Training data generation method, computer equipment and program product
By classifying and aggregating data packets by date, sequentially stored information data packets are transformed into high-quality deep learning model training datasets. This solves the problems of inconsistent data formats and high computational resource requirements, achieving efficient data conversion and data feature continuity, and generating reliable model training data.
Patent Information
- Application Number
- CN202411533925.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-30
- Publication Date
- 2024-11-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
How to efficiently convert sequentially stored information data packets into high-quality deep learning model training datasets, solving problems such as inconsistent data formats, noise handling, and high computational resource requirements.
By classifying data packets by date and processing them based on date, and using data from the user's own vehicle or other vehicles as the aggregation point, the original data that has been read is aggregated to generate aggregated data features for the user's own vehicle or other vehicles, and finally, model training data is generated.
It improves data conversion efficiency, ensures the continuity of data features, generates a reliable model training dataset, and supports efficient training of deep learning models.
Smart Images

Figure CN119046692A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of computer technology, and in particular to a training data generation method, a computer device, and a program product. Background Art
[0002] For the autonomous driving system from the vehicle's perception end to the decision-making end, it provides a highly integrated autonomous driving solution by integrating multiple sensors, advanced computer vision, deep learning models and control technologies. As an important module in the vehicle's autonomous driving system, the prediction accuracy of the deep learning model is related to the safety of autonomous driving. The prediction accuracy of the deep learning model can be improved through model training, and the training of the deep learning model requires a large number of data sets. The scale and quality of the training data in the data set affect the quality of the trained deep learning model.
[0003] A data packet that stores information sequentially refers to a collection of data recorded in time. This type of data packet maintains its original sequential characteristics when collected, stored, and processed. If the data in the data packet that stores information sequentially can be converted into the data in the training data set, it will be of great help in improving the quality of the data set. However, making a deep learning training data set is a complex and multi-step task involving multiple links such as data collection, preprocessing, label generation, and data set format conversion. Therefore, how to convert a data packet that stores information sequentially into a training data set has become an issue worthy of attention. Summary of the invention
[0004] The embodiments of the present disclosure at least provide a training data generation method, a computer device, and a program product.
[0005] In a first aspect, an embodiment of the present disclosure provides a method for generating training data, comprising: Acquire a data set to be converted, the data set comprising a plurality of data packets classified by date; the data packet comprising a plurality of sub-data packets, the sub-data packets comprising a plurality of original data stored in time sequence; For any of the data packets, in the process of reading data according to the storage order of the pieces of original data included in each of the sub-data packets in the data packet, whenever data of the own vehicle or another vehicle is read, the read and unaggregated original data of various types are aggregated according to the target data structure to obtain aggregated data of the own vehicle or another vehicle; Generating data features corresponding to the own vehicle or other vehicles according to the aggregated data of the own vehicle or other vehicles; Generate model training data based on the data features corresponding to the vehicle or other vehicles described in each item.
[0006] In a second aspect, an optional implementation of the present disclosure further provides a computer device, a processor, and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the processor is used to execute the machine-readable instructions stored in the memory, and when the machine-readable instructions are executed by the processor, the steps of the above-mentioned first aspect, or any possible implementation of the first aspect are performed.
[0007] In a third aspect, an optional implementation of the present disclosure further provides a computer program product, which, when executed, executes the steps of the above-mentioned first aspect, or any possible implementation of the first aspect.
[0008] For a description of the effects of the above-mentioned computer device and computer program product, please refer to the description of the above-mentioned training data generation method, which will not be repeated here.
[0009] The training data generation method, computer device and program product provided by the embodiments of the present disclosure use data packets classified by date to generate training data, which can realize parallel processing of data packets under different dates with date as a basic processing unit, which can not only improve the data conversion efficiency, but also ensure the continuity of the data features extracted from the data packets under each date. Taking each time to read the data of the vehicle or other vehicles as the data aggregation opportunity, the various types of raw data that have been read and not aggregated are aggregated, which can realize the aggregation of various data related to the vehicle at the same time or adjacent time, and then the aggregated data can be used to accurately extract data features suitable for deep learning model training. Finally, based on the data features corresponding to the aggregated data obtained by each aggregation, the historical data features as the model input and the future data features as the label can be obtained, so as to obtain model training data with reliable quality. In the whole process, by classifying the sequentially stored data packets by date and generating model training data with date as the basic processing unit, a model training data set with reliable quality can be generated while taking into account the continuity of data time series and the continuity of data features.
[0010] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.
[0012] Figure 1 A flow chart of a training data generation method provided by an embodiment of the present disclosure is shown; Figure 2 A schematic diagram of a data packet classified by date provided by an embodiment of the present disclosure is shown; Figure 3 A schematic diagram of a data structure of a sub-data packet provided by an embodiment of the present disclosure is shown; Figure 4 A flow chart of data aggregation and labeling provided by an embodiment of the present disclosure is shown; Figure 5 A schematic diagram showing a structure of a data feature provided by an embodiment of the present disclosure is shown; Figure 6 A schematic diagram showing features of various pieces of data with the same data identifier provided by an embodiment of the present disclosure is shown; Figure 7 A schematic diagram of feature storage in a target array provided by an embodiment of the present disclosure is shown; Figure 8 A schematic diagram of a training data set file in HDF5 format provided by an embodiment of the present disclosure is shown; Fig. 9 A schematic diagram of a sorted feature sequence provided by an embodiment of the present disclosure is shown; Fig.10 A schematic diagram showing an illegal situation provided by an embodiment of the present disclosure; Fig.11 A flow chart of forward search and backward search provided by an embodiment of the present disclosure is shown; Fig.12 A flow chart of generating a windowed feature sequence provided by an embodiment of the present disclosure is shown; Fig.13 A schematic diagram of the structure of a computer device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION
[0013] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure is not intended to limit the scope of the present disclosure claimed for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.
[0014] In addition, the terms "first", "second", etc. in the specification and claims of the embodiments of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. The "plurality or several" mentioned in this article refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.
[0015] Research has found that the training of deep learning models depends on a large number of data sets, because the model contains a large number of parameters. Only by training on a large-scale data set can complex patterns and features be effectively captured. The scale and quality of the data set directly affect the performance of the model. Therefore, in the process of making the data set, many difficulties need to be overcome and the diversity and representativeness of the data need to be ensured. Specifically, these difficulties include the inconsistency of the data format and structure of data from different sources, which need to be uniformly processed and aligned to extract the features required for model training from the data. There may be missing values and noise in the sequentially stored data, which must be effectively processed, otherwise it will affect the performance of the model. Sequentially stored data usually has a large amount of data, and processing and storing this data requires a lot of computing resources and storage space. Therefore, how to convert the data packets of sequentially stored information into high-quality training data sets on the basis of overcoming the above difficulties, and lay the foundation for efficient training and performance improvement of deep learning models, has become a technical issue worth thinking about.
[0016] Based on the above research, the present disclosure provides a training data generation method, a computer device and a program product. The training data is generated by using data packets classified by date, which can realize parallel processing of data packets under different dates with date as a basic processing unit, which can not only improve the data conversion efficiency, but also ensure the continuity of the data features extracted from the data packets under each date. Taking each time to read the data of the vehicle or other vehicles as the data aggregation opportunity, the various types of raw data that have been read and not aggregated are aggregated, which can realize the aggregation of various data related to the vehicle at the same time or adjacent time, and then the aggregated data can be used to accurately extract data features suitable for deep learning model training. Finally, based on the data features corresponding to the aggregated data obtained by each aggregation, the historical data features as the model input and the future data features as the label can be obtained, so as to obtain model training data with reliable quality. In the whole process, by classifying the sequentially stored data packets by date and generating model training data with date as the basic processing unit, a model training data set with reliable quality can be generated while taking into account the continuity of data time series and the continuity of data features.
[0017] The defects existing in the above solutions are the results obtained by the inventor after practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the present disclosure for the above problems below should be the contributions made by the inventor to the present disclosure during the disclosure process.
[0018] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0019] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.
[0020] It should be noted that the specific terms mentioned in the embodiments of the present disclosure include: Data packets storing information sequentially: Data packets storing information sequentially refer to data sets recorded in chronological order. Such data packets maintain their original sequential characteristics when collected, stored, and processed. They are usually used to represent continuous data such as time series data, log files, and sensor readings. Its characteristics include time series, that is, data is arranged in chronological order, and each piece of data has a timestamp; continuity, that is, the data in the data packet is recorded continuously, reflecting the dynamic changes of a process or phenomenon; dependency, that is, there is usually an inherent dependency between the pieces of data stored sequentially, and the previous and subsequent data may be related. Common examples include sensor data, such as temperature sensors that record temperature values once per second, generating time series data; financial data, such as financial market data recorded in chronological order; log data, such as server logs, system logs, etc., log information recorded in the order of events; audio data, such as audio signals in recording files, sound data stored in chronological order. In the disclosed embodiment, the data packet storing information sequentially can store data collected by various sensors on the vehicle. Data packets storing information sequentially are widely used in scientific research, financial analysis, the Internet of Things, the medical field, computer systems, etc.
