An automatic parking method, an automatic parking device, and a vehicle
By using an automated parking model based on conditional generative adversarial networks, the problem of poor adaptability of automated parking in existing technologies is solved, and safety prediction and optimization in multiple scenarios are achieved, thereby improving parking efficiency and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING JINGWEI HIRAIN TECH CO INC
- Filing Date
- 2023-12-19
- Publication Date
- 2026-06-05
AI Technical Summary
Existing automatic parking technology has poor adaptability, struggles to cope with unexpected situations, requires high sensor perception and computing power, results in unsatisfactory parking performance, and leads to a poor user experience.
An automatic parking model trained using a conditional generative adversarial network is adopted. Initial information is obtained through a vehicle data acquisition module, preprocessed, and then input into the model to achieve prediction and control of parking results. The model can be continuously optimized to adapt to different scenarios.
It improves the fault tolerance and efficiency of automatic parking, can predict multiple parking scenarios in advance, optimize the model to adapt to new environments, and provide users with a good parking experience.
Smart Images
Figure CN117681862B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of autonomous driving technology, and in particular to an automatic parking method, an automatic parking device, and a vehicle. Background Technology
[0002] With the rapid development of artificial intelligence technology, automatic parking technology is becoming increasingly popular. Automatic parking refers to the automatic parking of a car without human control. Existing automatic parking technology mainly relies on sensors distributed throughout the vehicle and its surrounding environment to acquire real-time data (such as the relative distance between the vehicle and surrounding objects). Then, various algorithms are used on the in-vehicle cloud platform to perform real-time calculations and control the vehicle system's steering or acceleration and deceleration, thereby achieving automatic parking and exiting of the vehicle.
[0003] However, current automatic parking systems have poor adaptability. Because the automatic parking process involves real-time detection and continuous environmental awareness, it places high demands on sensor perception and computing power. Current automatic parking functions are not ideal for processing data in unexpected situations, resulting in poor parking performance and an inability to handle unforeseen circumstances. In summary, current automatic parking systems are inefficient and perform poorly, leading to a negative user experience. Summary of the Invention
[0004] This application provides an automatic parking method, an automatic parking device, and a vehicle, which can improve automatic parking efficiency and bring a good user experience.
[0005] In a first aspect, this application provides an automatic parking method, the method comprising:
[0006] The vehicle-based data acquisition module acquires initial vehicle information, which is used to characterize the initial state of the vehicle.
[0007] The initial vehicle information is preprocessed to obtain the target vehicle information;
[0008] The target vehicle information is input into the automatic parking model to obtain the target parking result. The vehicle includes the automatic parking model, which is generated based on training with a conditional generative adversarial network.
[0009] In response to the operation of determining the target parking result, the vehicle is controlled to park.
[0010] Optionally, the automatic parking method further includes:
[0011] The conditional generative adversarial network is pre-trained based on a public parking dataset to obtain a pre-trained parking model.
[0012] The parameters of the pre-trained parking model are adjusted based on the training dataset to obtain the automatic parking model.
[0013] Optionally, the conditional generative adversarial network includes a generator and a discriminator, the training dataset includes conditional information and real samples, and the step of adjusting the parameters of the pre-trained parking model based on the training dataset to obtain the automatic parking model includes:
[0014] The condition information is input into the generator to obtain the generated sample;
[0015] The condition information and the sample to be input are input into the discriminator to obtain the parking result. The parking result is used to characterize the probability value that the sample to be input is the real sample and to characterize the matching degree between the sample to be input and the condition information. The sample to be input includes the real sample and the generated sample.
[0016] Calculate the loss function based on the parking results;
[0017] The automatic parking model is obtained by updating and iterating the parameters of the generator using the backpropagation algorithm and the loss function.
[0018] Optionally, the automatic parking method further includes:
[0019] The automatic parking model is optimized based on a preset optimization method to obtain an optimized parking model. The preset optimization method includes hyperparameter tuning, optimizer selection, or regularization.
[0020] Optionally, the automatic parking method further includes:
[0021] The optimized parking model is tested using the test dataset to obtain test results;
[0022] If the test results do not meet the preset requirements, the preset parking model is optimized using the preset optimization method to obtain an optimized parking model.
[0023] Optionally, the automatic parking method further includes:
[0024] A simulation test scenario is set up according to the test requirements. The simulation test scenario is used to simulate the parking environment in which the vehicle is located.
[0025] Create a simulation test dataset corresponding to the simulation test scenario and obtain the simulation target parking results;
[0026] Input the simulation test dataset into the automatic parking model to obtain the initial parking simulation results;
[0027] The initial parking result is evaluated based on the target parking result to obtain the evaluation result;
[0028] If the evaluation results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
[0029] Optionally, the automatic parking method further includes:
[0030] Obtain initial parking datasets for multiple parking scenarios;
[0031] The initial parking dataset is cleaned and preprocessed to obtain an intermediate parking dataset;
[0032] The intermediate parking dataset is processed using feature engineering techniques to obtain the target parking dataset.
[0033] The target parking dataset is divided according to a preset ratio to obtain a training dataset, a validation dataset, and a test dataset.
