Systems and methods to support fast iteration and quantized deployment of trajectory prediction algorithms

Abstract

The application discloses a kind of systems and methods for supporting trajectory prediction algorithm fast iteration and quantization deployment, its system includes: data sample processing module, for integrating multiple source heterogeneous data, generate standard format dynamic sample library, and generate standardized preprocessed trajectory prediction sample set;Model training and quantization module, for using preprocessed trajectory prediction sample set to carry out multi-stage training trajectory prediction model in model library, including first floating point model training, then carry out quantization perception training;Model deployment and test module, for trajectory prediction model after multi-stage training is carried out hardware-in-the-loop test and automatic driving real vehicle test, to further iteratively optimize trajectory prediction model according to test result.The trajectory prediction model of the system and method disclosed in the application shortens the iteration cycle of automatic driving trajectory prediction algorithm, reduces the model quantization precision loss, improves the real vehicle verification efficiency, improves the algorithm development efficiency and the reliability of deployment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to a system and method for supporting rapid iteration and quantitative deployment of trajectory prediction algorithm. BACKGROUND

[0002] Trajectory prediction technology is one of the key technologies of automatic driving, and is increasingly concerned by the academic and industrial circles. Current trajectory prediction mainly relies on deep learning technology. The model analyzes historical trajectory data, environmental information and traffic flow data to predict the future behavior of traffic participants such as vehicles or pedestrians. The improvement of trajectory prediction accuracy directly affects the key links such as decision-making and path planning of the automatic driving system, and is the basis for ensuring the safe and stable operation of automatic driving.

[0003] However, the development iteration and real vehicle deployment of existing trajectory prediction algorithms face challenges such as difficulty in integrating multi-source heterogeneous data, long iteration period of algorithms, serious precision loss in model quantitative deployment, single evaluation dimension, and imperfect model verification system for real vehicle deployment. Specifically, there are significant differences in format, coordinate system and annotation standard between public data sets and real vehicle data, making it difficult to be used for training together and the sample library difficult to be dynamically expanded; the traditional training process adopts a single mode of "supervised training-real vehicle tuning", which fails to design training and testing methods targeting real vehicle deployment; the performance test of the deployed model mainly relies on offline simulation test and real vehicle test, lacking a standard test mechanism combining hardware-in-the-loop (HIL) test and real vehicle closed-loop test, and the test time and hardware cost are high; these problems seriously restrict the rapid iteration, high-precision deployment and all-around performance test of trajectory prediction algorithms, and an all-link solution covering data processing, model training and verification, quantitative compression and deployment testing is urgently needed. SUMMARY

[0004] Therefore, the embodiments of the present application provide a system and method for supporting rapid iteration and quantitative deployment of trajectory prediction algorithm, aiming to improve the construction, training, evaluation and deployment efficiency of trajectory prediction model in automatic driving system.

[0005] The technical solution adopted by the present application is:

[0006] A system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithm is provided, comprising:

[0007] A data sample processing module is configured to integrate multi-source heterogeneous data, construct a dynamically expandable sample library through a standardized format conversion mechanism, and generate a standardized preprocessed trajectory prediction sample set; wherein the sample library includes a public data set and a real vehicle data set collected in stages;

[0008] The model training and quantization module is configured to perform multi-stage training on a pre-constructed trajectory prediction model of an autonomous vehicle using a preprocessed trajectory prediction sample set, including floating-point model training and quantization-aware training, wherein the quantization-aware training includes parameter quantization calibration of the trajectory prediction model after floating-point model training, insertion of a pseudo-quantization node in the trajectory prediction model to generate a pseudo-quantization model, fine-tuning of parameters of the pseudo-quantization model, and finally conversion into a fixed-point model.

[0009] The model deployment and testing module is configured to perform hardware-in-the-loop testing and autonomous vehicle testing on the trajectory prediction model after multi-stage training, to further optimize and iterate the trajectory prediction model according to the test results.