[0021] BEV raster image: an image obtained by rasterizing the BEV feature map predicted by a 3D Gaussian set.
[0022] NumPy: The full name is Numerical Python, which is a multidimensional array object that can be used to store and process large matrices.
[0023] HDF5, the full name of Hierarchical Data Format version 5, is a file format used to store and organize large amounts of data. It supports the ability to efficiently store and process large-scale scientific data. HDF5 files usually use the .h5 extension and are a binary file format.
[0024] To facilitate understanding of this embodiment, a training data generation method disclosed in an embodiment of the present disclosure is first introduced in detail. The execution subject of the training data generation method provided in the embodiment of the present disclosure is generally a terminal device or other processing device with certain computing capabilities, wherein the terminal device may be a user equipment (User Equipment, UE), a mobile device, a user terminal, a terminal, a personal digital assistant device (Personal Digital Assistant, PDA), a handheld device, a computer device, etc.; in some possible implementations, the training data generation method may be implemented by a processor calling a computer-readable instruction stored in a memory.
[0025] The training data generation method provided by the embodiment of the present disclosure is described below by taking a computer device as an example of an execution subject.
[0026] like Figure 1 As shown, it is a flowchart of a training data generation method provided by an embodiment of the present disclosure, which may include S101-S104: S101: Acquire a data set to be converted, where the data set includes a plurality of data packets classified by date; the data packet includes a plurality of sub-data packets, and the sub-data packets include a plurality of original data stored in time sequence.
[0027] Here, the data set to be converted is a data set to be converted into a model training data set corresponding to the deep learning model, and the data set to be converted includes multiple data packets classified by date. Each date can correspond to a total data packet. In the data set, the data packets corresponding to different dates can be stored in chronological order or in disorder, which is not specifically limited in the embodiments of the present disclosure.
[0028] Each data packet under a date may include multiple sub-data packets, and the number of sub-data packets included in different data packets may be different. Figure 2 FIG. 1 is a schematic diagram of a data packet classified by date provided by an embodiment of the present disclosure. Figure 2 In the figure, the data packets corresponding to the three dates (i.e., date 1, date 2, and date 3) are shown respectively. The data packet corresponding to date 1 includes three sub-data packets, the data packet corresponding to date 2 includes two sub-data packets, and the data packet corresponding to date 3 includes four sub-data packets. The sub-data packets for each date are different.
[0029] Each sub-data packet is a data packet that stores information sequentially, including multiple pieces of raw data stored in time sequence, and the raw data can be data collected by any sensor on the vehicle. The sub-data packets included in the same data packet can be stored in the data packet in time sequence or in any order. The raw data can be data collected by using sensors on the vehicle according to the functions of the deep learning model. When training different functional models, the raw data is different. For example, in the case of training a deep learning model for vehicle trajectory prediction, the raw data can be collected vehicle driving data; in the case of training a deep learning model for medical data analysis, the raw data can be collected medical monitoring data. The following text takes the case where the raw data is vehicle driving data as an example to illustrate the training data generation method provided in the embodiment of the present disclosure. One piece of raw data corresponds to one timestamp. Since the accuracy of the timestamp often reaches microseconds or milliseconds, the timestamp of each piece of raw data is different.
[0030] A sub-data packet can be called a record file, which is often used to store the original data collected in sequence in an acquisition time period, and different sub-data packets can correspond to different acquisition time periods. For example, each sub-data packet is a 60s data packet, which includes the original data stored in time sequence, and the maximum time span corresponding to each original data is 60s.
[0031] In a specific implementation, different initial data can be obtained, and the initial data here can include sub-data packets and / or original data that sequentially store information. Then, according to the date corresponding to the initial data, the initial data can be classified by date, and the initial data can be classified and repackaged to obtain a data packet corresponding to each date, and each data packet includes multiple sub-data packets. In this way, the data set to be converted can be obtained. Among them, in the case where the initial data is raw data, when the initial data is packaged, the initial data can be first combined into sub-data packets according to the time period length corresponding to each sub-data packet in time sequence, and then the combined sub-data packets can be classified into a data packet corresponding to a certain date. In this way, by classifying the data packets according to the date, it can be ensured that there are sub-data packets of continuous time periods under each date.
[0032] S102: For any data packet, in the process of reading data according to the storage order of each original data included in each sub-data packet in the data packet, whenever the data of the own vehicle or other vehicles is read, the various types of original data that have been read and not aggregated are aggregated according to the target data structure to obtain the aggregated data of the own vehicle or other vehicles.
[0033] Here, for each piece of original data in the sub-data package, each piece of original data has a topic information, and the topic information can be understood as the data type of the original data. Figure 3 FIG. 1 is a schematic diagram of a data structure of a sub-data packet provided by an embodiment of the present disclosure. Figure 3 , which respectively correspond to seven pieces of raw data, are Traffic Light information, Perception information, Perception Obstacle information, Chassis information, Localization information, Human interaction information, and RoadTopo information. Each piece of raw data may include a variety of data related to the corresponding topic information. For example, for the raw data corresponding to the chassis information, the raw data may include a timestamp, throttle information, brake information, etc. For the raw data corresponding to the human interaction information, the raw data may include various interactive operations performed by the user.
[0034] according to Figure 3 It can be seen that in a sub-data packet, one topic information corresponds to one piece of original data, so the topic information is also stored in chronological order.
[0035] The data of the self-vehicle or other-vehicle is the data of the vehicle type in each piece of raw data included in the data packet. The vehicle type may include the self-vehicle type and the perceived obstacle type. The raw data of the self-vehicle type may specifically include the self-vehicle identification, vehicle type (such as sedan, SUV, commercial vehicle, etc.), timestamp and other information. The raw data of the perceived obstacle type may specifically include the obstacle identification, obstacle number, obstacle type (such as pedestrian, vehicle), obstacle size, timestamp and other information. A piece of self-vehicle data or a piece of other-vehicle data is a piece of vehicle type data read from a certain sub-data packet. That is, when reading the raw data from the sub-data packet, if the data of the self-vehicle type is read, the data is taken as the self-vehicle data, and if the data of the obstacle type is read, the data is taken as the other-vehicle data.
[0036] The target data structure is a data storage structure, which can be set based on experience. Exemplarily, the target data structure can be a frame structure.
[0037] The various types of raw data that have been read and not aggregated may be data such as Traffic Light, Perception, chassis, localization, Hminode, RoadTopo, etc. that are read before the current self-vehicle or other-vehicle data is read and after the last aggregation. Among them, the various types of raw data that have been read and not aggregated may include the current self-vehicle or other-vehicle data. The aggregated data of the self-vehicle or other-vehicle is the aggregated data corresponding to the self-vehicle or other-vehicle data.
[0038] In specific implementation, after obtaining each data packet, data aggregation processing can be performed on each data packet, and the features related to the self-vehicle or other-vehicle data are aggregated on the timestamp corresponding to the self-vehicle or other-vehicle data, and the model training data is generated based on the aggregated data obtained by aggregation. Among them, one data packet can be processed using one thread, and multiple data packets can be processed in parallel using multiple threads. When data aggregation, for any data packet, the original sequence of each sub-data packet can be read in sequence according to the storage order of each original data included in each sub-data packet in the data packet. In the process of sequentially reading data, whenever the corresponding self-vehicle or other-vehicle data is read, the frame structure can be used to aggregate the original data of each type that has been read and not aggregated with the current self-vehicle or other-vehicle data to obtain an aggregated data corresponding to the current self-vehicle or other-vehicle data. That is, in the process of sequentially reading data, whenever the original data of the vehicle type is read (whether it is the original data of the self-vehicle type or the original data of the perceived obstacle type), the original data corresponding to each topic information that has been read and not aggregated can be aggregated to obtain an aggregated data.
[0039] Here we explain why we aggregate raw data such as positioning information, map information, lane information, chassis information, etc. whenever we read the data of our own vehicle or other vehicles. Because for the deep learning model used to predict the vehicle trajectory, one is to take the own vehicle as the midpoint, and input the environmental information around the own vehicle (such as map information, lane information, perception information) and the driving information of the own vehicle (such as chassis information, positioning information, speed information, etc.) as features into the model, so as to realize the planning of the trajectory of the own vehicle; the other is to take the obstacle vehicle as the center, and input the environmental information around the obstacle vehicle and the driving information of the obstacle vehicle as features into the model, so as to realize the planning of the obstacle vehicle trajectory. Therefore, it can be seen that whether it is the prediction of the own vehicle or the prediction of the obstacle vehicle, the required training data is consistent, for example, the historical trajectory features of the vehicle in the past period of time (including features related to the historical trajectory, historical environmental information, historical vehicle status, etc.), the current vehicle status, the current environmental information and the real trajectory features of the future period of time as labels are required. Therefore, when generating training data, it is beneficial to use the raw data of the vehicle type read as the opportunity to trigger data aggregation.
