moment can be determined based on the position and angle of the feature map matched in the global map, which is used for the subsequent positioning of the vehicle after entering the parking lot.

[0071] Furthermore, after step 10, the method also includes: determining the initial real vehicle pose of the target vehicle on the preloaded map based on the feature map of the target vehicle and the parking lot entrance.

[0072] When matching a vehicle to a parking lot entrance, the accurate time and pose of the vehicle at the entrance are obtained to form the initial true vehicle pose for subsequent vehicle localization after entering the parking lot. Based on the initial true vehicle pose, a local map is cropped for subsequent vehicle localization and image matching.

[0073] Since the vehicle localization process in the parking lot at each moment requires predicting the virtual vehicle pose based on the real pose of the previous moment, and then obtaining the real vehicle pose through image alignment, the vehicle pose at the parking lot entrance is used as the initial real vehicle pose at the initial moment when entering the parking lot, and is used for subsequent position prediction and image alignment processes.

[0074] Optionally, the methods for obtaining the preloaded map of the parking lot that the vehicle entered may also include: third-party map service APIs such as Baidu Maps, Tencent Maps, or Gaode Maps; downloaded and integrated local map data such as maps stored on connected storage devices, map files output by the parking lot's local electronic map system or map editing tools; or other network-based map services such as cloud-based system server map information or browser online maps.

[0075] Step 20: Obtain the feature map of the target vehicle at the current moment and the actual vehicle pose of the target vehicle on the preloaded map at the previous moment.

[0076] The vehicle's vision equipment acquires environmental semantic information such as lane lines, parking lines, ground markings, wheel chocks, speed bumps, and the outlines of obstacles around the target vehicle, and projects this information into a bird's-eye view (BVE) to form a feature map of semantics centered on the target vehicle at the current moment.

[0077] Specifically, the bird's-eye view image stitching adopts an end-to-end network structure, ignoring the image stitching of the panoramic cameras, and directly outputs the bird's-eye view perception results from independent multi-view camera perspectives. Through the end-to-end network structure, the model directly extracts features from the original images of each camera and integrates them in the bird's-eye view space. This allows the model to automatically learn how to integrate information from different cameras without relying on precise intrinsic and extrinsic parameter calibration, reducing dependence on intrinsic and extrinsic parameter calibration and minimizing the impact of calibration errors on the final result.

[0078] The description of a real vehicle's map pose includes both its position and attitude. Position is represented by three coordinate values ​​(x, y, z) in a specific coordinate system, while attitude is described by three rotational directions (roll, pitch, and yaw) and their corresponding angle values. This method of description accurately reflects the vehicle's actual position and attitude in three-dimensional space.

[0079] By combining the feature map of the target vehicle at the current moment with the actual vehicle pose obtained from the preloaded map at the previous moment, the location of the vehicle in the map at the next moment can be calculated.

[0080] Starting with the initial true vehicle pose of the target vehicle at the parking lot entrance, the true vehicle pose of the target vehicle on the preloaded map can be obtained at each subsequent moment.

[0081] Step 30: Based on the target vehicle's real vehicle pose on the preloaded map at the previous moment, predict the target vehicle's virtual vehicle pose on the preloaded map at the current moment.

[0082] Based on the target vehicle's actual vehicle pose on the preloaded map at the previous time step, a local map corresponding to the actual vehicle pose at the previous time step is cropped from the map. This local map is used for the vehicle matching process on the map at the next time step. Based on the cropped local map, the virtual vehicle pose (map pose) of the target vehicle on the preloaded map at the current time step is predicted.

[0083] One of the most common sources for predicting the current vehicle pose is a motion model. Optionally, the motion model can be provided by an inertial measurement unit (IMU) or obtained through simpler methods such as wheel odometers or constant-speed models. Based on the actual vehicle pose at the previous moment, the motion model can be used to provide an initial guess of the current vehicle pose. This initial guess can include small movements in position and minor changes in attitude. Prediction accuracy depends on various factors, including the accuracy of the motion model, the reliability of the sensor data, and the effectiveness of iterative optimization. In practical applications, prediction parameters and algorithms need to be adjusted according to specific circumstances to obtain more accurate prediction results.