[0010] In the above technical solution, the floating-point model training specifically includes pre-training of a vehicle trajectory encoder and a map encoder in the trajectory prediction model through a self-supervised learning mechanism, followed by supervised training on a public data set, and finally transfer learning using a real vehicle data set to train floating-point parameters in the trajectory prediction model that are adapted to the real vehicle data domain.

[0011] In the above technical solution, the hardware-in-the-loop testing specifically includes pre-validation of model inference latency and resource occupation through an embedded twin environment; and the autonomous vehicle testing specifically includes closed-loop scene adaptability testing of the trajectory prediction model through a vehicle-mounted computing platform and a line control execution mechanism, and road scene data collected during the vehicle testing process is fed back to the data sample processing module for sample library expansion.

[0012] In the above technical solution, the multi-source heterogeneous data includes public road data and self-collected vehicle data, and the data sample processing module generates a public data set by integrating the public road data, and integrates real vehicle data collected in real scenes by vehicle-mounted sensing devices into a real vehicle data set.

[0013] In the above technical solution, the vehicle-mounted sensing device is configured to perceive surrounding environment information and publish raw observation data in real time, and mainly includes a camera, a laser radar, an inertial measurement unit, and a wheel speed sensor.

[0014] In the above technical solution, the model training and quantization module further includes performance evaluation, which is configured to perform performance evaluation in stages during floating-point model training and quantization-aware training, monitor the training effect and convergence of the model during the process, and provide intuitive evaluation and optimization direction for the prediction ability of the model in complex scenarios.

[0015] According to the technical scheme, during the hardware-in-the-loop test, a data backfill program is run to obtain and replay the newly collected perception result data, including motor vehicle and non-motor vehicle perception results, vehicle state information and road structure information, from the real vehicle data set; the perception result data is preprocessed on a pre-constructed twin computing platform, the preprocessed data is trajectory predicted by a multi-stage trained trajectory prediction model, the trajectory prediction result is post-processed, and the trajectory prediction result data is published.

[0016] According to the technical scheme, during the hardware-in-the-loop test, a model performance evaluation program is also run to evaluate the model prediction performance of the prediction result and the perception result data, including quantitative performance evaluation and qualitative performance evaluation; wherein the quantitative performance evaluation is calculated according to three indexes of trajectory prediction accuracy, trajectory prediction time domain and trajectory prediction time delay; the qualitative performance evaluation is visualized by using a specific rendering tool for real-time perception results and prediction results.

[0017] The application also provides a method for rapid iteration and quantitative deployment of an automatic driving vehicle trajectory prediction algorithm, comprising the following steps:

[0018] The multi-source heterogeneous data is integrated, a dynamic and expandable sample library is constructed through a standardized format conversion mechanism, and a standardized preprocessed trajectory prediction sample set is generated; wherein the sample library includes a public data set and a real vehicle data set collected in stages;

[0019] The preprocessed trajectory prediction sample set is used to perform multi-stage training on a pre-constructed automatic driving vehicle trajectory prediction model, including first floating point model training, and then quantitative perception training, wherein the quantitative perception training includes quantizing and calibrating the parameters of the trajectory prediction model after floating point model training, inserting a pseudo-quantization node in the trajectory prediction model to generate a pseudo-quantization model, fine-tuning the parameters of the pseudo-quantization model, and finally converting into a fixed point model;

[0020] The trajectory prediction model after multi-stage training is tested in a hardware-in-the-loop and automatic driving real vehicle to further optimize and iterate the trajectory prediction model according to the test results.

[0021] The application also provides a computer storage medium having a computer program stored therein, which can be executed by a processor, and the computer program executes the method for supporting rapid iteration and quantitative deployment of a trajectory prediction algorithm according to the technical scheme.