[0040] In this way, when reading each piece of original data in the sub-data packet, data aggregation is performed every time the self-vehicle or other-vehicle data is read, and the aggregated data corresponding to each piece of self-vehicle or other-vehicle data in the sub-data packet can be obtained. For example, the sub-data packet includes 5 pieces of self-vehicle or other-vehicle data, and the 5 pieces of self-vehicle or other-vehicle data have different timestamps. After the data aggregation is completed, the aggregated data corresponding to the 5 pieces of self-vehicle or other-vehicle data can be obtained.
[0041] S103: Generate data features corresponding to the own vehicle or other vehicles based on the aggregated data of the own vehicle or other vehicles.
[0042] Here, data features are features extracted from aggregated data, which are related to the prediction function of the deep learning model. For example. In the case of training a deep learning model for vehicle trajectory prediction, the extracted features may include determining whether the obstacle is a pedestrian or a vehicle based on the perception information, determining whether it is currently possible to move forward based on the traffic light information, and so on. At the same time, when using the self-vehicle data aggregation, some trajectory samplers can also be used to generate some future executable trajectories for the self-vehicle based on the aggregated data corresponding to the self-vehicle data, as data features for subsequent model training, and the surrounding environment perceived by the self-vehicle can be drawn onto a picture (such as a bev raster image), and this can also be used as one of the data features for subsequent deep learning training, and so on.
[0043] The timestamp of the aggregated data corresponding to the own vehicle or the other vehicle may be the timestamp of the data of the own vehicle or the other vehicle, and the identifier of the aggregated data may be the data identifier of the data of the own vehicle or the other vehicle. For example, the data identifier of the aggregated data corresponding to the own vehicle may be the number of the own vehicle, and the data identifier of the aggregated data corresponding to the other vehicle may be the number of the other vehicle.
[0044] The data features corresponding to the own vehicle or other vehicles are the data features corresponding to the own vehicle or other vehicle data. One piece of own vehicle or other vehicle data corresponds to one piece of aggregated data, one piece of aggregated data corresponds to one data feature, so one piece of own vehicle or other vehicle data also corresponds to one data feature. The data feature has the same timestamp and data identifier as the corresponding aggregated data. For example, when the timestamp of own vehicle data 1 is timestamp 1 and the data identifier is number 1, the timestamp of aggregated data 1 corresponding to own vehicle data 1 is timestamp 1 and the data identifier is number 1, and the timestamp of the data feature corresponding to aggregated data 1 is timestamp 1 and the data identifier is number 1.
[0045] In the specific implementation, for any data packet, after obtaining the aggregated data corresponding to each piece of self-vehicle or other-vehicle data in the data packet, for each piece of aggregated data, feature extraction can be performed on the aggregated data according to the prediction task of the deep learning module, so as to obtain various features matching the prediction task, and these features are stored according to a specific data structure to obtain the data features corresponding to each piece of self-vehicle or other-vehicle data.
[0046] Among them, for each data feature, a combination of vehicle number (such as the vehicle number or obstacle vehicle number (i.e. other vehicle number)) and timestamp can be used for unique retrieval.
[0047] S104: Generate model training data according to the data features corresponding to each vehicle or other vehicle.
[0048] In specific implementation, for the data features corresponding to each piece of self-vehicle data, the data features corresponding to each self-vehicle number can be divided according to the self-vehicle number. Then, for each self-vehicle number, at least one model training data corresponding to the self-vehicle number can be generated according to each piece of data features with the self-vehicle number. For the data features corresponding to each piece of other-vehicle data, the data features corresponding to each other-vehicle number can be divided according to the other-vehicle number. Then, for each other-vehicle number, at least one model training data corresponding to the other-vehicle number can be generated according to each piece of data features with the other-vehicle number.
[0049] In this way, by classifying the sequentially stored data packets by date and using date as the basic processing unit to generate model training data, a model training data set with reliable quality can be generated while taking into account both the data temporal continuity and the data feature continuity.
[0050] In one embodiment, the above S104 can be implemented according to the following steps: S104-1: Use the data identifier and timestamp of the data of the own vehicle or the other vehicle as the data identifier and timestamp of the data feature corresponding to the own vehicle or the other vehicle.
[0051] Exemplarily, the data features corresponding to the data of the own vehicle or other vehicles may have a data identifier and a timestamp of the data of the own vehicle or other vehicles. Optionally, the step of determining the data identifier and timestamp of the data features may be performed when generating the data features corresponding to the own vehicle or other vehicles, or may be performed when generating model training data, which is not specifically limited in the embodiments of the present disclosure.
[0052] S104-2: Determine a feature sequence corresponding to the same data identifier according to the data identifier of the data feature and the data features corresponding to each own vehicle or other vehicle.
[0053] Here, when the data identifiers are the same, it can be explained that the data features belong to the features of the same vehicle.
[0054] In specific implementation, after obtaining the data identifiers corresponding to each piece of self-vehicle or other-vehicle data in the data packet, data features can be aggregated. Specifically, according to each piece of data feature and the data identifier of the data feature, each piece of data feature belonging to the same data identifier can be aggregated together to obtain a feature sequence corresponding to each data identifier.
[0055] Optionally, when aggregating data features, only the data features corresponding to each piece of own vehicle or other vehicle data in the same data packet may be aggregated, or the data features corresponding to each piece of own vehicle or other vehicle data in multiple data packets with adjacent dates or date intervals less than a preset interval may be aggregated.
[0056] S104-3: Generate at least one model training data corresponding to each data identifier according to the timestamp and target duration of each data feature in each feature sequence; the model training data includes multiple target data features that are consecutive in time sequence, and the maximum time difference between the multiple target data features is less than or equal to the target duration.
[0057] Here, the target data feature is the data feature included in the model training data, and a target data feature has a timestamp. A model training data includes multiple target data features that are continuous in time series. Here, the time series continuity means that the time interval between the previous target data feature and the next target data feature in every two adjacent target data features is less than the time interval with other target data features, and the timestamp of the next target data feature is later than the timestamp of the previous target data feature. The maximum time difference between multiple target data features in a model training data is the time difference between the two target data features with the earliest and latest timestamps in the model training data, and the duration is less than or equal to the target duration. The target duration can be set based on experience, and the embodiments of the present disclosure do not specifically limit it.
[0058] For example, for any data identifier, after obtaining the feature sequence corresponding to the data identifier, a sliding window with a target duration can be used to perform sliding screening on multiple data features of the feature sequence, and the multiple data features screened each time can be used as target data features. That is, each data feature located in the sliding window during each screening can be combined into a model training data.
[0059] Optionally, for any data identifier, after obtaining the feature sequence corresponding to the data identifier, for each data feature in the feature sequence, the data feature can be used as a benchmark to forward screen out multiple data features as target data features, and backward screen out multiple data features as target data features. The forward screened target data features and the benchmark data features are used as the input part of the model training data, and the backward screened target data features are used as the label part of the training data, thereby obtaining a model training data corresponding to the benchmark data feature. When forward screening and backward screening of data features, the sum of the durations in the two directions is equal to the target duration. For example, if the target duration is 4.5s, search forward for 1.5s of historical data features, and search backward for 3s of future data features, and combine these data features into a training data.
[0060] In this way, by using data identifiers and timestamps, we can obtain a temporally continuous feature sequence corresponding to the same data identifier. By then generating training data based on the feature sequence, we can obtain historical target data features as model inputs and future target data features as labels, thereby obtaining a training data set with reliable quality.
[0061] In one embodiment, the vehicle data or other vehicle data may be raw data of vehicle type. The step of aggregating the raw data to obtain aggregated data in S102 may be implemented as follows: S102-1: In the process of reading data according to the storage order of each piece of original data included in each sub-data packet in the data packet, if the currently read original data does not belong to the own vehicle or other vehicle data, the original data is stored in a storage queue of the corresponding type according to the type of the original data.
[0062] Here, the storage queue may specifically be a storage container set locally, and different data types correspond to different storage containers, that is, one topic information corresponds to one storage container. For any data packet, each data type corresponding to the data packet may have a storage queue.