[0084] For example, the actual vehicle pose of the target vehicle on the preloaded map in the previous frame is obtained. Based on the image of the previous frame and the vehicle's motion trajectory and speed, the position and angle of the vehicle in the next frame are predicted, and the predicted position and angle are used as the virtual vehicle pose of the target vehicle on the preloaded map at the current moment.

[0085] For example, based on the vehicle's pose in the images captured within the first 3 seconds, the vehicle's motion trajectory is generated. The real vehicle pose at the moment the last image was generated is used as the reference pose. The vehicle's motion trajectory is then used to simulate motion from the reference pose to form the virtual vehicle pose of the target vehicle on the preloaded map at the current moment.

[0086] Furthermore, the visual memory parking method also includes: collecting historical driving data of the target vehicle and data output by the inertial measurement unit at the current moment to generate the odometry pose of the target vehicle at the current moment.

[0087] Historical driving data can include parameters such as wheel speed or steering wheel angle during vehicle operation, which are used to construct a motion model of the target vehicle and predict the vehicle's position and posture at the current moment.

[0088] An inertial measurement unit (IMU) is a device capable of measuring an object's three-axis attitude angles or angular rates, as well as acceleration. It can provide real-time data such as a vehicle's acceleration, angular velocity, and attitude angles. At any given moment, the IMU can acquire real-time data on the target vehicle's acceleration, angular velocity, and attitude angles.

[0089] Historical driving data and the data output by the inertial measurement unit at the current moment can be used to adjust the vehicle's positioning results to obtain more accurate positioning predictions.

[0090] Furthermore, step 30 also includes: predicting the virtual vehicle pose of the target vehicle on the preloaded map at the current moment based on the actual vehicle pose of the target vehicle on the preloaded map at the previous moment and the odometer pose of the target vehicle at the previous moment.

[0091] Based on the vehicle's position and orientation at the previous moment, using historical driving data and data output from the inertial measurement unit at the current moment, the displacement and orientation of the vehicle relative to its starting point at the previous moment are calculated.

[0092] Specifically, by calculating the data output by the inertial measurement unit (IMU), the displacement changes of the target vehicle in the three directions (X, Y, Z) compared to the previous moment can be obtained. The angular velocity in the IMU output data is integrated to obtain the vehicle's directional change. Adding the calculated displacement and directional changes to the actual vehicle pose (including position and orientation) from the previous moment yields the predicted virtual vehicle pose for the current moment. Since the integration process of the IMU data may lead to error accumulation, especially over long periods or with complex trajectories, it is necessary to periodically correct the IMU calculation results using constant motion data.

[0093] By combining historical driving data with the data output from the inertial measurement unit at the current moment, the vehicle's positioning results can be adjusted to obtain more accurate positioning predictions.

[0094] Step 40: Based on the predicted virtual vehicle pose of the target vehicle on the preloaded map at the current moment, crop out the local map of the target vehicle on the preloaded map.

[0095] Centered on the predicted current position of the vehicle on the preloaded map, a portion of the map is cropped according to a preset cropping range to obtain a local map of the target vehicle on the preloaded map.

[0096] For example, in an underground parking garage with a three-level structure, each level measuring 300m x 500m, a vehicle is located at a specific position on the second level. From the parking garage map on the second level, a rectangular map measuring 6m x 8m is cropped out with the vehicle's position as the geometric center to serve as a local map for subsequent matching.

[0097] Alternatively, the map can be cropped in the following ways: cropping a circle with a radius of 5 meters centered on the vehicle's location; cropping a rectangle with a border of 3.5 meters in front of the vehicle's front, 2.5 meters behind the vehicle's rear, and 3 meters outside the sides of the vehicle's doors; or other cropping methods that can obtain a local map that facilitates local matching.