[0022] The beneficial effects generated by the present application are: the present application trains the floating point model and the quantization perception of the trajectory prediction model of the automatic driving car, converts the trajectory prediction model with floating point weight precision into a fixed point model through multi-stage training, wherein the quantization perception training includes three steps of quantization parameter calibration, pseudo quantization fine tuning and fixed point model conversion, ensures efficient conversion of the floating point precision model to the low bit fixed point model, while minimizing the accuracy loss in the quantization process, improves the real vehicle verification efficiency, greatly shortens the iteration period of the automatic driving trajectory prediction algorithm, and significantly improves the algorithm development efficiency and the reliability of the deployment.

[0023] Further, the floating point model training is used for self-supervised pre-training, to provide initial weights for downstream public dataset supervised training and real vehicle data transfer learning; and the public dataset supervised training is used to provide a trajectory prediction model with strong generalization ability for downstream real vehicle data transfer learning; finally, the real vehicle data transfer learning is used to fully utilize the diversified driving scene data collected by the real vehicle, to reduce the performance deviation caused by the distribution difference between the public dataset and the real vehicle data based on the weight regularization technology, and to improve the sampling probability of long-tail scene samples based on the data resampling strategy of sample equalization and diversification processing, and further improve the performance of the trajectory prediction model on the real vehicle dataset.

[0024] Further, the automatic driving real vehicle test specifically realizes the closed-loop scene adaptability test of the trajectory prediction model through the vehicle-mounted computing platform and the line control execution mechanism, and the road scene data collected in the real vehicle test process is fed back to the data sample processing module, so as to be used for dynamically expanding the sample library.

[0025] Further, by adding performance evaluation in the model training and quantization module, the training effect and convergence of the model can be monitored during the floating point model training and quantization perception training, and the intuitive evaluation and optimization direction of the prediction ability of the model under complex scenes can be provided, to ensure the comparability and universality of the model under the standard benchmark (uniform dataset and evaluation index).

[0026] Of course, any product implementing the present application does not necessarily need to achieve all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0028] Figure 1 is the system block diagram of the present application for supporting fast iteration and quantization deployment of the trajectory prediction model;

[0029] Figure 2 is a data sample processing method flowchart of an embodiment of the present application;

[0030] Figure 3 is a model training and quantization flowchart of an embodiment of the present application;

[0031] Figure 4 is a data backfilling-based offline test flowchart of an embodiment of the present application;

[0032] Figure 5 is a real vehicle platform online test flowchart of an embodiment of the present application;

[0033] Figure 6 is a method flowchart of an embodiment of the present application supporting trajectory prediction algorithm rapid iteration and quantization deployment. DETAILED DESCRIPTION

[0034] In order to make the objects, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0035] It should be noted that the diagrams provided in the embodiments of the present application only schematically illustrate the basic concept of the present application, and therefore only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component can be arbitrarily changed in shape, number and proportion, and the layout pattern of the components can also be more complex.

[0036] In the present application, it should also be noted that, if terms such as "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like appear, the indicated orientation or position relationship is based on the orientation or position relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, if the terms "first", "second" appear, they are only for description and distinction purposes and cannot be understood as indicating or implying relative importance.

[0037] In addition, it should also be noted that the features of various embodiments of the present application can be combined or integrated partially or entirely, and as understood by those skilled in the art, can interact and operate in different ways. Each embodiment can be implemented independently of each other, or in an associated relationship.

[0038] Embodiment 1

[0039] AsFigure 1 As shown, the system supporting fast iteration and quantitative deployment of trajectory prediction algorithm includes three main modules: data sample processing module, model training and quantization module, and model deployment and testing module.

[0040] The data sample processing module is used to integrate multi-source heterogeneous data, construct a dynamically expandable sample library through a standardized format conversion mechanism, and generate standardized preprocessed trajectory prediction sample sets. The sample library includes public data sets and real vehicle data sets collected in stages.