[0063] Exemplarily, for any data packet, each sub-data packet can be read in sequence according to the storage order of each sub-data packet in the data packet. For the sub-data packet currently read, it is necessary to read each piece of original data in the sub-data packet. When reading, each piece of original data can be read in sequence according to the storage order of each piece of original data in the sub-data packet. For the original data currently read, according to the data type of the original data, it is determined whether the original data currently read is the data of the own vehicle or other vehicles, that is, it is determined whether the data type of the original data currently read is the vehicle type. If not, it can be determined whether there is a storage queue corresponding to the data type of the original data currently read; if so, the original data can be stored in the storage queue. If not, a storage queue corresponding to the data type can be created and the original data can be stored in it. If the data type of the original data currently read is the vehicle type, the following S102-2 can be executed.
[0064] S102-2: If the data of the own vehicle or other vehicles is currently read, the unextracted original data currently in the storage queue is extracted from each type of storage queue respectively.
[0065] Here, the raw data in the storage queue that has not been extracted belongs to the raw data that has been read but not aggregated.
[0066] For example, if the raw data currently being read is data of the own vehicle or other vehicles, the raw data that has not been extracted and is currently in the storage queue can be extracted from each type of storage queue. Here, for any raw data, after the raw data is extracted from the storage queue, the storage queue will no longer store the raw data.
[0067] Optionally, if the self-vehicle or other-vehicle data is currently read, the self-vehicle or other-vehicle data currently read may be first stored in the storage queue corresponding to the vehicle type. Then, the unextracted raw data currently in the storage queue is extracted from the storage queues corresponding to each data type. The extracted raw data includes the self-vehicle or other-vehicle data and the raw data of other data types.
[0068] S102 - 3 : According to the target data structure, aggregate the own vehicle or other vehicle data and each piece of original data extracted from the storage queue to obtain aggregated data of the own vehicle or other vehicle.
[0069] Exemplarily, the currently read self-vehicle or other-vehicle data and each piece of original data extracted from the storage queue can be aggregated according to the target data structure to obtain an aggregated data corresponding to the currently read self-vehicle or other-vehicle data. Further, the step of reading each piece of original data in the sub-data packet in sequence according to the storage order of each piece of original data in the sub-data packet can be returned until all pieces of original data in the sub-data packet are read, and the aggregated data corresponding to each piece of self-vehicle or other-vehicle data in the sub-data packet is obtained. Then, the step of reading the sub-data packets in sequence according to the order of each sub-data packet in the data packet can be returned until all pieces of sub-data packet in the data packet are read, and the aggregated data corresponding to each piece of self-vehicle or other-vehicle data in the data packet is obtained.
[0070] Here, since the original data is stored in the sub-data packet in time sequence and is also read in time sequence when read, the unextracted original data in each storage queue belongs to the data with a time close to or the same as the current read self-vehicle or other-vehicle data, and these data have not been aggregated. However, since during model prediction, by inputting the relevant data of the same vehicle at the same time or a similar time into the model, the model can fully explore the connection between the data and then make an accurate prediction, so when the original data of the vehicle type currently read, the unextracted original data extracted from the storage queues of each data type can be aggregated to obtain an aggregated data corresponding to the original data of the vehicle type, and the timestamp and data identifier of the original data of the vehicle type are used as the timestamp and data identifier of the aggregated data, so as to realize the aggregation of multiple data under the same timestamp. After the aggregated data of the original data of the vehicle type currently read is obtained by aggregation, the aggregated data can also be data labeled to obtain the data features corresponding to the original data of the vehicle type currently read. Among them, the step of data labeling the aggregated data is to execute the above-mentioned step S103. Alternatively, after obtaining the aggregated data corresponding to the original data of each vehicle type in the data packet, the step of labeling each piece of aggregated data may be performed.
[0071] For example, when reading each piece of raw data in the sub-data in sequence, if the currently read Perception Obstacle information, it can be said that the currently read other car data, at this time, the unextracted raw data and other car data in the storage queue corresponding to each data type can be aggregated to obtain an aggregated data related to the other car data. Then, the aggregated data can be annotated to obtain the data features corresponding to the aggregated data.
[0072] In one embodiment, in order to improve the temporal continuity of each piece of original data to be aggregated and the data of the own vehicle or other vehicles, the step of aggregating to obtain data features in the above S102-3 may also be implemented according to the following steps: Based on the timestamp of the currently read self-vehicle or other-vehicle data and the timestamp of each piece of original data extracted from the storage queue, determine whether the time difference between each piece of original data extracted from the storage queue and the self-vehicle or other-vehicle data meets the preset time difference; if so, aggregate the self-vehicle or other-vehicle data and each piece of original data extracted from the storage queue according to the target data structure to obtain the aggregated data of the self-vehicle or other-vehicle.
[0073] Here, the preset time difference can be set based on experience, and the embodiment of the present disclosure does not specifically limit it. For example, the preset time difference can be 0.1s, etc.
[0074] In specific implementation, for any self-vehicle or other-vehicle data, after extracting the unextracted raw data from the storage queue of each data type, the time difference between each extracted raw data and the corresponding self-vehicle or other-vehicle data can be determined according to the timestamp of the self-vehicle or other-vehicle data and the timestamp of each extracted raw data (here, if data is also read from the storage queue of the vehicle type, the currently read self-vehicle or other-vehicle data will be read at this time, and the self-vehicle or other-vehicle data can be ignored at this time). Then, it is determined whether each time difference is less than or equal to the preset time difference. If not, the self-vehicle or other-vehicle data and each extracted raw data can be ignored, data aggregation is not performed, and the next raw data in the sub-data packet is continued to be read. If so, it is determined that each extracted raw data has a strong continuity with the self-vehicle or other-vehicle data in time, and the self-vehicle or other-vehicle data and each raw data extracted from the storage queue can be integrated into a frame structure to obtain the corresponding integrated data of the self-vehicle or other-vehicle data.
[0075] Furthermore, after obtaining the aggregated data corresponding to the currently read self-vehicle or other-vehicle data, or after annotating the aggregated data corresponding to the currently read self-vehicle or other-vehicle data and obtaining the data features, the step of reading the next original data in the currently read sub-data packet can be returned to execute until all the original data in the currently read sub-data packet are read, and the aggregated data corresponding to each self-vehicle or other-vehicle data in the currently read sub-data packet can be obtained.
[0076] Furthermore, after all the original data in the currently read sub-data packet are read, the next sub-data packet in the current data packet can be read until all the sub-data packets in the current data packet are read, that is, the aggregated data corresponding to each piece of self-vehicle or other-vehicle data in the current data packet can be obtained.
[0077] like Figure 4 As shown, a flowchart of data aggregation and labeling provided by an embodiment of the present disclosure may include the following steps: S401: Read sub-data packets in sequence according to the order of each sub-data packet in the data packet.
[0078] S402: For the sub-data packet currently read, read each piece of original data in sequence according to the storage order of each piece of original data in the sub-data packet.
[0079] S403: According to the data type of the currently read original data, the original data is stored in a storage queue of the corresponding data type.
[0080] S404: Determine whether the currently read raw data is the data of the own vehicle or another vehicle.
[0081] That is, it is determined whether the data type of the currently read raw data is a vehicle type.
[0082] If yes, execute the following S405~S406; if no, execute the following S407.
[0083] S405: Extracting the unextracted original data currently in the storage queue from the storage queues of the respective data types.
[0084] S406: According to the target data structure, based on the timestamp and data identifier of the currently read own vehicle or other vehicle data, aggregate the own vehicle or other vehicle data and each piece of original data extracted from the storage queue to obtain aggregated data of the own vehicle or other vehicle.
[0085] Here, after executing S406, S407 may be executed.
[0086] S407: Whether all the original data in the sub-data packet currently read has been read.
[0087] If yes, execute the following S408; if no, return to execute S402 and continue to read the next original data in the currently read sub-data packet.
[0088] S408: Check whether all sub-data packets in the data packet have been read.
[0089] If yes, the following S409 can be executed; if no, the process returns to S401 to continue reading the next sub-data packet.
[0090] S409: Data labeling is performed on the aggregated data corresponding to the own vehicle or other vehicles to obtain data features corresponding to the own vehicle or other vehicles, and each data feature corresponding to the own vehicle or other vehicles is stored in a feature container.
[0091] Regarding the specific implementation steps of the above S401 to S409, reference may be made to the above embodiments and will not be repeated here.
[0092] In one embodiment, the above S103 may be implemented according to the following steps: S103-1: According to the training task, feature extraction is performed on the aggregated data of the own vehicle or other vehicles to obtain initial features related to the training task.
[0093] Here, the training task is used to characterize the functions that the deep learning model needs to achieve. The initial features are the features extracted from the aggregated data.