[0098] Using cropped local maps for matching can reduce the amount of data that vehicles need to calculate during map matching, refine the accuracy of map matching, improve the alignment between feature maps and local maps, and improve the accuracy and efficiency of local vehicle positioning on the map.

[0099] Step 50: Align the feature map and local map under the deep learning network, and output the real vehicle pose of the target vehicle in the preloaded map at the current time.

[0100] The input to a deep learning network for image alignment is a feature map and a local map. The output after alignment is an aligned feature map and a pose transformation matrix. Based on the aligned feature map and pose transformation matrix, the predicted vehicle pose at the current time is corrected to obtain a more accurate true vehicle pose of the target vehicle in the preloaded map at the current time.

[0101] Feature learning networks typically include pyramid-shaped feature hierarchy methods for object detection and image segmentation, feature alignment modules for feature matching and image transformation, and deep feature synthesis features that fuse low-level and high-level features. When training and testing deep learning models, a training set is used to train the model, a validation set is used to tune the model's parameters and hyperparameters, and a test set is used to evaluate the model's performance. It is crucial to ensure that the dataset is appropriately divided to avoid overfitting or underfitting, and to select appropriate activation and loss functions for computation.

[0102] By combining methods such as feature pyramids, feature alignment modules, and deep feature synthesis, along with the proper use of training, validation, and test sets, and the selection of appropriate activation and loss functions, better performance can be achieved.

[0103] Further, see Figure 3 , Figure 3 This is a flowchart of the steps in Embodiment 2 of the deep learning-based vision and parking method provided in this application. Step 50 further includes:

[0104] Step 51: Align the feature map and the local map under the deep learning network to obtain the vehicle pose of the target vehicle in the local map.

[0105] First, the feature map and local map are input into the deep learning network, and the target vehicle's pose in the local map is output.

[0106] Specifically, the deep learning network includes a first spatial transformation network (STN), a second spatial transformation network (STN), and a target ground value (target gt).

[0107] Spatial Transformer Networks (STNs) are deep learning network architectures used to perform spatial transformations on input data within a neural network, such as translation, rotation, scaling, or shearing. This allows the network to learn how to locate or recognize targets under these transformations. STNs insert differentiable modules into traditional convolutional neural networks, allowing the network to dynamically transform input data to adapt to objects of different scales and shapes.

[0108] The target is the output value that the model learns and predicts. For classification tasks, the target might be the label of a certain category; for regression tasks, the target might be a continuous numerical value. For example, in an image classification task, if there is an image of lane lines and we know that the image belongs to the category "lane markings," then "lane markings" is the target of this image. As another example, in an image classification task, if there is an image of parking space "N165" and we know that the image belongs to the category "parking space number," then "parking space number" is the target of this image.

[0109] The ground truth is the true value or label provided in the dataset. When training the model, the model's predictions are compared with the ground truth to calculate the loss and update the model parameters.

[0110] In the learning process of a deep learning model, the accuracy of predictions by the deep learning module is improved by minimizing the difference between the predicted value and the true value.

[0111] Specifically, see Figure 4 Step 51 specifically includes:

[0112] Step 511: Input the feature map and local map into the first spatial transformation network STN to obtain the first feature map warp and the first transformation parameter estimate (T_estimate).

[0113] The input feature map and local map are fed into the first spatial transformation network (STN) for coarse image matching. Between the STN-processed feature map and the local map, a similarity metric such as cosine similarity or Euclidean distance is applied for coarse matching. This helps to quickly filter out candidate regions similar to the local map, providing a foundation for subsequent fine matching operations.