[0041] The model training and quantization module is used to use the preprocessed trajectory prediction sample set to perform multi-stage training on the pre-constructed trajectory prediction model of the autonomous vehicle, including first floating point model training, then quantitative perception training. The quantitative perception training includes quantizing and calibrating the parameters of the trajectory prediction model after floating point model training, inserting a pseudo-quantization node in the trajectory prediction model to generate a pseudo-quantization model, fine-tuning the parameters of the pseudo-quantization model, and finally converting to a fixed-point model.

[0042] The model deployment and testing module is used to perform hardware-in-the-loop testing and autonomous vehicle testing on the trajectory prediction model after multi-stage training, and further optimize and iterate the trajectory prediction model according to the test results.

[0043] The data sample processing module integrates multi-source heterogeneous data and constructs a dynamically expandable sample library through a standardized format conversion mechanism. The sample library is based on multiple public road data sets and combined with real vehicle data collected in stages to ensure that the sample library can be dynamically expanded. In addition, the data sample processing module supports generating standardized preprocessed trajectory prediction samples for specific model input requirements, providing data support for subsequent training and evaluation.

[0044] Specifically, the data sample processing module includes a public road dataset, a self-collected real vehicle database, a standardized traffic flow data format conversion unit, a specific model sample generation unit, and a training and test sample database. The public road dataset is used to integrate multiple large trajectory prediction open source datasets to provide various publicly available road and traffic flow data for model training and testing. The self-collected real vehicle database is used to store real vehicle data collected by the vehicle-mounted sensing equipment of the automatic driving real vehicle test platform in real scenes, including the speed, position, and lane line information of the ego vehicle and surrounding vehicles, and can be dynamically expanded according to the incremental data collected in the subsequent real vehicle test running process. The standardized traffic flow data format conversion unit is used to convert the original data in the public road dataset and the self-collected real vehicle database into a unified format, thereby facilitating the model to call data from different sources for training and testing. The specific model sample generation unit is used to generate sample data required by a specific model for training according to the standardized unified format data, meeting the input requirements of different algorithms. The training and test sample database is used to store the data set processed by the specific model sample generation unit, and supports providing the required data for the model in the form of standard input or streaming input during the model training stage.

[0045] Among them, the model training and quantization module uses the training and verification samples from the data sample processing module to perform multi-stage training strategies including self-supervised pre-training, supervised training, transfer learning, and quantization-aware training on the pre-constructed trajectory prediction model, and includes a comprehensive model performance evaluation unit combining qualitative and quantitative methods, and finally outputs the low-bit fixed-point model after parameter quantization for deployment and testing.

[0046] Specifically, the model training and quantization module includes three parts of a floating point model training unit, a quantization perception training unit and a performance evaluation unit. The floating point model training unit is used to train a floating point model (the parameters in the pre-constructed trajectory prediction model are floating point, which can be called a floating point model) based on a public data set and a real vehicle data set, and to provide a model with strong generalization performance in multiple scenarios for the downstream quantization perception training unit. The quantization perception training unit includes quantization parameter calibration, quantization perception training and fixed point model conversion. The quantization parameter calibration is used to calibrate the floating point model parameters, and the quantization parameters are determined through training of a small amount of data set. The quantization perception training inserts a pseudo quantization node in the model, so that the model perceives the impact of quantization during the training process, and fine tunes the model parameters to minimize the loss of accuracy as much as possible when converting to a fixed point model downstream. The fixed point model conversion is used to convert the pseudo quantization model to a fixed point model. The performance evaluation unit includes a validation set index self-test program, a key scene visualization program and an official index test program of a public data set, which is used to monitor the training effect and convergence of the trajectory prediction model during the floating point model training and quantization perception training process, to provide intuitive evaluation and optimization direction of the model prediction ability in complex scenarios, and to ensure the comparability and universality of the model under the standard benchmark (uniform data set and evaluation index).