[0094] In specific implementation, for the aggregated data corresponding to any self-vehicle or other-vehicle data, the feature extraction task can be used to extract features from the aggregated data according to the training task of the deep learning model to obtain various initial features that match the training task. Among them, the role of the feature extraction task is to label the aggregated data corresponding to the obstacle vehicle (i.e., other vehicle) or the self-vehicle. The initial features may include, for example, determining whether the obstacle is a pedestrian or a vehicle based on the perception information, determining whether it is possible to move forward based on the traffic light information, generating some future executable trajectories for the self-vehicle based on the aggregated data corresponding to the target original data, and drawing a bev raster map based on the surrounding environment perceived by the self-vehicle, etc.
[0095] For example, in the case of training a deep learning model for vehicle trajectory prediction, initial features related to the historical trajectory of the vehicle, label initial features related to the future driving trajectory of the vehicle, current road environment features of the vehicle, current driving state features of the vehicle, bev raster map features, etc. can be extracted based on the aggregated information. Among them, the initial features related to the historical trajectory of the vehicle may include the xy coordinates of each trajectory point on the historical trajectory, the direction of the vehicle, the speed and acceleration of the vehicle, etc. The label initial features may include the xy coordinates of each trajectory point on the future trajectory, the direction of the vehicle, the speed and acceleration of the vehicle, etc.
[0096] S103-2: Combine the initial features into data features corresponding to the own vehicle or other vehicles.
[0097] During specific implementation, each initial feature corresponding to the data of the own vehicle or other vehicle can be stored according to a specific data structure (the specific structure can be a preset structure) to obtain the data features corresponding to the data of the own vehicle or other vehicle, and at the same time, the timestamp and data identifier of the data of the own vehicle or other vehicle can be used as the timestamp and data identifier corresponding to the data feature. After obtaining the data features corresponding to the data of the own vehicle or other vehicle, the data features can be stored in a local feature container. For example, the data features can be stored in a feature container according to the data identifier of the data features. A data packet can correspond to a feature container, and the data features corresponding to each piece of data of the own vehicle or other vehicle in a data packet can all be stored in the feature container corresponding to the data packet.
[0098] like Figure 5 FIG. 1 is a schematic diagram of a structure of a data feature provided by an embodiment of the present disclosure. Figure 5 The data features corresponding to multiple pieces of self-vehicle or other-vehicle data (i.e., 1, 2, 3, ...) are shown in FIG. Each piece of data feature is stored in a feature container. For any data feature, the data structure of the feature includes data identification, timestamp, speed, acceleration, future trajectory, historical trajectory, bev raster image, and a series of other features.
[0099] In one embodiment, the above S104-3 may be implemented according to the following steps: S104-3-1: Sort the data features in the feature sequence in chronological order according to the timestamp of each data feature to obtain a sorted feature sequence.
[0100] Here, for any vehicle, whether it is the vehicle itself or an obstacle vehicle, after obtaining each data feature with the same data identifier from the feature container according to the data identifier, the obtained data features are not necessarily arranged in time sequence, so they need to be sorted. Figure 6The figure is a schematic diagram of various data features with the same data identifier provided by an embodiment of the present disclosure, wherein each rectangular box shown represents a data feature, and the data identifiers of the various data features are consistent. The number under each data feature (e.g. Figure 6 0, 30, 49, etc.) is set up as an example to visualize the timestamp corresponding to each data feature. In actual implementation, the accuracy of the timestamp is much higher than Figure 6 According to the example in Figure 6 It can be seen that the timestamps of each data feature are out of order, so for subsequent feature extraction, it is necessary to sort them. Among them, feature is the data feature.
[0101] For example, the data features can be sorted in chronological order according to the timestamps of the data features in the feature sequence to obtain a sorted feature sequence that is monotonous in time scale. In the sorted feature sequence, each data feature is strictly arranged in chronological order.
[0102] S104-3-2: Take each data feature in the sorted feature sequence as a benchmark data feature in turn, take the benchmark data feature as the search starting point, and search forward and backward in the sorted feature sequence according to the timestamp and target duration of the benchmark data feature to obtain the target data features corresponding to the benchmark data feature.
[0103] During specific implementation, for each data feature in the sorted feature sequence, each data feature can be traversed in turn, and the currently traversed data feature can be used as the benchmark data feature. For the benchmark data feature, the benchmark data feature can be used as the search starting point for feature search, and forward search and backward search can be performed in the sorted feature sequence respectively to obtain the target data features searched in two directions. Among them, when searching forward and backward, the sum of the time spans of the searches in the two directions is equal to the target time, and the time spans of the searches in the two directions can be different. For example, when the target time is 4.5s, the time span of the forward search can be 1.5s, and the time span of the backward search can be 3s, that is, the data features within 1.5s are searched forward, and the data features within 3s are searched backward. Among them, the timestamp of the target data feature obtained by the forward search is earlier than the timestamp of the benchmark data feature; the timestamp of the target data feature obtained by the backward search is later than the timestamp of the benchmark data feature.
[0104] At the same time, if the time difference between the farthest data feature in the corresponding direction (that is, the data feature with the largest time difference with the benchmark data feature) and the benchmark data feature is less than the time span of the direction when searching forward or backward, it is determined that there are not enough data features in the direction, and the search in both directions can be stopped and the benchmark data feature can be ignored. Then the next benchmark data feature can be determined in the sorted feature sequence, and the new benchmark data feature can be processed. For example, if a data feature within 1.5s cannot be searched when searching forward, or if a data feature within 3s cannot be searched when searching backward, the current benchmark data feature can be ignored.
[0105] S104-3-3: Using the target array, each benchmark data feature and each corresponding target data feature are stored to obtain a model training data corresponding to each benchmark data feature.
[0106] Here, the target array is used to store each data feature in the model training data. Specifically, the target array can be a numpy array.
[0107] Exemplarily, for any benchmark data feature, after being able to find each target data feature that meets the target duration, the benchmark data feature and each target data feature corresponding to the benchmark data feature can be input into a numpy array and combined to obtain a model training data corresponding to the benchmark data feature.
[0108] Furthermore, in order to be able to better distinguish the characteristics of the values at each position in the target array, the type of characteristics to be stored in each position in the array can also be specified during input. For example, it can be pre-specified that the 0th to ith positions in the target array are used to store the index corresponding to the bev raster image, the i+1th to jth positions are used to store data characteristics related to historical trajectories, the j+1th to mth positions are used to store data characteristics related to future trajectories, and the m+1th to nth positions are used to store current environmental characteristics, etc. Figure 7 As shown, it is a schematic diagram of a feature storage in a target array provided by an embodiment of the present disclosure, in Figure 7 In the target array, bits 0 to 44 are used to store data features related to historical trajectories, where these data features are stored as (x, y, heading), where heading indicates the direction of the vehicle. Bits 45 to 134 are used to store data features related to future trajectories, where these data features are also stored as (x, y, heading). Regarding the features stored in other bits in the target array, Figure 7 Not explained one by one, you can set it as needed during the specific implementation.
[0109] In one embodiment, after the model training data is generated, it is also possible to: generate at least one training data set file using the model training data corresponding to each data packet.
[0110] In specific implementation, after generating the model training data corresponding to each data packet, in order to facilitate the rapid reading of each model training data when training the deep learning model, the model training data can also be output in the data set file format.
[0111] For example, for any date, multiple model training data corresponding to the data packet under the date can be packaged into a training data set file corresponding to the date. In this way, a training data set file corresponding to each date can be obtained. Alternatively, multiple model training data corresponding to the data packets under adjacent dates or multiple dates with a date interval less than a preset interval can be packaged into a training data set file. In this way, a training data set file corresponding to multiple dates can be obtained.
[0112] Among them, the data set file can be a file in a target format, and the target format can be set according to experience, and the embodiment of the present disclosure does not specifically limit it. For example, the target format can be in HDF5 format. Exemplarily, each model training data can be stored using a numpy array, and then the numpy arrays corresponding to the multiple model training data can be stacked to obtain a training data set file in HDF5 format. After obtaining the training data set file, the training data set file can be used as a training data set for a deep learning model in an autonomous driving system from the perception end to the decision end of the vehicle.
[0113] like Figure 8 As shown, it is a schematic diagram of a training data set file in HDF5 format provided by an embodiment of the present disclosure, wherein the training data set file includes n model training data stored in numpy arrays (such as Figure 8 numpy1~numpyn in ), n is greater than 1.
[0114] In one embodiment, S104-3-2 may be implemented as follows: S104-3-2-1: Taking the benchmark data feature as the search starting point, for any target direction in the forward and backward directions in the sorted feature sequence, search the sorted feature sequence according to the target direction to obtain the next data feature of the benchmark data feature.
[0115] In specific implementation, for any benchmark data feature, the benchmark data feature can be used as the search starting point, and for the sorted feature sequence, the next data feature of the benchmark data feature is searched in each target direction in the forward and backward directions from the search starting point.