[0034] Optionally, the step of performing feature processing on the intermediate parking dataset based on feature engineering techniques to obtain the target parking dataset includes:
[0035] The intermediate parking dataset is processed using feature engineering techniques to obtain a feature parking dataset.
[0036] The dimensionality of the feature parking dataset is reduced to obtain the target parking dataset.
[0037] Secondly, this application also provides an automatic parking device, the device comprising:
[0038] An acquisition unit is used to acquire initial vehicle information based on the vehicle's data acquisition module, wherein the initial vehicle information is used to characterize the initial state of the vehicle.
[0039] The preprocessing unit is used to preprocess the initial vehicle information to obtain the target vehicle information;
[0040] The obtaining unit is used to input the target vehicle information into the automatic parking model and obtain the target parking result. The vehicle includes the automatic parking model, which is generated based on training of a conditional generative adversarial network.
[0041] A control unit is configured to control the vehicle to park in response to an operation that determines the target parking result.
[0042] Optionally, the automatic parking device further includes: a model training unit, the model training unit comprising:
[0043] The pre-training module is used to pre-train the conditional generative adversarial network based on a public parking dataset to obtain a pre-trained parking model.
[0044] The training module is used to adjust the parameters of the pre-trained parking model based on the training dataset to obtain the automatic parking model.
[0045] Optionally, the conditional generative adversarial network includes a generator and a discriminator, the training dataset includes conditional information and real samples, and the training module is specifically used for:
[0046] The condition information is input into the generator to obtain the generated sample;
[0047] The condition information and the sample to be input are input into the discriminator to obtain the parking result. The parking result is used to characterize the probability value that the sample to be input is the real sample and to characterize the matching degree between the sample to be input and the condition information. The sample to be input includes the real sample and the generated sample.
[0048] Calculate the loss function based on the parking results;
[0049] The automatic parking model is obtained by updating and iterating the parameters of the generator using the backpropagation algorithm and the loss function.
[0050] Optionally, the training module is further configured to:
[0051] The automatic parking model is optimized based on a preset optimization method to obtain an optimized parking model. The preset optimization method includes hyperparameter tuning, optimizer selection, or regularization.
[0052] Optionally, the training module is further configured to:
[0053] The optimized parking model is tested using the test dataset to obtain test results;
[0054] If the test results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
[0055] Optionally, the training module is further configured to:
[0056] A simulation test scenario is set up according to the test requirements. The simulation test scenario is used to simulate the parking environment in which the vehicle is located.
[0057] Create a simulation test dataset corresponding to the simulation test scenario and obtain the simulation target parking results;
[0058] Input the simulation test dataset into the automatic parking model to obtain the initial parking simulation results;
[0059] The initial parking result is evaluated based on the target parking result to obtain the evaluation result;
[0060] If the evaluation results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
[0061] Optionally, the automatic parking device further includes a data acquisition unit, which is specifically used for:
[0062] Obtain initial parking datasets for multiple parking scenarios;
[0063] The initial parking dataset is cleaned and preprocessed to obtain an intermediate parking dataset;
[0064] The intermediate parking dataset is processed using feature engineering techniques to obtain the target parking dataset.
[0065] The target parking dataset is divided according to a preset ratio to obtain a training dataset, a validation dataset, and a test dataset.
[0066] Optionally, the step of performing feature processing on the intermediate parking dataset based on feature engineering techniques to obtain the target parking dataset includes:
[0067] The intermediate parking dataset is processed using feature engineering techniques to obtain a feature parking dataset.
[0068] The dimensionality of the feature parking dataset is reduced to obtain the target parking dataset.
[0069] Thirdly, this application also provides a vehicle that includes the automatic parking device provided in the second aspect above.
[0070] Therefore, this application has the following beneficial effects:
[0071] This application provides an automatic parking method, an automatic parking device, and a vehicle. The method includes: acquiring initial vehicle information through a vehicle data acquisition module; preprocessing the initial vehicle information to obtain target vehicle information; inputting the target vehicle information into an automatic parking model to obtain a target parking result. The vehicle includes the automatic parking model, which is generated based on conditional generative adversarial network training; and controlling the vehicle to park in response to an operation that determines the target parking result. Thus, by acquiring vehicle information through the vehicle data acquisition module and inputting it into the automatic parking model to obtain a parking result, automatic parking is achieved. Furthermore, the automatic parking model, pre-generated based on conditional generative adversarial network training, can predict the safety of multiple parking scenarios in advance, improving the fault tolerance of automatic parking. In addition, the automatic parking model can continuously collect vehicle information to optimize and adapt to new parking scenarios, improving automatic parking efficiency and providing a better user experience. Attached Figure Description
[0072] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.
[0073] Figure 1 This is a flowchart illustrating an automatic parking method according to an embodiment of this application;
[0074] Figure 2 This is a flowchart illustrating one embodiment of an automatic parking method according to this application.
[0075] Figure 3 A model structure diagram of the automatic parking model provided in the embodiments of this application;
[0076] Figure 4 This is a schematic diagram of the structure of an automatic parking device 400 provided in an embodiment of this application. Detailed Implementation
[0077] The "multiple" mentioned in the embodiments of this application refers to two or more. It should be noted that in the description of the embodiments of this application, terms such as "first" and "second" are used only for the purpose of distinguishing descriptions and should not be construed as indicating or implying relative importance, nor should they be construed as indicating or implying order.