[0047] Further, based on the floating point model training unit provided by the present application, the representation ability and generalization performance of the model are gradually improved through a multi-stage training strategy. The floating point model training unit includes self-supervised pre-training, public data set supervised training and real vehicle data transfer learning. The self-supervised pre-training pre-trains the trajectory encoder and the map encoder through various mask mechanisms and reconstruction strategies, which is used to efficiently extract the latent semantic features of trajectories and maps, and provides initial weights for the downstream public data set supervised training and real vehicle data transfer learning. The public data set supervised training uses a label-based supervised learning strategy on the public data set, and combines data enhancement techniques (adding noise, random data loss and multi-scene feature combination training strategy, etc.), to provide a trajectory prediction model with strong generalization ability for downstream real vehicle data transfer learning. The real vehicle data transfer learning fully utilizes the diversified driving scene data collected by the real vehicle, reduces the performance deviation caused by the distribution difference between the public data set and the real vehicle data based on weight regularization technology, improves the sampling probability of long-tail scene samples based on sample equalization and diversified processing data resampling strategy, and further improves the performance of the trajectory prediction model on the real vehicle data set.

[0048] The model deployment and testing module is connected with the model training and quantization module, and includes a hardware-in-the-loop testing platform and an automatic driving real vehicle testing platform. The hardware-in-the-loop testing platform performs pre-validation of model inference delay and resource occupation through an embedded twin environment, and the real vehicle testing platform realizes closed-loop scene adaptability testing of the prediction model through a vehicle-mounted computing platform and a line control execution mechanism. In addition, the road scene data collected in the automatic driving real vehicle testing process can be fed back to the data sample processing module for expanding the trajectory prediction training sample library.

[0049] Specifically, the model deployment and testing module includes a hardware-in-the-loop testing platform and an automatic driving real vehicle testing platform. The hardware-in-the-loop testing platform includes a data backfill program, a trajectory prediction program and a model performance evaluation program, and is used for offline testing based on backfill data. The data backfill program runs on the host computer platform, uses a specific software tool to obtain and play back the newly collected perception result data from the self-collected real vehicle database, including motor vehicle and non-motor vehicle perception results, vehicle state information, road structure information, etc. The trajectory prediction program runs on the embedded platform, and the embedded platform and the vehicle-mounted computing platform constitute a twin computing platform. The trajectory prediction program subscribes to the perception result data, pre-processes the perception result data, performs trajectory prediction model inference, post-processes the model inference result, and publishes the model prediction result data. The model performance evaluation program runs on the host computer platform, receives the prediction result sent by the trajectory prediction program, and performs model prediction performance evaluation in combination with the perception result data, including quantitative performance evaluation and qualitative performance evaluation. The quantitative performance evaluation evaluates the prediction model performance through the calculation results of three indexes of prediction accuracy, prediction time domain and prediction delay. The qualitative performance evaluation visualizes the real-time perception result and the prediction result by using a specific rendering tool, thereby evaluating the prediction model performance.

[0050] Specifically, the automatic driving real vehicle test platform is installed with vehicle-mounted sensing equipment, and runs upstream perception programs, trajectory prediction programs and downstream control programs, and also has a line control execution mechanism for online testing based on the real vehicle platform. The vehicle-mounted sensing equipment includes a camera, a laser radar, an inertial measurement unit and a wheel speed meter, etc., which are used to perceive the surrounding environment information and publish raw observation data in real time. The upstream perception program runs on the vehicle-mounted computing platform, subscribes to the raw observation data published by the vehicle-mounted sensing equipment in real time, processes the perception model and publishes the model perception result data. The trajectory prediction program (i.e. predicting the trajectory of the surrounding traffic participants in the driving scene of the autonomous vehicle through a trajectory prediction model) runs on the vehicle-mounted computing platform, receives the perception result data published by the upstream perception program, performs data preprocessing, model inference and inference result post-processing, and publishes the model prediction result data. The downstream control program runs on the vehicle-mounted computing platform, receives the prediction result data published by the trajectory prediction program, performs control algorithm processing, and publishes the control result data. The line control execution mechanism receives the control result sent by the downstream control program and drives the vehicle to run.