[0116] Taking the target direction as the forward direction as an example, you can search forward from the benchmark data feature in the sorted feature sequence to obtain the previous data feature of the benchmark data feature, and use this data feature as the next data feature corresponding to the benchmark data feature in the forward direction. Taking the target direction as the backward direction as an example, you can search backward from the benchmark data feature in the sorted feature sequence to obtain the next data feature of the benchmark data feature, and use this data feature as the next data feature corresponding to the benchmark data feature in the backward direction.
[0117] S104-3-2-2: If the time difference between the reference data feature and the next data feature is less than or equal to the preset time difference, the next data feature is used as the first data feature, and the next data feature of the first data feature is searched in the sorted feature sequence.
[0118] Here, the preset time difference can be set based on experience, and the embodiment of the present disclosure does not perform specific settings. For example, the preset time difference can be 0.1s.
[0119] Exemplarily, after searching for the next data feature of the baseline data feature, it can be determined whether the time difference between the baseline data feature and the next data feature searched is greater than a preset time difference. If not, the next data feature can be used as the first data feature, and the next data feature of the first data feature can be searched in the sorted feature sequence; if so, the baseline data feature is ignored, and S104-3-2 is returned to continue traversing the sorted feature sequence to find a new baseline data feature, and the feature search is performed with the new baseline data feature as the search starting point.
[0120] S104-3-2-3: If the time difference between the first data feature and the next data feature of the first data feature is less than or equal to the preset time difference, return to the step of taking the next data feature as the first data feature until any target condition among the multiple loop end conditions is met.
[0121] Among them, multiple loop end conditions include that the time difference between the first data feature and the next data feature is greater than the preset time difference, the time difference between the next data feature and the benchmark data feature is greater than or equal to the search time corresponding to the target direction, and the next data feature of the first data feature cannot be searched in the sorted feature sequence.
[0122] Here, one search direction corresponds to one search duration, and the search durations corresponding to different search directions may be different. The sum of the search durations corresponding to the two search directions is equal to the target duration. For example, the search duration corresponding to the forward direction may be 1.5s, and the search duration corresponding to the backward direction may be 3s, that is, the forward search is for historical data features within 1.5s, and the backward search is for future data features within 3s. It is understandable that in specific applications, the search durations corresponding to different search directions may be set as needed, and are not specifically limited here.
[0123] like Fig. 9 FIG. 1 is a schematic diagram of a sorted feature sequence provided by an embodiment of the present disclosure, wherein each box in the feature sequence represents a data feature. Fig. 9 , a schematic diagram of a benchmark data feature as an example and a data feature obtained by searching in the forward and backward directions of the benchmark data feature is shown. Among them, the corresponding search duration in the forward direction is 1.5s, and the corresponding search duration in the backward direction is 3s. That is, in the disclosed embodiment, when searching for data features from the sorted feature sequence, each data feature in the sorted feature sequence can be used as a benchmark data feature in turn, and then the benchmark data feature is used as the zero point, and the search duration corresponding to the backward direction in the backward direction of the zero point is used as the end point, and the search duration corresponding to the forward direction in the forward direction of the zero point is used as the end point, so as to obtain a search window with the benchmark data feature as the zero point and the duration span as the target duration. Then, model training data can be generated according to each data feature in the search window. Among them, after determining the search window corresponding to the benchmark data feature, whether the front end or the back end of the search window exceeds the sorted feature sequence, the benchmark data feature can be ignored, and the next benchmark data feature and its corresponding search window can be searched.
[0124] For any target direction, when searching for the target data feature corresponding to the benchmark data feature in the sorted feature sequence according to the target direction, you can first determine whether there is a next data feature in the target direction. If not, you can stop the search, ignore the current benchmark data feature, and process the next benchmark data feature. If so, you can search for data features that meet the search duration of the target direction in sequence in the target direction. For each data feature searched, you can determine whether the following conditions are met: the time difference between the data feature searched each time and the benchmark data feature is greater than or equal to the search duration of the target direction, and the time difference between the data feature searched each time and the data feature searched the previous time in the target direction is greater than the preset time difference. If either of the two conditions is met, the search can be ended. If both conditions are not met, you can continue to search for the next data feature.
[0125] Among them, for each data feature searched, if the time difference between the data feature and the data feature searched in the target direction the previous time is greater than the preset time difference, it means that the data feature searched the current time is unreasonable and illegal. Therefore, the current benchmark data feature can be directly ignored and the next benchmark data feature can be searched, that is, the next round of window processing can be performed. Fig.10 FIG. 1 is a schematic diagram of an illegal situation provided by an embodiment of the present disclosure, wherein: Fig.10 The two data features in are two data features obtained by two adjacent searches in the target direction. The time difference between the two data features is greater than the preset time difference, so it is an illegal situation.
[0126] Exemplarily, after searching for the next data feature corresponding to the first data feature, it can be determined whether the time difference between the data feature and the first data feature is greater than a preset time length. If so, the current baseline data feature can be ignored. If not, it can return to the step of "taking the next data feature as the first data feature" in S104-3-2-2, and continue to search for a new next data feature until any target condition among the multiple loop end conditions is met.
[0127] S104-3-2-4: Obtain target data features corresponding to the benchmark data features based on the satisfied target conditions and the data features obtained through the search.
[0128] For example, after a search satisfies the target condition, it can be determined whether the baseline data feature needs to be ignored based on the satisfied target condition; if so, the next baseline data feature can be processed; if not, each target data feature obtained from the search can be used as the target data feature.
[0129] In one embodiment, the target duration may include search durations corresponding to different target directions; the above S104-3-2-4 may be implemented according to the following steps: S104-3-2-4-1: When the target condition is satisfied that the time difference between the next data feature and the reference data feature is greater than or equal to the search duration corresponding to the target direction, determine the initial data feature in the target direction based on each data feature obtained by the search; Here, when the target condition is met that the time difference between the first data feature and the next data feature is greater than the preset time difference, it can be explained that the time difference between the two adjacent data features searched twice in the target direction is greater than the preset time difference, so the benchmark data feature can be ignored, and the search and search for the next benchmark data feature can be performed. When the target condition is met that the next data feature corresponding to the first data feature cannot be searched in the sorted feature sequence, it means that the data feature with the largest time difference with the benchmark data feature in the target direction is less than the search time in the target direction, and the number of data features in the target direction is insufficient to meet the requirements for building model training data, so the benchmark data feature can also be ignored, and the search and search for the next benchmark data feature can be performed.
[0130] When the target condition is met that the time difference between the next data feature and the baseline data feature is greater than or equal to the search time corresponding to the target direction, it means that a sufficient number of data features have been searched in the target direction, so the search can be stopped and the initial data feature in the target direction is determined based on the data features obtained from the search.
[0131] Here, the reason why the time difference between the next data feature and the benchmark data feature is greater than the corresponding search time in the target direction is that the intervals between each data feature in the sorted feature sequence are not the same, and are not strictly in accordance with the preset time difference. Therefore, it may happen that the time difference between the currently searched data feature and the benchmark data feature is greater than the corresponding search time in the target direction, while the time difference between the previously searched data feature and the benchmark data feature is less than the corresponding search time in the target direction. At this time, the currently searched data feature can be used as the initial data feature, or the data features searched in other times except the currently searched data feature can be used as the initial data feature. It is understandable that if the time difference between the next data feature and the benchmark data feature is equal to the corresponding search time in the target direction, then each searched data feature can be used as the initial data feature in the target direction.
[0132] Exemplarily, if the length of the sorted feature sequence is 500, that is, it includes 500 data features, then when generating model training data, 500 data features can be traversed in sequence. That is, you can start from the first data feature and look forward for historical data features within 1.5s. Since it is the data feature, there is no historical data feature, so the data feature is discarded. Assuming that when traversing to the 16th data feature, look forward for historical data features within 1.5s. At this time, not only must it be ensured that the time difference between the timestamp of the farthest historical data feature and the timestamp of the currently traversed data feature is greater than or equal to 1.5s, but it must also be ensured that the time difference between adjacent data features cannot be greater than the preset time difference. Otherwise, the data feature will be considered as an illegal feature and then discarded. Similarly, when traversing to 469 data features, look backward for 30 data features. Only when it is ensured that the time difference between the timestamp of the farthest future data feature and the timestamp of the currently traversed data feature is greater than or equal to 3s, the data feature is a legal feature, otherwise the data feature is discarded.
[0133] S104-3-2-4-2: Determine the target data features corresponding to the benchmark data features based on the benchmark data features and the initial data features in different target directions.
[0134] Exemplarily, for any reference data feature, each initial data feature obtained by searching for the reference data feature in different target directions may be directly used as each target data feature corresponding to the reference data feature.
[0135] like Fig.11 As shown, a flowchart of a forward search and a backward search provided by an embodiment of the present disclosure may include the following steps: S1101: Determine a feature sequence corresponding to the same data identifier.