[0114] The core of STN (Spatial Transformer) is learning the spatial transformation parameters of the input feature map. These parameters define a transformation matrix used to adjust the spatial position of the feature map. By applying the transformation matrix, STN can transform the input feature map to a new spatial location, making it better aligned with the local map. STN can perform transformations on the entire feature map and can include operations such as scaling, cropping, rotation, and non-rigid deformation. This allows STN to handle input feature maps of different scales and shapes and transform them to a unified coordinate system, thereby simplifying subsequent matching operations.

[0115] Transformation parameter estimation describes how to spatially transform the input feature map, while feature map warping is the output feature map obtained after spatially transforming the input feature map according to these transformation parameters. These two outputs together constitute the core functionality of STN, enabling it to achieve spatial invariance and improve robustness to different spatial transformations.

[0116] Step 512: Input the first feature map distortion and feature map into the second spatial transformation network STN to obtain the second feature map distortion and the second transformation parameter estimate.

[0117] The feature map distortion and feature map after coarse matching by the first spatial transformation network STN are input into the second spatial transformation network STN for further matching. The second spatial transformation network STN outputs the second feature map distortion and the second transformation parameter estimate after further matching.

[0118] Through two transformations, a more accurate processed image, namely the second feature map distortion, is obtained. Feature map distortion is the core output of STN, containing feature information after spatial transformation. This information is more useful and easier to process for subsequent network layers. By using feature map distortion, STN can achieve spatial invariance of the model, improving its robustness to different spatial transformations.

[0119] Step 513: Based on the second feature map distortion, the first transformation parameter estimation, the second transformation parameter estimation, and the target ground truth, form the total residual of the deep learning network.

[0120] After matching the feature maps and correcting the ground truth values ​​using the deep learning network, the total residual output by the deep learning network is obtained. This residual corrects the predicted vehicle pose of the target vehicle at the current time, which helps to improve the accuracy of localization.

[0121] Specifically, the target ground truth includes the ground truth map and the ground truth pose.

[0122] The ground truth map represents the true distribution of specific information or attributes in an image and is used to train and optimize the network; while the ground truth pose represents the true position and orientation of the target object or camera in three-dimensional space and is used to guide the network to learn how to predict poses and achieve accurate image alignment.

[0123] By combining ground truth maps and ground truth poses, the performance and accuracy of image alignment deep learning networks can be improved.

[0124] See Figure 5 , Figure 5 This is a flowchart illustrating Embodiment 5 of the visual memory parking method based on deep learning networks provided in this application. Step 513 further includes:

[0125] Step 5131: Based on the distortion of the truth map and the second feature map, a first residual is formed.

[0126] The second feature map output by the second spatial transformation network STN is distorted and forms the first residual of the deep learning network with the ground truth map.

[0127] The first residual is the difference between the distortion of the second feature map output by the second spatial transformation network STN and the ground truth map. This difference can be quantified by calculating some distance or loss function between the two, and used as a reference metric for optimizing deep learning networks.

[0128] Deep learning networks continuously adjust the parameters of the first residual through optimization methods such as backpropagation and gradient descent to minimize the first residual, thereby improving the network's performance in image alignment or feature extraction tasks.

[0129] Step 5132: Based on the product of the true pose and the first and second transformation parameter estimates, form the second residual.

[0130] The product of the first transformation parameter estimate output by the first spatial transformation network STN and the second transformation parameter estimate output by the second spatial transformation network STN is used to form the second residual with the coordinate vector parameters of the true pose.

[0131] The second residual is the difference between the estimated transformation parameters output by the first and second spatial transformation networks (STN) and the true pose parameters.

[0132] Deep learning networks improve their performance on image alignment or feature extraction tasks by continuously adjusting the parameters of the second residual to minimize it.

[0133] Step 5132: Sum the first residual and the second residual to obtain the total residual of the deep learning network.

[0134] The sum of the first residual and the second residual is taken as the total residual of the deep learning network as a whole.