[0051] It can be seen that, by performing floating point model training and quantized perception training on the trajectory prediction model of the autonomous vehicle, the trajectory prediction model with floating point weight precision is converted into a fixed point model through multi-stage training, wherein the quantized perception training includes three steps of quantized parameter calibration, pseudo-quantized fine-tuning and fixed point model conversion, ensuring efficient conversion of the floating point precision model to the low-bit fixed point model, while minimizing the precision loss in the quantization process, improving the real vehicle verification efficiency, greatly shortening the iteration period of the autonomous driving trajectory prediction algorithm, and significantly improving the algorithm development efficiency and deployment reliability.

[0052] Embodiment 2

[0053] The method for rapid iteration and quantized deployment of the autonomous vehicle trajectory prediction algorithm of this embodiment is based on the above system, as shown in Figure 6 The method mainly includes the following steps:

[0054] Step 1, data collection and processing: by integrating multi-source heterogeneous data, a dynamically expandable sample library is constructed through a standardized format conversion mechanism, and a standardized preprocessed trajectory prediction sample set is generated; wherein the sample library includes public data sets and real vehicle data sets collected in stages.

[0055] Step 2, multi-stage training of the trajectory prediction model: using the pre-processed trajectory prediction sample set to perform multi-stage training on the pre-constructed trajectory prediction model of the autonomous vehicle, including first performing floating-point model training, and then performing quantization perception training, wherein the quantization perception training includes quantization calibration of the parameters in the trajectory prediction model after floating-point model training, inserting a pseudo-quantization node in the trajectory prediction model to generate a pseudo-quantization model, fine-tuning the parameters of the pseudo-quantization model, and finally converting to a fixed-point model. This step ensures efficient conversion of the floating-point precision model to the low-bit fixed-point model, while minimizing the accuracy loss during quantization.

[0056] Step 3, model deployment and testing: performing hardware-in-the-loop testing and autonomous vehicle testing on the trajectory prediction model after multi-stage training, and further optimizing and iterating the trajectory prediction model based on the test results.

[0057] In Step 1, real-time collection of multi-source heterogeneous data is performed, including public road data sets and self-collected vehicle data, and a standardized format conversion mechanism is used to convert the data into a unified format, constructing a dynamic and expandable traffic segment standard database. The sample preprocessing and dynamic updating process includes model-specific (i.e., the trajectory prediction model of the present application) training and validation sample preprocessing of the traffic segment standard database, including scene information extraction, trajectory information extraction, and map feature extraction, and writing a data loading sampler specific to the training method.

[0058] In Step 2, the model can also be qualitatively and quantitatively evaluated using a validation set self-test program, a key scene visualization program, and an official indicator test program for public data sets, real-time monitoring of the training effect of the model, evaluation of its performance in complex scenarios, and ensuring that the model has strong generalization.

[0059] In Step 3, offline backfill testing is performed on a hardware-in-the-loop testing platform to verify the inference ability of the model on an embedded platform; then online inference testing is performed on an autonomous vehicle testing platform to verify the accuracy and real-time performance of the model in actual driving scenarios.

[0060] The trajectory prediction model algorithm can be further optimized and iterated based on the test results, specifically by optimizing the model structure, training strategy, quantization parameters, etc. based on the model test result feedback, and ensuring that the trajectory prediction model runs stably on different hardware platforms, thereby accelerating the iteration and deployment of the algorithm.

[0061] Embodiment 3

[0062] This embodiment is based on embodiment 1, which realizes the full-process optimization of the automatic driving trajectory prediction model from data collection to deployment landing by constructing a complete algorithm development closed loop. First, in the data collection and processing stage, the system obtains multi-source heterogeneous data from public road data sets (such as Argoverse, nuScenes) and self-collected vehicle databases, converts the original data into a standardized representation form containing timestamps, coordinate systems, traffic participant attributes, and lane topology relationships through a standardized traffic flow data format conversion unit, and constructs a dynamically expandable traffic segment standard database. In the sample preprocessing and dynamic updating stage, based on the standardized database, scene information extraction, trajectory information extraction and map feature extraction operations are performed to generate standardized samples containing historical trajectory sequences, high-precision map features and future trajectory labels, and a sampler supporting streaming data loading is developed to realize the dynamic fusion of real-time incremental data in the training process.