[0078] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the embodiments of this application will be further described in detail below with reference to the accompanying drawings and specific implementation methods. It should be understood that the specific embodiments described herein are merely for explaining this application and are not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this application are shown in the accompanying drawings, not the entire structure.
[0079] Current automated parking technologies exhibit poor adaptability. Because automated parking involves real-time detection and continuous environmental perception, it places high demands on sensor perception and computing power. Existing automated parking functions are less than ideal at handling unexpected situations, resulting in poor parking performance and an inability to cope with unforeseen circumstances. Furthermore, users must select a parking space in a semi-enclosed area such as a park or residential area, or a parking lot (covered by a high-precision map), via the in-vehicle central control screen or mobile app. The vehicle then acquires information about lane markings, road signs, surrounding vehicles, and other pedestrians on these semi-enclosed roads. This places very high demands on the user; without user guidance, the car struggles to obtain additional information to assist in parking. In conclusion, current automated parking technologies are inefficient and perform poorly, leading to a negative user experience.
[0080] Based on this, embodiments of this application provide an automatic parking method, which includes: acquiring initial vehicle information based on a vehicle data acquisition module; preprocessing the initial vehicle information to obtain target vehicle information; inputting the target vehicle information into an automatic parking model to obtain a target parking result, wherein the vehicle includes the automatic parking model, and the automatic parking model is generated based on conditional generative adversarial network training; and controlling the vehicle to park in response to the operation of determining the target parking result.
[0081] In this way, vehicle information is obtained through the vehicle data acquisition module and input into the automatic parking model to obtain parking results and achieve automatic parking. Moreover, the automatic parking model is pre-trained based on conditional generative adversarial networks, which can predict the safety of multiple parking scenarios in advance and improve the fault tolerance of automatic parking. In addition, the automatic parking model can continuously collect vehicle information to optimize and adapt to new parking scenarios, improve the efficiency of automatic parking, and bring a good user experience.
[0082] First, let me explain the terms that may be used in this application:
[0083] 1) Automatic parking refers to a car automatically parking itself without manual control;
[0084] 2) Generative Adversarial Networks (GANs) are a type of deep learning model that produces fairly good outputs through the game-like learning between (at least) two modules in the framework: a generative model and a discriminative model. The embodiments of this application use Conditional Generative Adversarial Networks (CGANs).
[0085] 3) The generator is used to generate simulated data and continuously optimize it so that the discriminator cannot distinguish it as simulated data. Its main function is to generate samples with the same distribution from the training data. For an input x and a class label y, it estimates the joint probability distribution (the probability distribution of a random vector composed of two or more random variables) in the generative model.
[0086] 4) Discriminator, used to determine the conditional probability distribution of an input sample belonging to a certain class;
[0087] 5) The loss function is used to measure the degree of inconsistency between the model's predicted value f(x) and the true value Y. It is a non-negative real-valued function, usually represented by L(Y,f(x)). The smaller the loss function, the better the robustness of the model.
[0088] To facilitate understanding of the specific implementation of the automatic parking method provided in the embodiments of this application, the following description will be provided in conjunction with the accompanying drawings.
[0089] It should be noted that the main body implementing this automatic parking method can be the automatic parking device provided in the embodiments of this application, or it can be carried in an electronic device or a functional module of an electronic device. The electronic device in the embodiments of this application can be any device capable of implementing the automatic parking method in the embodiments of this application, such as an Internet of Things (IoT) device.
[0090] Please see Figure 1 , Figure 1 This is a flowchart illustrating an automatic parking method provided in an embodiment of this application. This method can be applied to an automatic parking device, which may be, for example, an automatic parking device... Figure 4 The automatic parking device 400 shown.
[0091] In this embodiment of the application, the following steps may be included, for example:
[0092] S101: The vehicle-based data acquisition module acquires initial vehicle information.
[0093] It should be noted that the data acquisition module may include sensors (such as ultrasonic sensors, cameras, inertial sensors, radar sensors, or lidar sensors) or positioning systems (such as Global Positioning System (GPS)). The data acquisition module can acquire vehicle information in multiple parking scenarios (such as parallel parking, perpendicular parking, or angled parking). Initial vehicle information is used to characterize the initial state of the vehicle. In some implementations, the initial vehicle information is unprocessed vehicle information obtained from the data acquisition module. The vehicle information may include the vehicle's own state information or road condition information of the vehicle's surrounding environment. For example, it may be information about the position of objects around the vehicle, the vehicle's speed, direction, and the surrounding environment. Of course, the vehicle information may also be other information used to obtain the parking result of the vehicle, which does not affect the implementation of the embodiments of this application.
[0094] S102: Preprocess the initial vehicle information to obtain the target vehicle information.
[0095] In some implementations, preprocessing may include data normalization, standardization, and feature scaling to ensure data consistency and comparability. Target vehicle information may refer to preprocessed information that meets the requirements of the input automatic parking model.
[0096] S103: Input the target vehicle information into the automatic parking model to obtain the target parking result.
[0097] It should be noted that the vehicle includes an automatic parking model, which is generated based on training with a conditional generative adversarial network.