[0135] The total residual is the difference between the true pose and the region pose. The total residual can be used to correct the virtual vehicle pose of the target vehicle at the current time, thereby improving the accuracy of the vehicle's pose in the map at the current time.

[0136] Step 514: Based on the target ground truth and the total residual, calculate the regional pose of the target vehicle in the local map.

[0137] The total residual is the difference between the ground truth pose and the region pose. By using the total residual and the virtual vehicle pose of the target vehicle at the current moment, we can obtain the accurate real vehicle pose of the target vehicle at the current moment as a reference center for cropping the local map.

[0138] Step 52: Based on the clipping position of the local map in the preloaded map, convert the vehicle pose in the local map into the vehicle pose in the preloaded map.

[0139] Obtain the clipping position of the local map in the preloaded map, use the vehicle pose of the local map as the reference center point, restore the local map to the preloaded map, and transform the vehicle pose in the local map into the preloaded map.

[0140] Step 60: Based on the real vehicle pose in the preloaded map and the selected target parking space, obtain the parking path corresponding to the target parking space from the cloud memory and park the car.

[0141] During vehicle cruising, the system updates the memorized path and target parking space in real time. When a target parking space is selected, it retrieves the corresponding parking path from the cloud memory and guides the vehicle to park automatically based on the actual vehicle position of the target vehicle.

[0142] Specifically, after selecting a target parking space, the onboard system sends a request to the cloud to obtain a pre-planned parking path for that space. The cloud system, based on information such as the target parking space's location, vehicle type, and vehicle size, retrieves the most suitable parking path from its database and sends it to the vehicle. Upon receiving the parking path from the cloud, the onboard system, considering the vehicle's current position, orientation, and speed, plans and adjusts the path accordingly. The system then guides the vehicle automatically along the planned path by controlling its steering, acceleration, and braking. During the journey, onboard cameras continuously monitor the surrounding environment to ensure the vehicle safely and accurately reaches the target parking space.

[0143] By using cloud memory to obtain the parking path corresponding to the target parking space, parking convenience and safety are improved.

[0144] Further, see Figure 6 Step 60 also includes:

[0145] Step 61: Obtain the parking path corresponding to the target parking space from the cloud memory.

[0146] In a memory parking system, the parking path and related information remembered by the vehicle are stored in the cloud. Even if the vehicle changes location or a long time has passed, the vehicle can still retrieve this memory information from the cloud when it automatically parks according to the remembered path.

[0147] The parking path information stored in the cloud can specifically include: driving trajectory, vehicle speed, steering angle, etc., which is used to accurately reproduce the previous parking process.

[0148] By retrieving parking routes corresponding to target parking spaces from cloud memory, existing parking data is used to assist drivers in parking, reducing errors caused by human operation and improving parking safety.

[0149] Step 62: Convert the parking path into a parking path in odometer coordinates based on the target vehicle's current vehicle pose and odometer pose.

[0150] When a driver uses the memory parking function, the system will assist the driver in parking the vehicle into a parking space that has been memorized by the system, based on the previously memorized route.

[0151] Since the memory path is the vehicle pose on the map, it is necessary to use the current vehicle pose and odometer pose to convert the memory path to the odometer coordinate system so that downstream modules can read it and complete the parking operation.

[0152] Step 63: Guide the target vehicle to park based on the parking path under the odometer coordinates.

[0153] An odometer coordinate system is established using the vehicle's actual position and orientation before it begins to move as the origin. During vehicle movement, sensors such as wheel speed pulse sensors measure the distance and direction of each trip. The odometer continuously collects information such as mileage, acceleration, and angular velocity. Based on the collected data, a kinematic model calculates the vehicle's position change relative to the previous moment, and these changes are accumulated to obtain the vehicle's current position relative to the initial moment. Based on the current vehicle position and the target parking space position, combined with the vehicle's kinematic characteristics such as turning radius and maximum speed, the optimal path from the current position to the target parking space is planned for parking.