[0136] S1102: Sort each data feature in the feature sequence in chronological order to obtain a sorted feature sequence.
[0137] S1103: Traverse each data feature in the sorted feature sequence.
[0138] S1104: for the currently traversed benchmark data feature, determine whether the data feature searched in the forward search and the backward search meets the conditions. The conditions include that the time difference between the data feature currently searched in the target direction and the benchmark data feature is greater than or equal to the search duration, and the time difference between the data feature currently searched in the target direction and the data feature searched in the target direction last time is not greater than a preset time difference.
[0139] If so, execute the following S1105; if not, ignore the currently traversed benchmark data features and return to S1103.
[0140] S1105: taking each piece of data features obtained through the search as an initial data feature in the target direction, and determining each piece of target data features corresponding to the benchmark data feature according to the benchmark data feature and the initial data features in different target directions.
[0141] S1106: Determine whether the traversal of the sorted feature sequence is complete.
[0142] If yes, the process ends; if no, the process returns to S1103.
[0143] Regarding the specific implementation process of the above S1101~S1106, reference can be made to the above embodiments, which will not be repeated here.
[0144] In one embodiment, the above S104-3-2-4-2 can be implemented according to the following steps: A1: For any target direction, the reference data feature is used as the search starting point, the reference data feature is used as the second data feature, and a target timestamp having a preset time difference with the second data feature in the target direction is determined.
[0145] Exemplarily, regardless of whether the target direction is forward or backward, since the time interval between the initial data features corresponding to the benchmark data features is not strictly the preset time difference, in order to obtain target data features with equal time intervals, the initial data features obtained by the search can also be interpolated.
[0146] Exemplarily, for any target direction, the baseline data feature can be used as the search starting point and the baseline data feature can be used as the second data feature, and then the target timestamp with a preset time length from the second data feature can be determined based on the timestamp of the second data feature and the preset time difference.
[0147] A2: If no initial data feature with a target timestamp is found in the target direction, the target data feature corresponding to the target timestamp is determined by using two initial data features whose timestamps are adjacent to the target timestamp.
[0148] Here, the initial data feature with the target timestamp is the initial data feature with the same timestamp as the target timestamp. The two initial data features adjacent to the target timestamp may include an initial data feature with a timestamp earlier than the target timestamp and the smallest time difference between the timestamp and the target timestamp, and an initial data feature with a timestamp later than the target timestamp and the smallest time difference between the timestamp and the target timestamp.
[0149] In specific implementation, after obtaining the target timestamp, it can be determined whether there is an initial data feature with the same timestamp as the target timestamp according to the timestamps of each initial data feature in the target direction. If not, two initial data features with timestamps adjacent to the target timestamp can be found from each initial data feature according to the timestamps of each initial data feature in the target direction. Then, data difference is performed using these two initial data features to obtain the target data feature corresponding to the target timestamp, that is, to obtain the target data feature with the timestamp of the target timestamp, which has a preset time difference with the second data feature.
[0150] Among them, when the data is interpolated, the features related to the trajectory (such as xy coordinates, speed, direction, etc.) in the two initial data features used for the interpolation can be averaged to obtain the post-interpolation features, and the post-interpolation features are used as the track-related features in the target data features. For the features that are not scalars in the two initial data features used for the interpolation, the feature of the initial data feature with the closest timestamp to the target timestamp can be directly used as the target data feature.
[0151] For example, taking the preset time difference of 0.1s as an example, when the time difference between each initial data feature corresponding to the second data feature and the second data feature is not 0.1s, the features related to the trajectory in the two initial data features adjacent to 0.1s can be used, and the features related to the trajectory corresponding to the target data feature at 0.1s can be obtained by taking the average or the like. At the same time, non-scalar features such as environmental features and bev grating images in the initial data features closest to 0.1s can be used as non-scalar features corresponding to the target data feature at 0.1s.
[0152] In another embodiment, if an initial data feature with a target timestamp is searched in the target direction, the initial data feature can be used as a target data feature with a target timestamp, and the target data feature can be used as a new second data feature, and the step of determining a target timestamp with a preset time difference with the second data feature in the target direction is returned until a target data feature is obtained whose time difference with the baseline data feature is the search duration corresponding to the target direction.
[0153] A3: Take the target data feature as a new second data feature, and return to the step of determining a target timestamp having a preset time difference with the second data feature in the target direction, until a target data feature having a time difference with the reference data feature equal to the search duration corresponding to the target direction is obtained.
[0154] Exemplarily, after obtaining a target data feature having a preset time difference with the second data feature, the target data feature can be used as a new second data feature, and then the step of determining the target timestamp in A1 is returned to execute until a target data feature having a time difference with the benchmark data feature corresponding to the search duration in the target direction is obtained. At this time, the time intervals between each adjacent target data feature in the target direction are the preset time difference, and the time interval between the target data feature adjacent to the benchmark data feature and the benchmark data feature is also the preset time difference. In this way, after the target data features are determined both forward and backward, a windowed feature sequence composed of forward historical data features and backward future data features arranged at equal intervals can be obtained.
[0155] For example, taking the backward search for future data features of 3s with a preset time difference of 0.1s as an example, the above A1~A3 is executed cyclically until the target data feature with a time difference of 3s with the benchmark data feature is determined, and the determined target data features are used as the target data features in the backward direction. That is, for the benchmark data feature, in the backward direction, from the benchmark data feature, there may be a target data feature at every 0.1s, and the target data feature with the maximum time interval with the benchmark data feature is the target data feature at 3s. The same is true for the forward search. Taking the forward search for future data features of 1.5s with a preset time difference of 0.1s as an example, the above A1~A3 is executed cyclically until the target data feature with a time difference of 1.5s with the benchmark data feature is determined, and the determined target data features are used as the target data features in the forward direction. That is, for the reference data feature, in the forward direction, there may be a target data feature at every 0.1 s from the reference data feature forward, and the target data feature having the maximum time interval with the reference data feature is the target data feature at 1.5 s.
[0156] Still taking the example of the length of the above sorted feature sequence being 500, assuming that the 50th data feature is currently traversed, the historical data features of this data feature are 30-49, and the future data features are 51-85. At this point, it can be ensured that the time difference between the timestamp of the furthest historical data feature and future data feature and the timestamp of the currently traversed data feature is greater than 1.5s and 3s respectively, so the historical data features 30-49 and the future data features 51-85 can be used as initial data features. However, the interval between each initial data feature is not strictly equal to 0.1s, so it is necessary to interpolate based on each historical initial data feature and each future initial data feature to obtain the historical target data features and future target data features with equal time intervals.
[0157] For any benchmark data feature, after obtaining the target data features corresponding to the benchmark data feature in the forward and backward directions respectively, each target data feature in the forward direction and the benchmark data feature can be used as input data of the deep learning model, and each target data feature in the backward direction can constitute the label data of the input data.
[0158] like Fig.12 As shown, a flowchart of generating a windowed feature sequence provided by the present embodiment may include the following steps: S1201: Discard window features whose forward features and backward features do not meet the conditions, and obtain initial window features.
[0159] The initialization window feature is composed of the benchmark data feature and the initial data feature corresponding to the benchmark data feature in the forward and backward directions respectively. The window feature that does not meet the conditions can be understood as follows: for the benchmark data feature obtained through traversal, when searching for data features, if the time difference between the data feature currently searched in the target direction and the data feature last searched in the target direction is greater than the preset time difference, then the benchmark data feature is the window feature that does not meet the conditions; and when the next data feature of the benchmark data feature cannot be searched in the target direction, the benchmark data feature is the window feature that does not meet the conditions.
[0160] S1202: For any target direction of any initial windowed feature, take the reference data feature as the search starting point, use the reference data feature as the second data feature, and determine a target timestamp with a preset time difference with the second data feature in the target direction.
[0161] S1203: If no initial data feature with the target timestamp is found in the target direction, two initial data features whose timestamps are adjacent to the target timestamp are used to generate a target data feature corresponding to the target timestamp.
[0162] S1204: Take the target data feature as a new second data feature, and return to the step of determining a target timestamp having a preset time difference with the second data feature in the target direction, until a target data feature having a time difference with the reference data feature equal to the search duration corresponding to the target direction is obtained.
[0163] S1205: Obtain equally spaced target window features based on the target data features obtained from different target directions.
[0164] Regarding the specific implementation process of S1201~S1205, reference may be made to the above embodiments and will not be repeated here.
[0165] After obtaining the equally spaced target window features, the target window features can be input into a numpy array for combination to obtain a model training data, thereby realizing data format conversion and output.
[0166] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.