[0098] In some implementations, when deploying an automated parking model on a vehicle, the following aspects can be considered:
[0099] First, deployment method selection: embedded devices, cloud servers, or edge devices.
[0100] Because embedded devices typically possess sufficient computing power and can be integrated with hardware and software systems through models, models can be deployed in dedicated embedded devices within vehicles, such as calculators or high-performance embedded processors. To leverage the massive computing power and storage resources of the cloud and facilitate model updates and management, models can be deployed on cloud servers, communicating with the vehicle via a network connection. Specifically, the vehicle can upload sensor data to the cloud server, the model performs inference on the cloud server, and then returns control commands to the vehicle. To achieve inference and control locally within the vehicle, reduce communication latency with cloud servers, and ensure better privacy and security, models can be deployed in edge devices within the vehicle, such as onboard computers or edge servers.
[0101] Second, hardware integration and software development.
[0102] In order to communicate with the vehicle control system, driver development programs are written using appropriate protocols (such as CAN bus), and software is developed to process sensor data streams; necessary sensor devices are linked, and pre-trained conditional generative adversarial networks are loaded and run using deep learning frameworks (Tensorflow or PyTorch).
[0103] Third, model optimization.
[0104] It should be noted that the automatic parking model provided in this application embodiment is a continuously iterative model. During the operation of the vehicle, the required data can be continuously collected and returned to the automatic parking model or remote service personnel in various ways to facilitate subsequent optimization and upgrades (such as quantizing and pruning the model to adapt to vehicle resource constraints, or converting the model to TensorRT format for acceleration).
[0105] Fourth, real-time performance and predictability.
[0106] It should be noted that sensor data can be input into the model for inference to generate corresponding control commands, or information such as the destination can be used to provide commands or auxiliary suggestions in advance in a pre-trained automatic parking model.
[0107] Fifth, security and reliability are guaranteed.
[0108] It should be noted that extensive field testing can be conducted after deployment, including testing under various road and environmental conditions, and handling of abnormal situations should be considered.
[0109] S104: In response to an operation that determines the target parking result, control the vehicle to park.
[0110] It should be noted that the target parking result obtained in this embodiment is not mandatory. Instead, it can be provided to the driver via voice or screen display in an opinion-based manner, based on information. In response to the driver confirming the target parking result, the vehicle is controlled to park. Thus, the model pre-judges, assists in judgment, or mandates parking based on the parking situation to meet the optimal judgment requirements of various complex road conditions and unexpected situations.
[0111] Thus, the automatic parking model, pre-trained based on a conditional generative adversarial network, can predict the safety of multiple parking scenarios in advance, improving the fault tolerance of automatic parking. In addition, the automatic parking model can continuously collect vehicle information to optimize and adapt to new parking scenarios, improving the efficiency of automatic parking and bringing a good user experience.
[0112] Please see Figure 2 , Figure 2 A flowchart illustrating another automatic parking method provided in an embodiment of this application may include, for example, the following steps:
[0113] S201: Pre-train the conditional generative adversarial network based on the public parking dataset to obtain a pre-trained parking model.
[0114] It should be noted that because the public parking dataset contains diverse sample categories and has high data quality after multiple cleaning and labeling processes, it is advisable to pre-train the conditional generative adversarial network using the public parking dataset first. This can improve the generalization ability of the generative model and reduce data loss caused by subsequent model training.
[0115] In some implementations, public parking datasets may include Waymo Open Dataset, ApolloScape, KITTI Vision Benchmark Suite, or nuScenes. Waymo Open Dataset is a large-scale autonomous driving dataset released by the autonomous driving company Waymo, containing high-resolution images, point clouds, IMU data, and vehicle status and annotation information from multiple sensors. ApolloScape is a dataset released by Baidu's Apollo autonomous driving platform, containing multi-sensor data from multiple cameras and LiDAR, scene annotations, and 3D object detection and tracking results. KITTI Vision Benchmark Suite is an autonomous driving dataset that includes data from onboard cameras, Velodyne LiDAR, and GPS / IMU, providing images, LiDAR point cloud data, and 2D / 3D object detection and tracking annotations for multiple scenes. nuScenes is an autonomous driving dataset released by nuTonomy (now Aptiv), containing high-resolution video, point clouds, and vehicle status data from multiple sensors, as well as annotated 2D / 3D object, scene, and behavior information.
[0116] S202: Adjust the parameters of the pre-trained parking model based on the training dataset to obtain an automatic parking model.
[0117] It should be noted that the training dataset consists of vehicle information collected in multiple parking scenarios beforehand. Since the pre-trained parking model is generated based on the pre-trained deep learning model of conditional generative adversarial network, the process of obtaining the automatic parking model is also regarded as an adjustment of the conditional generative adversarial network deep learning model.
[0118] It should be noted that, because automatic parking involves environmental conditions, vehicle position, and other features, the model in this embodiment of the application uses a conditional generative adversarial network.
[0119] It's important to note that in Conditional Generative Adversarial Networks (GANs), both the generator and discriminator receive conditional inputs. By generating and evaluating parking results incorporating these conditional information, the generator can produce more accurate and realistic parking outcomes based on the given conditions, while the discriminator can better distinguish between real and generated parking samples. Furthermore, GANs can handle multimodal data; regardless of the data type input, a one-to-one correspondence between the generator and discriminator can be established between the data and the conditions.