[0154] Parking paths based on odometer coordinates can effectively guide target vehicles during parking operations. This improves parking accuracy and efficiency while reducing driver fatigue and operational difficulty.

[0155] See Figure 7 , Figure 7 This is a flowchart illustrating Embodiment Seven of the visual memory parking method based on deep learning networks provided in this application. The visual memory parking method further includes:

[0156] Step 70: Convert the parking path in the odometer coordinates memorized during the parking process of the target vehicle into a parking path in the global map coordinate system.

[0157] Record the parking path and corresponding parking space and parking lot information of the target vehicle during each parking process, and convert this information from the parking path under the odometer coordinates of the vehicle to the parking path under the global map coordinate system.

[0158] The odometer coordinate system uses the vehicle's current actual vehicle pose as the initial position and origin. The vehicle's current position and attitude are determined by the displacement information measured by the odometer during vehicle movement, which serves as the position information under the odometer.

[0159] The global map coordinate system is based on the parking lot map information and does not change due to vehicle movement. Since the parking routes stored in the cloud are included in the parking lot map information, it is necessary to convert the acquired stored routes from the odometer coordinate system to the map coordinate system so that they can be retrieved along with the parking lot map information.

[0160] When a user uses the system for the first time, they need to manually drive the vehicle through the target parking lot or environment, or follow the automatic parking guidance to complete a parking route learning process. Specifically, this includes the vehicle scanning and recording information about the surrounding environment, such as parking lines, obstacles, and turning points. After the driver manually drives the vehicle through the parking route or performs automatic parking according to the guidance, the vehicle completes the "route memorization" process so that it can autonomously recognize, reproduce, or update the route in the future.

[0161] For parking spaces that already have a memorized path, automatic parking can be guided by vision. During the parking process, the memorized path is corrected based on the actual parking environment. The vehicle scans and records the surrounding environment, such as parking lines, obstacles, and turning points, to enhance the matching ability of the deep learning model.

[0162] The parking path recorded in the odometer coordinates is converted to the global map coordinate system and stored together with the parking lot map. This allows the remembered path to be retrieved the next time the parking lot map is accessed. Updating the parking path in the global map coordinate system allows it to be provided to the user simultaneously with the map being accessed. Therefore, converting the remembered parking path to the global map coordinate system facilitates data storage and simplifies subsequent retrieval.

[0163] Step 80: Upload the parking path in the global map coordinate system to the cloud and update the preloaded map of the corresponding parking lot.

[0164] The system memorizes the parking route and updates it to the pre-loaded map of the corresponding parking lot in the cloud. When the driver needs to use the memory parking function again, the system will assist the driver in parking the vehicle into the parking space that has been memorized by the system, based on the previously memorized route.

[0165] Furthermore, the cloud platform optimizes and adjusts the parking path in the parking space based on the newly added parking path, and uses the optimized parking path as the parking path for the next automatic parking in the parking space.

[0166] For example, parking space "063" in a parking lot has its memory path stored in the system's cloud. When a driver selects parking space 063 for automatic parking, the system retrieves the memory path for that parking space from the cloud and parks accordingly. However, a new pillar is added next to parking space 063. To avoid the pillar during automatic parking, the vehicle changes its original parking path. After parking is complete, the updated path is stored in the cloud for the portion of parking space 063 in that parking lot.

[0167] Since the data recorded by this system is primarily based on vision, the historical data uploaded to the cloud can also be used by vehicles with vision on other platforms, breaking the limitation that traditional memory data cannot be shared by vehicles on different platforms due to different sensor types.

[0168] See Figure 8 , Figure 8 This is a schematic diagram of an embodiment of the storage medium provided in this application.

[0169] The storage medium 300 stores program data 301, which, when executed by the processor, implements, as follows: Figure 1 The steps of the visual memory parking method described.