[0063] The floating point model training stage adopts a three-stage progressive strategy: first, a trajectory encoder and a map encoder (i.e. the encoding part of the trajectory prediction model) are constructed through self-supervised pre-training, and general feature representations are extracted through trajectory mask reconstruction and map contrastive learning tasks; then supervised training is performed on public datasets, and a multi-modal loss function is used to optimize prediction accuracy; finally, the pre-trained model is adapted to the real vehicle data domain through transfer learning, and domain adaptation technology is used to reduce data distribution differences. In the fixed-point model quantization stage, the model weight distribution characteristics are analyzed through the quantization parameter calibration module to determine the optimal quantization scale factor of each layer, pseudo quantization operators are inserted during the quantization-aware training process to simulate low-precision calculation, and a progressive fine-tuning strategy is used to balance model accuracy and inference efficiency, finally generating an 8-bit fixed-point model adapted to embedded platforms. In the training process evaluation stage, the ADE / FDE indicators are calculated by running the validation set indicator self-test program, the three-dimensional trajectory rendering analysis of complex scenes such as intersections and dense traffic is performed through the key scene visualization program, and the official benchmark test platform is regularly used for official benchmark test to ensure that the model performance reaches the industry benchmark level.

[0064] In the model deployment and testing stage, an embedded twin environment is constructed using a hardware-in-the-loop test platform, the perception data stream collected by the real vehicle is reproduced through the data backfill program, and the latency and memory occupation indicators of the fixed-point model on the target hardware are verified; in the real vehicle test stage, online inference is performed on the vehicle-mounted computing platform, the prediction results are input into the control algorithm to generate control instructions, and a closed-loop iteration mechanism of scene data collection-model optimization-deployment verification is established. Finally, based on the quantization evaluation report output by the test platform, the model structure, training hyperparameters, and quantization strategy are optimized, and a new version of the model is quickly verified and deployed through an automated pipeline, forming a sustainable algorithm iteration system.

[0065] Figure 2The specific process of the data sample processing module of the system is shown. Specifically, the data to be processed is composed of two parts, namely the public road data set and the self-collected real vehicle database. The public road data set includes nuScenes, nuPlan, Argoverse and other large automatic driving data sets, which are used to provide a large number of sample data for model training; the data in the self-collected real vehicle database is divided into existing real vehicle collection data and new real vehicle collection data, wherein the existing real vehicle collection data is the real driving scene data collected and stored in the historical real vehicle experiment; the new real vehicle collection data is incremental data obtained in the latest real vehicle test running, which is obtained by the camera, laser radar and other vehicle-mounted sensing devices carried by the automatic driving real vehicle test platform in the running process, and is preliminarily processed and optimized after data preprocessing, and is arranged as map data and agent data. The map data mainly includes the attributes and information of map elements such as lane lines and pedestrian crossings; the agent data includes the physical attributes, trajectories, speeds and orientations of the ego vehicle and surrounding traffic participants.