[0120] In some implementations, the conditional generative adversarial network includes a generator and a discriminator, and the training dataset includes conditional information and real samples. The embodiment of this application provides a method for adjusting the parameters of a pre-trained parking model based on the training dataset to obtain an automatic parking model, which may include:
[0121] Combination Figure 3 The schematic diagram of the automatic parking model shown can include speed and time data as conditional information, sensor data, location data and map data collected by the automatic parking system as real samples, and parking results as parking trajectory, control commands or comprehensive environmental perception results.
[0122] In some implementations, to make the generated samples more suitable and increase their diversity, the generator input can also include latent variables. Latent variables are low-dimensional vector representations of the input data. They are typically obtained by mapping high-dimensional input data to a low-dimensional space using an encoder. This low-dimensional vector contains the main feature information of the input data and can be used to generate new samples or perform other tasks.
[0123] S301: Input the condition information into the generator to obtain the generated sample.
[0124] S302: Input the condition information and the sample to be input into the discriminator to obtain the parking result.
[0125] The parking result is used to characterize the probability that the input sample is a real sample and to characterize the matching degree between the input sample and the condition information. The input sample includes real samples and generated samples.
[0126] In some implementations, a discriminator can be trained to perform binary or multi-ary classification based on labels controlled by humans (known parking results, human intervention, etc.).
[0127] S303: Calculate the loss function based on the parking results.
[0128] S304: The generator parameters are updated and iterated through backpropagation algorithm and loss function to obtain the automatic parking model.
[0129] In model training, the generator and discriminator interact strongly. The generator takes speed and time as input, and after several iterations, the generated samples begin to approximate the discriminator infinitely, making it impossible for the discriminator to accurately distinguish between real and generated samples. It's important to note that the generated samples are not discarded directly; instead, they are used together to train the discriminator and perform adversarial learning to generate the most accurate and realistic model. Furthermore, during model iteration, data can be collected and used for training in real time, and the trained model also has predictive performance.
[0130] It should be noted that adjustments and optimizations can be made during model training, therefore the method provided in this application embodiment may further include:
[0131] S401: Input the condition information into the generator to obtain the generated sample.
[0132] S402: Input the condition information and the sample to be input into the discriminator to obtain the parking result.
[0133] The parking result is used to characterize the probability that the input sample is a real sample and to characterize the matching degree between the input sample and the condition information. The input sample includes real samples and generated samples.
[0134] S403: Calculate the loss function based on the parking results.
[0135] S404: The generator parameters are updated and iterated through backpropagation algorithm and loss function to obtain an automatic parking model.
[0136] S405: Optimize the automatic parking model based on the preset optimization method to obtain an optimized parking model.
[0137] It should be noted that the preset optimization methods include hyperparameter tuning, optimizer selection, or regularization. Among them, hyperparameter tuning improves the performance and stability of the model by adjusting its hyperparameters, the optimizer is used to optimize the algorithm, and regularization is used to discard vectors that deviate from the discriminator's latent space.
[0138] As an example, we can optimize the model using regularization based on KL divergence. The function can be Regularized Objective = -log p(D|θ) + λ*KL(p(θ)||p0(θ)), where log p(D|θ) is the objective function given the data (usually maximum likelihood estimation); KL(p(θ)||p0(θ)) is the KL divergence between the model distribution p(θ) and the prior distribution p0(θ); and λ is the regularization parameter used to control the weight of the regularization term. By minimizing the above regularization objective function, we can simultaneously maximize the likelihood of the data and reduce the difference between the model and the prior distribution, thereby achieving a balance and robustness in model optimization or parameter estimation. As another example, the cosine distance loss function is used to achieve hyperparameter tuning. Specifically, the cosine distance loss function is Cosine Similarity=(A·B) / (||A||*||B||), where A·B represents the inner product (dot product) of vectors A and B, and ||A|| and ||B|| represent the norms (lengths) of vectors A and B, respectively.
[0139] S406: Call the test dataset to test the optimized parking model and obtain the test results.
[0140] S407: If the test results do not meet the preset requirements, continue to execute S405.
[0141] It should be noted that after the model has been trained to a certain level, it needs to be initially tested using a test set. Then, based on the test results, the model can be further optimized in a personalized way. In this way, through continuous iteration and optimization, the performance and reliability of the model can be further improved.
[0142] It should be noted that automatic parking faces a variety of environments, and traditional automatic parking systems struggle to simultaneously achieve prediction, real-time operation, and auxiliary judgment. Therefore, simulation testing is needed to further refine the model and improve its generalization ability. Thus, after S501 to S504, the method provided in this application embodiment may further include:
[0143] S501: Input the condition information into the generator to obtain the generated sample.
[0144] S502: Input the condition information and the sample to be input into the discriminator to obtain the parking result.
[0145] The parking result is used to characterize the probability that the input sample is a real sample and to characterize the matching degree between the input sample and the condition information. The input sample includes real samples and generated samples.
[0146] S503: Calculate the loss function based on the parking results.