[0170] The program data 301 is stored in a storage medium 300 and includes several instructions for causing a network device (which may be a router, personal computer, server, or other network device) or processor to execute all or part of the steps of the methods described in the various embodiments of this application.

[0171] Optionally, the storage medium 300 can be any medium capable of storing program data, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), disk, or optical disc.

[0172] See Figure 9 , Figure 9 This is a schematic diagram of the structure of an embodiment of the vehicle-mounted computer device provided in this application.

[0173] The vehicle-mounted computer device 400 is installed in the vehicle and works in conjunction with various sensor modules and control modules in the vehicle to realize the functions of path memory and automatic parking. The vehicle-mounted computer device 400 includes a processor 420 and a memory 410 that are connected to each other. The memory 410 stores a computer program. When the processor 420 executes the computer program, it implements the steps of the visual memory parking method described above.

[0174] A vehicle 500 includes an onboard computer device 400 as described above. During operation, the vehicle 500 can complete a visual memory automatic parking process under the control of the onboard computer device 400.

[0175] Unlike existing technologies, this application provides a deep learning-based visual memory parking method, medium, device, and vehicle. This application constructs an initial vehicle position for reference by acquiring a pre-loaded map of the parking lot and the feature map of the target vehicle at the current moment. This eliminates the need for the system to memorize complex road topology information and semantic lanes; it only needs to maintain the memorized route and a lightweight semantic map, resulting in small storage space and low maintenance costs. Based on the initial vehicle and the target vehicle's actual vehicle pose on the pre-loaded map at the previous moment, the virtual vehicle pose of the target vehicle on the pre-loaded map at the next moment is predicted. The actual vehicle pose of the target vehicle on the pre-loaded map is then calculated and corrected. Using the predicted virtual vehicle pose of the target vehicle on the pre-loaded map at the current moment, a local map of the target vehicle on the pre-loaded map is cropped to improve the accuracy of the target vehicle's position matching on the map. Aligning the feature map containing environmental information and the cropped local map under a deep learning network yields a more accurate actual vehicle pose of the target vehicle on the pre-loaded map at the current moment, used in subsequent matching processes, further improving the vehicle's positioning accuracy on the map. The use of only visual sensing reduces the types of sensors required for the memory parking method. It also memorizes new parking paths during the parking process and updates cloud data for use by vehicles with vision on other platforms, thus improving the applicability of the memorized data.

[0176] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the storage medium embodiments and computer device embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0177] This application can be used in a wide range of general-purpose or specialized in-vehicle computing system environments or configurations. Examples include: personal computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, network PCs, minicomputers, and distributed computing environments including any of the above systems or devices.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative; multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed.

[0179] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A visual memory parking method based on deep learning networks, characterized in that, The visual memory parking method includes: Get the preloaded map of the parking lot; Obtain the feature map of the target vehicle at the current moment and the actual vehicle pose of the target vehicle on the preloaded map at the previous moment; Based on the actual vehicle pose of the target vehicle on the preloaded map at the previous moment, predict the virtual vehicle pose of the target vehicle on the preloaded map at the current moment; Based on the virtual vehicle pose of the target vehicle on the preloaded map at the current moment, a partial map of the target vehicle on the preloaded map is cropped out. Align the feature map and the local map using a deep learning network, and output the true vehicle pose of the target vehicle in the preloaded map at the current time. Based on the actual vehicle positions and the selected target parking space in the preloaded map, the parking path corresponding to the target parking space is obtained from the cloud memory and then used for parking.

2. The visual memory parking method according to claim 1, characterized in that, The step of predicting the virtual vehicle pose of the target vehicle on the preloaded map at the current moment based on the target vehicle's actual vehicle pose on the preloaded map at the previous moment includes: Based on the target vehicle's actual vehicle pose on the preloaded map at the previous moment and the target vehicle's odometer pose at the previous moment, predict the target vehicle's virtual vehicle pose on the preloaded map at the current moment.