[0167] Based on the same technical concept, the embodiment of the present disclosure also provides a computer device. Fig.13 FIG. 1 is a schematic diagram of a computer device provided in an embodiment of the present disclosure, including: Processor 1301, memory 1302 and bus 1303. The memory 1302 stores machine-readable instructions executable by the processor 1301, and the processor 1301 is used to execute the machine-readable instructions stored in the memory 1302. When the machine-readable instructions are executed by the processor 1301, the processor 1301 performs the following steps: S101: Acquire a data set to be converted, the data set includes multiple data packets classified by date; the data packet includes multiple sub-data packets, and the sub-data packets include multiple original data stored in time sequence; S102: For any data packet, in the process of reading data according to the storage order of each original data included in each sub-data packet in the data packet, whenever the self-vehicle or other-vehicle data is read, according to the target data structure, the various types of original data that have been read and not aggregated are aggregated to obtain the aggregated data of the self-vehicle or other-vehicle; S103: Generate data features corresponding to the self-vehicle or other-vehicle according to the aggregated data of the self-vehicle or other-vehicle; and S104: Generate model training data according to each data feature corresponding to the self-vehicle or other-vehicle.
[0168] The above-mentioned memory 1302 includes internal memory 1321 and external memory 1322; the memory 1321 here is also called internal memory, which is used to temporarily store the calculation data in the processor 1301, as well as the data exchanged with the external memory 1322 such as the hard disk. The processor 1301 exchanges data with the external memory 1322 through the internal memory 1321. When the computer device is running, the processor 1301 communicates with the memory 1302 through the bus 1303, so that the processor 1301 executes the execution instructions mentioned in the above-mentioned method embodiment.
[0169] The present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the training data generation method described in the above method embodiment are executed. The storage medium can be a volatile or non-volatile computer-readable storage medium.
[0170] The computer program product of the training data generating method provided in the embodiments of the present disclosure includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the steps of the training data generating method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.
[0171] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (SDK).
[0172] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.
[0173] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0174] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0175] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0176] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.
[0177] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure is described in detail with reference to the aforementioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the aforementioned embodiments within the technical scope disclosed in the present disclosure, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be based on the protection scope of the claims.
Claims
1. A method for generating training data, characterized in that: include: Acquire a data set to be converted, the data set comprising a plurality of data packets classified by date; The data packet includes a plurality of sub-data packets, each of which includes a plurality of original data stored in a time sequence; For any of the data packets, in the process of reading data according to the storage order of the pieces of original data included in each of the sub-data packets in the data packet, whenever data of the own vehicle or another vehicle is read, the read and unaggregated original data of various types are aggregated according to the target data structure to obtain aggregated data of the own vehicle or another vehicle; Generating data features corresponding to the own vehicle or other vehicles according to the aggregated data of the own vehicle or other vehicles; Generate model training data based on the data features corresponding to the vehicle or other vehicles described in each item.
2. The method according to claim 1, characterized in that The generating model training data according to the data features corresponding to the self-vehicle or other vehicles in each of the items includes: Using the data identifier and timestamp of the data of the own vehicle or the other vehicle as the data identifier and timestamp of the data feature corresponding to the own vehicle or the other vehicle; Determine a feature sequence corresponding to the same data identifier according to the data identifier of the data feature and each piece of data feature corresponding to the own vehicle or the other vehicle; According to the timestamp and target duration of each data feature in each of the feature sequences, at least one model training data corresponding to each of the data identifiers is generated; the model training data includes multiple target data features that are continuous in time sequence, and the maximum time difference between the multiple target data features is less than or equal to the target duration.
3. The method according to claim 1, characterized in that In the process of reading data according to the storage order of each piece of original data included in each sub-data packet in the data packet, whenever the data of the own vehicle or other vehicle is read, the various types of original data that have been read and not aggregated are aggregated according to the target data structure to obtain the aggregated data of the own vehicle or other vehicle, including: In the process of reading data according to the storage order of each piece of original data included in each sub-data packet in the data packet, if the original data currently read does not belong to the vehicle data or other vehicle data, the original data is stored in a storage queue of a corresponding type according to the type of the original data; If the data of the own vehicle or other vehicles is currently read, the unextracted original data currently in the storage queue is extracted from each type of storage queue respectively; According to the target data structure, the self-vehicle or other-vehicle data and each piece of original data extracted from the storage queue are aggregated to obtain the aggregated data of the self-vehicle or other-vehicle.
4. The method according to claim 3, characterized in that The step of aggregating the self-vehicle or other-vehicle data and each piece of original data extracted from the storage queue according to the target data structure to obtain the aggregated data of the self-vehicle or other-vehicle includes: Determine, based on the timestamp of the currently read self-vehicle or other-vehicle data and the timestamp of each piece of original data extracted from the storage queue, whether the time difference between each piece of original data extracted from the storage queue and the self-vehicle or other-vehicle data satisfies a preset time difference; If so, the self-vehicle or other-vehicle data and each piece of original data extracted from the storage queue are aggregated according to the target data structure to obtain the aggregated data of the self-vehicle or other-vehicle.
5. The method according to any one of claims 1 to 4, characterized in that The step of generating data features corresponding to the own vehicle or other vehicles according to the aggregated data of the own vehicle or other vehicles includes: According to the training task, feature extraction is performed on the aggregated data of the own vehicle or other vehicles to obtain initial features related to the training task; The initial features are combined into data features corresponding to the own vehicle or the other vehicle.
6. The method according to claim 2, characterized in that The step of generating at least one model training data corresponding to each of the data identifiers according to the timestamp and target duration of each of the data features in each of the feature sequences comprises: According to the timestamps of the data features in the feature sequence, the data features are sorted in chronological order to obtain a sorted feature sequence; Taking each data feature in the sorted feature sequence as a benchmark data feature in turn, taking the benchmark data feature as a search starting point, searching forward and backward in the sorted feature sequence according to the timestamp of the benchmark data feature and the target duration, respectively, to obtain each target data feature corresponding to the benchmark data feature; By using the target array, each of the reference data features and the corresponding target data features are stored to obtain a piece of model training data corresponding to each of the reference data features.
7. The method according to claim 6, characterized in that The method of taking the benchmark data feature as the search starting point and searching forward and backward in the sorted feature sequence according to the timestamp of the benchmark data feature and the target duration to obtain each target data feature corresponding to the benchmark data feature includes: Taking the reference data feature as a search starting point, searching in the sorted feature sequence according to the target direction for any one of the forward and backward target directions in the sorted feature sequence to obtain the next data feature of the reference data feature; If the time difference between the reference data feature and the next data feature is less than or equal to the preset time difference, the next data feature is used as the first data feature, and the next data feature of the first data feature is searched in the sorted feature sequence; If the time difference between the first data feature and the next data feature of the first data feature is less than or equal to the preset time difference, return to the step of taking the next data feature as the first data feature until any target condition among the multiple loop end conditions is met; According to the satisfied target conditions and the data features obtained by searching, target data features corresponding to the reference data features are obtained.
8. The method according to claim 7, characterized in that The target duration includes the search duration corresponding to different target directions; The step of obtaining each target data feature corresponding to the reference data feature according to the target condition satisfied and each data feature obtained by searching includes: When the target condition is satisfied that the time difference between the next data feature and the reference data feature is greater than or equal to the search duration corresponding to the target direction, determining the initial data feature in the target direction according to each data feature obtained by the search; According to the reference data feature and the initial data features in different target directions, each target data feature corresponding to the reference data feature is determined.
9. The method according to claim 8, characterized in that The determining, according to the reference data feature and the initial data features in different target directions, each target data feature corresponding to the reference data feature comprises: For any of the target directions, taking the reference data feature as a search starting point, taking the reference data feature as a second data feature, and determining a target timestamp having a preset time difference with the second data feature in the target direction; If no initial data feature with the target timestamp is found in the target direction, the target data feature corresponding to the target timestamp is determined by using two initial data features whose timestamps are adjacent to the target timestamp. The target data feature is used as a new second data feature, and the step of determining a target timestamp having a preset time difference with the second data feature in the target direction is returned until a target data feature is obtained whose time difference with the benchmark data feature is a search duration corresponding to the target direction.
10. A computer device, characterized in that: include: A processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the processor is used to execute the machine-readable instructions stored in the memory. When the machine-readable instructions are executed by the processor, the processor executes the steps of the training data generation method according to any one of claims 1 to 9.
11. A computer program product, which, when executed by a computer device, performs the steps of the training data generation method according to any one of claims 1 to 9.
Citation Information
Patent Citations
Sensor aggregation framework for autonomous driving vehicles
CN110377025A
Training sample generation method and device in elevated scene, equipment and storage medium
CN114358196A
To-be-trained data generation method and device, electronic equipment and storage medium
CN116304834A
Driving data processing method and device, computer equipment and storage medium
CN118506473A
Systems and methods for responding to predicted events in time-series data using synthetic profiles created by artificial intelligence models trained on non-homogonous time-series data
US20240193401A1