[0147] S504: The generator parameters are updated and iterated through backpropagation algorithm and loss function to obtain the automatic parking model.
[0148] S505: Set up simulation test scenarios according to test requirements.
[0149] Among them, the simulation test scenario is used to simulate the parking environment in which the vehicle is located. One or more simulation test scenarios can be defined according to different test needs and purposes. Different simulation test scenarios can include different parking spaces, road conditions (such as narrow roads, slopes, etc.), different obstacles (such as other vehicles, pedestrians, etc.) and various other environmental parameters.
[0150] In some implementations, suitable simulation software or virtual environments can be selected for testing to provide the functions of creating, editing, and controlling scenarios, as well as simulating the physical characteristics of vehicles and the environment.
[0151] S506: Create a simulation test dataset corresponding to the simulation test scenario and obtain the simulation target parking results.
[0152] It should be noted that, in order to ensure that the test scenarios cover the expected usage as much as possible and to conduct simulation tests, a simulation test dataset can be created. This dataset includes input condition data that matches the simulation test scenarios, such as sensor data (e.g., camera images, LiDAR data) and other environmental information (e.g., map data, vehicle location, etc.). This data will be used as input to the automatic parking model to test the automatic parking model and obtain results.
[0153] S507: Input the simulation test dataset into the automatic parking model to obtain the initial parking simulation results.
[0154] S508: Evaluate the initial parking result based on the target parking result in the simulation and obtain the evaluation result.
[0155] Specifically, the initial parking results from the simulation are compared with the target parking results to evaluate the model's performance and quality. The model's performance in simulation tests is analyzed, and the generated parking results are quantitatively and qualitatively evaluated. Evaluation metrics may include parking accuracy, speed, or smoothness. Furthermore, it is important to record and analyze the model's behavioral differences under different scenarios and conditions.
[0156] S509: If the evaluation results do not meet the preset requirements, the automatic parking model will be optimized based on the preset optimization method to obtain an optimized parking model.
[0157] It should be noted that the automatic parking model can be tuned and optimized based on the evaluation results and preset optimization methods. This allows for adjustments to the model's architecture, hyperparameters, and optimization functions, and multiple iterations can be performed until a satisfactory automatic parking model is obtained.
[0158] In some implementations, an initial parking dataset can be collected first, and then the initial parking dataset can be processed to obtain a target parking dataset that can be used as input conditions in a generative adversarial network. Finally, the target parking dataset can be divided to obtain the training dataset.
[0159] Therefore, prior to S201 to S202, the method provided in the embodiments of this application may include:
[0160] S601: Obtain the initial parking dataset for multiple parking scenarios;
[0161] It should be noted that the initial parking dataset may include vehicle sensor data, location data, speed data, angle data, map data, or timing data. During the data collection process, the data can be initially screened according to preset collection rules. This is to ensure the accuracy and consistency of the data so that the sensors can be correctly calibrated, the data can be synchronized in time and aligned in position, and to ensure the real-time performance of the data so that the algorithm and model can be effectively applied in actual parking operations.
[0162] The sensor data can include data collected by various types of sensors. Specifically, ultrasonic sensors can detect the distance and position of obstacles around the vehicle; reversing cameras can acquire rear-view visual feedback images; radar sensors can acquire obstacle detection feedback, distance, and speed; and lidar sensors can acquire high-precision obstacle recognition information and environmental perception feedback. Location data can include data acquired by GPS and IMU. Specifically, GPS can acquire the vehicle's current and target positions for path planning and navigation; and IMU can acquire accelerometer and gyroscope data to measure the vehicle's acceleration, angular velocity, and attitude before and during parking. Vehicle speed data can include measurements of the vehicle's speed, acceleration, and deceleration. Angle data can include steering wheel angle data to determine the vehicle's steering intention before parking. Map data can include high-precision map data (e.g., road layout, parking information, or road restrictions) acquired in real-time by the in-vehicle navigation system. Timing data can include recording the time points of data acquisition and processing for time-related operations and synchronization.
[0163] S602: Perform data cleaning and preprocessing on the initial parking dataset to obtain an intermediate parking dataset.
[0164] It should be noted that data cleaning can include removing image noise, handling missing values, outliers, and isolated points; setting thresholds to check whether sensor measurements are within a reasonable range, and eliminating outliers introduced by sensor malfunctions or data transmission errors. Data preprocessing includes data normalization, standardization, and feature scaling. Data preprocessing ensures data consistency and comparability. Specifically, data normalization can be max-min normalization, used to scale feature data to a specified range, typically [0, 1], to eliminate differences caused by different units between data features. The formula is: x_scaled = (xx min ) / (x max -x min Then, Z-score standardization transforms the data into a distribution with a mean of 0 and a standard deviation of 1, reflecting the degree of fluctuation of the data relative to the mean. Here, x represents the original data, μ is the mean of the dataset, σ is the standard deviation of the dataset, and x_scaled represents the standardized data. min x is the minimum value in the dataset. max This represents the maximum value in the dataset.