3. The visual memory parking method according to claim 2, characterized in that, The visual memory parking method also includes: The historical driving data of the target vehicle and the data output by the inertial measurement unit at the current moment are collected to generate the odometer pose of the target vehicle at the current moment.

4. The visual memory parking method according to claim 1, characterized in that, The step of aligning the feature map and the local map under a deep learning network and outputting the real vehicle pose of the target vehicle in the preloaded map at the current time includes: The feature map and the local map are aligned using a deep learning network to obtain the vehicle pose of the target vehicle in the local map. Based on the cropping position of the local map in the preloaded map, the vehicle pose in the local map is converted into the vehicle pose in the preloaded map.

5. The visual memory parking method according to claim 4, characterized in that, The deep learning network includes a first spatial transformation network (STN), a second spatial transformation network (STN), and a target ground value; The step of aligning the feature map and the local map under a deep learning network to obtain the vehicle pose of the target vehicle in the local map includes: Input the feature map and the local map into the first spatial transformation network (STN) to obtain the first feature map distortion and the first transformation parameter estimate; Input the first feature map distortion and the feature map into the second spatial transformation network STN to obtain the second feature map distortion and the second transformation parameter estimate; Based on the second feature map distortion, the first transformation parameter estimation, the second transformation parameter estimation, and the target ground truth, the total residual of the deep learning network is formed; Based on the target ground truth and the total residual, the regional pose of the target vehicle in the local map is calculated.

6. The visual memory parking method according to claim 5, characterized in that, The target truth value includes a truth map and a truth pose; The step of forming the total residual of the deep learning network based on the second feature map distortion, the first transformation parameter estimation, the second transformation parameter estimation, and the target ground truth includes: Based on the distortion of the truth map and the second feature map, a first residual is formed; The second residual is formed based on the product of the true pose and the first and second transformation parameter estimates. The total residual of the deep learning network is obtained by summing the first residual and the second residual.

7. The visual memory parking method according to claim 1, characterized in that, The preloaded map includes: a global map stitched together from the feature maps during each learning process of the deep learning network, parking lot entrance information, and memorized parking routes.

8. The visual memory parking method according to claim 7, characterized in that, The process of obtaining the preloaded map of the parking lot includes: Retrieve a preloaded map of the parking lot closest to the target vehicle from the cloud; Match the feature map of the target vehicle at the parking lot entrance with the parking lot entrance information to determine the parking lot that the target vehicle entered and load the corresponding preloaded map.

9. The visual memory parking method according to claim 8, characterized in that, After obtaining the preloaded map of the parking lot, the process also includes: Based on the feature map of the target vehicle and the parking lot entrance, the initial true vehicle pose of the target vehicle on the preloaded map is determined.

10. The visual memory parking method according to claim 1, characterized in that, The step of obtaining the parking path corresponding to the target parking space from cloud memory and parking the car includes: Obtain the parking path corresponding to the target parking space from the cloud memory; The parking path is converted into a parking path in odometer coordinates based on the current vehicle pose and odometer pose of the target vehicle. Based on the parking path under the odometer coordinates, the target vehicle is guided to park.

11. The visual memory parking method according to claim 10, characterized in that, After parking the car using the parking path corresponding to the target parking space obtained from the cloud memory, the process further includes: The parking path in the odometer coordinates memorized during the parking process of the target vehicle is converted into a parking path in the global map coordinate system; The parking path in the global map coordinate system is uploaded to the cloud to update the preloaded map of the corresponding parking lot.

12. A storage medium storing program data thereon, characterized in that, When the program data is executed by the processor, it implements the steps of the visual memory parking method as described in any one of claims 1-11.

13. A vehicle-mounted computer device, characterized in that, It includes an interconnected processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the visual memory parking method as described in any one of claims 1-11.

14. A vehicle, characterized in that, The vehicle includes the on-board computer equipment as described in claim 13.

Images