[0165] Furthermore, since vehicles involve multiple sensors, data can be correlated and fused to obtain overall vehicle perception. This can be achieved by correlating and fusing cleaned and preprocessed data. For example, an extended Kalman filter can be used to fuse data from an inertial measurement unit (IMU), GPS, and lidar to provide a more accurate vehicle state estimate. The two key aspects of the Kalman filter are prediction and updating. ① The state prediction formula in the prediction step is: x_k = f(x_{k-1}, u_k), which predicts the vehicle's state (x_k) at the current time k. The current state estimate is calculated by combining the previous state (x_{k-1}) and the input (u_k) using a nonlinear state transition function f(). ② The state update formula in the update step (measurement update) is: x_k = x_k + K_k y_k. This formula is used to update the state (x_k) at the current time k so that it is closer to the actual measurement. K_k is the optimal gain and y_k is the measurement residual (the difference between the actual measurement value and the measurement prediction value). In this way, by correcting the state estimate, the state estimate and the actual measurement achieve the best fit.
[0166] S603: Based on feature engineering techniques, perform feature processing on the intermediate parking dataset to obtain the target parking dataset.
[0167] It should be noted that the embodiments of this application can employ different feature engineering techniques for feature processing based on different data types in the intermediate parking dataset. For time-type data, time series analysis and spectral analysis are used to calculate the mean, variance, standard deviation, and maximum and minimum values of various speeds to provide information on the vehicle's average motion state, degree of change, and dynamic range. Alternatively, for time-type data, Fourier transform can be used to convert the speed to the frequency domain, and relevant features, such as the main frequency distribution, can be extracted based on the spectral distribution. These features can reflect the periodic motion of the vehicle. For sensor data, geometric features are extracted to detect and classify different types of obstacles, providing information for path planning and obstacle avoidance decisions. For image-type data, end-to-end features can be extracted through texture analysis and convolutional neural networks, and co-occurrence matrices and grayscale histograms can be applied to target detection and tracking.
[0168] In some implementations, to increase data volume and remove data redundancy, the data obtained after feature engineering can be further subjected to dimensionality reduction. Therefore, the feature engineering-based feature processing of the intermediate parking dataset to obtain the target parking dataset provided in this application embodiment can include: performing feature processing on the intermediate parking dataset using feature engineering to obtain a feature parking dataset; and performing dimensionality reduction on the feature parking dataset to obtain the target parking dataset. Specifically, principal component analysis (PCA) can be used to reduce data dimensionality.
[0169] In addition, the dataset can be rotated, flipped, and scaled to enhance data diversity and help the model improve its generalization ability.
[0170] S604: Divide the target parking dataset into training dataset, validation dataset, and test dataset based on a preset ratio.
[0171] In some implementations, to ensure the performance and reliability of the model, the target parking dataset can be divided into a preset ratio of 7:1:2 to obtain training, validation, and test sets, which are then used for different stages of model training. Specifically, 70% of the data is used as the training set to train the model; 10% of the data is used as the validation set to verify the model's performance and to select and adjust the model during training; and 20% of the data is used as the test set to evaluate the model's generalization ability.
[0172] See Figure 4 This application also provides an automatic parking device 400, the device 400 comprising:
[0173] The acquisition unit 401 is used to acquire initial vehicle information based on the vehicle's data acquisition module, wherein the initial vehicle information is used to characterize the initial state of the vehicle.
[0174] Preprocessing unit 402 is used to preprocess the initial vehicle information to obtain target vehicle information;
[0175] The obtaining unit 403 is used to input the target vehicle information into the automatic parking model to obtain the target parking result. The vehicle includes the automatic parking model, which is generated based on training of a conditional generative adversarial network.
[0176] Control unit 404 is configured to control the vehicle to park in response to an operation that determines the target parking result.
[0177] Optionally, the automatic parking device 400 further includes: a model training unit, the model training unit comprising:
[0178] The pre-training module is used to pre-train the conditional generative adversarial network based on a public parking dataset to obtain a pre-trained parking model.
[0179] The training module is used to adjust the parameters of the pre-trained parking model based on the training dataset to obtain the automatic parking model.
[0180] Optionally, the conditional generative adversarial network includes a generator and a discriminator, the training dataset includes conditional information and real samples, and the training module is specifically used for:
[0181] The condition information is input into the generator to obtain the generated sample;
[0182] The condition information and the sample to be input are input into the discriminator to obtain the parking result. The parking result is used to characterize the probability value that the sample to be input is the real sample and to characterize the matching degree between the sample to be input and the condition information. The sample to be input includes the real sample and the generated sample.
[0183] Calculate the loss function based on the parking results;
[0184] The automatic parking model is obtained by updating and iterating the parameters of the generator using the backpropagation algorithm and the loss function.
[0185] Optionally, the training module is further configured to:
[0186] The automatic parking model is optimized based on a preset optimization method to obtain an optimized parking model. The preset optimization method includes hyperparameter tuning, optimizer selection, or regularization.
[0187] Optionally, the training module is further configured to:
[0188] The optimized parking model is tested using the test dataset to obtain test results;
[0189] If the test results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
[0190] Optionally, the training module is further configured to:
[0191] A simulation test scenario is set up according to the test requirements. The simulation test scenario is used to simulate the parking environment in which the vehicle is located.
[0192] Create a simulation test dataset corresponding to the simulation test scenario and obtain the simulation target parking results;
[0193] Input the simulation test dataset into the automatic parking model to obtain the initial parking simulation results;
[0194] The initial parking result is evaluated based on the target parking result to obtain the evaluation result;
[0195] If the evaluation results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
[0196] Optionally, the automatic parking device 400 further includes a data acquisition unit, which is specifically used for:
[0197] Obtain initial parking datasets for multiple parking scenarios;
[0198] The initial parking dataset is cleaned and preprocessed to obtain an intermediate parking dataset;
[0199] The intermediate parking dataset is processed using feature engineering techniques to obtain the target parking dataset.
[0200] The target parking dataset is divided according to a preset ratio to obtain a training dataset, a validation dataset, and a test dataset.
[0201] Optionally, the step of performing feature processing on the intermediate parking dataset based on feature engineering techniques to obtain the target parking dataset includes:
[0202] The intermediate parking dataset is processed using feature engineering techniques to obtain a feature parking dataset.
[0203] The dimensionality of the feature parking dataset is reduced to obtain the target parking dataset.
[0204] It should be noted that the specific implementation method and the technical effects achieved by the device 400 can be found in [reference needed]. Figure 2 or Figure 3The relevant descriptions in the methods shown.
[0205] Furthermore, this application also provides a vehicle including an automatic parking device, which may be, for example, the one described above. Figure 4 The automatic parking device 400 is described above.
[0206] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as a read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.
[0207] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The device embodiments described above are merely illustrative. Modules described as separate components may or may not be physically separate. 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 units. Some or all of the modules can be selected to achieve the objectives of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0208] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. An automatic parking method, characterized in that, The method includes: The vehicle-based data acquisition module acquires initial vehicle information, which is used to characterize the initial state of the vehicle. The initial vehicle information is preprocessed to obtain the target vehicle information; The target vehicle information is input into the automatic parking model to obtain the target parking result. The vehicle includes the automatic parking model, which is generated based on training with a conditional generative adversarial network. In response to the operation of determining the target parking result, control the vehicle to park; The method further includes: The conditional generative adversarial network is pre-trained based on a public parking dataset to obtain a pre-trained parking model. The parameters of the pre-trained parking model are adjusted based on the training dataset to obtain the automatic parking model; The conditional generative adversarial network includes a generator and a discriminator; the training dataset includes conditional information and real samples; and the step of adjusting the parameters of the pre-trained parking model based on the training dataset to obtain the automatic parking model includes: The condition information is input into the generator to obtain the generated sample; The condition information and the sample to be input are input into the discriminator to obtain the parking result. The parking result is used to characterize the probability value that the sample to be input is the real sample and to characterize the matching degree between the sample to be input and the condition information. The sample to be input includes the real sample and the generated sample. Calculate the loss function based on the parking results; The automatic parking model is obtained by updating and iterating the parameters of the generator using the backpropagation algorithm and the loss function.
2. The automatic parking method according to claim 1, characterized in that, The method further includes: The automatic parking model is optimized based on a preset optimization method to obtain an optimized parking model. The preset optimization method includes hyperparameter tuning, optimizer selection, or regularization.
3. The automatic parking method according to claim 2, characterized in that, The method further includes: The optimized parking model is tested using the test dataset to obtain test results; If the test results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
4. The automatic parking method according to claim 1, characterized in that, The method further includes: A simulation test scenario is set up according to the test requirements. The simulation test scenario is used to simulate the parking environment in which the vehicle is located. Create a simulation test dataset corresponding to the simulation test scenario and obtain the simulation target parking results; Input the simulation test dataset into the automatic parking model to obtain the initial parking simulation results; The initial parking result is evaluated based on the target parking result to obtain the evaluation result; If the evaluation results do not meet the preset requirements, the automatic parking model is optimized based on the preset optimization method to obtain an optimized parking model.
5. The automatic parking method according to claim 1, characterized in that, The method further includes: Obtain initial parking datasets for multiple parking scenarios; The initial parking dataset is cleaned and preprocessed to obtain an intermediate parking dataset; The intermediate parking dataset is processed using feature engineering techniques to obtain the target parking dataset. The target parking dataset is divided according to a preset ratio to obtain a training dataset, a validation dataset, and a test dataset.
6. The automatic parking method according to claim 5, characterized in that, The feature engineering technique is used to perform feature processing on the intermediate parking dataset to obtain the target parking dataset, including: The intermediate parking dataset is processed using feature engineering techniques to obtain a feature parking dataset. The dimensionality of the feature parking dataset is reduced to obtain the target parking dataset.
7. An automatic parking device, characterized in that, The apparatus for performing the automatic parking method of claim 1, comprising: An acquisition unit is used to acquire initial vehicle information based on the vehicle's data acquisition module, wherein the initial vehicle information is used to characterize the initial state of the vehicle. The preprocessing unit is used to preprocess the initial vehicle information to obtain the target vehicle information; The obtaining unit is used to input the target vehicle information into the automatic parking model and obtain the target parking result. The vehicle includes the automatic parking model, which is generated based on training of a conditional generative adversarial network. A control unit is configured to control the vehicle to park in response to an operation that determines the target parking result.
8. A vehicle, characterized in that, Includes the automatic parking device as described in claim 7.