[0066] On the basis of the self-collected real vehicle database and the integrated public road data set, the two are processed by the standardized traffic flow data format conversion unit. The format conversion process is related to the type of data set, and each type of large automatic driving data set in the public road data set and the self-collected real vehicle database is processed by a specific data set format conversion program. Specifically, the format conversion program first reads the scene data from the specified original data set file path, and returns the scene list and scene data instance according to the set number of working threads; then, in the scene data conversion stage, scene information extraction, trajectory information extraction, and map feature extraction are performed in turn. The scene information extraction is mainly to extract all frame data related to the scene from the given scene, including the identifiers, position information, types and speeds of all objects in each frame of the scene, and return these frame data and their corresponding scene information; the trajectory information extraction further processes the trajectory data of all objects based on the extracted scene frame data, and obtains complete trajectory data by removing invalid data, filling missing frames, and interpolating position, speed and orientation; the map feature extraction mainly extracts specified map element information from the corresponding map data according to the scene information, such as drivable lane boundaries, pedestrian crossing boundaries, lane boundaries and lane identifiers. After the scene data conversion, the processed unified format data is saved as a binary file and stored in the specified path. In addition, the system also summarizes all processed scene data and generates an index file, which records the metadata such as the identifier, timestamp and storage path of the processed scene data of each scene, so as to facilitate quick retrieval and access in the model training stage.

[0067] After the original data is converted into unified format data, the specific model sample generation unit will execute different data processing scripts according to the model type, and generate cache files that can be used for training and verification of different models based on the formatted data. First, load the file from the specified path, read the scene data; then, in the specific specific model data processing stage, three operations of map data processing, agent data processing and sample feature extraction are carried out respectively. Among them, the map data processing will perform equidistant interpolation and cutting operation on the map elements in the scene such as lane line, pedestrian crossing, etc., and convert the coordinates to the coordinate system centered on the target agent; the agent data processing will filter the target agent according to the condition mask, and perform coordinate transformation operation on the historical trajectory and future trajectory of the target agent. The processed agent data also includes trajectory validity mask, agent index and other information; sample feature extraction is to evaluate a single sample data, compare the Kalman filter prediction with the real trajectory in the future, and evaluate the accuracy of the Kalman filter prediction as the Kalman difficulty of the sample. In addition, this step will also classify the agent trajectory according to the speed, position, heading angle and other information of the starting point and ending point of the trajectory, and mark the trajectory type attribute of the sample.

[0068] The data processed by the specific model sample generation unit will be stored in the training test sample database, and will be divided into training set, validation set or test set according to the preset functional requirements. The data file name contains scene name, frame index and agent name, so the database not only supports standard input format, but also meets the needs of streaming input. Specifically, the model will load the relevant data of the target agent in the same scene as input in time sequence, so as to constrain the stability by using the historical prediction result, and ensure that the prediction result in the same scene basically remains unchanged.

[0069] Figure 3 The specific details of the model training and quantization module of the system are shown. First, the self-supervised pre-training stage needs to take the mask strategy of trajectory data, randomly mask some time step data, so that the trajectory sequence forms an incomplete historical input in the time dimension. The trajectory encoder completes these missing time steps by time inference. Its loss function is actually

[0070]

[0071] wherein, represents the masked trajectory sequence, is the completed trajectory generated by the model.

[0072] For the mask strategy of the map, in order to make the map encoder better capture the local structure characteristics, some key point features (such as lane center line, intersection, etc.) in the map are randomly masked, and the encoder is required to reconstruct these masked features. Its loss function is actually

[0073]

[0074] where, represents the local feature of the masked map point, is the reconstructed map point feature of the model.

[0075] Through the trajectory points and map point features of the mask part, the model learns to infer the missing information from the context and overall features, so as to obtain an encoder that has an overall understanding of the historical trajectory and map. The overall loss function of the self-supervised pre-training of the encoder is

[0076]

[0077] Then, the supervised training phase on large public datasets is carried out. First, the weights of the trajectory encoder and the map encoder generated in the self-supervised pre-training phase are loaded. These weights have been pre-trained through various masking mechanisms and reconstruction strategies, and can effectively extract the latent semantic features of the trajectory and the map. On this basis, a trajectory prediction network is constructed, in which the outputs of the trajectory encoder and the map encoder are fused through a feature fusion module to generate global context features. Subsequently, through a prediction head based on Laplace distribution, the model models the probability distribution of the future trajectory. The supervised training is based on the following Laplace distribution negative log-likelihood (NLL) loss function:

[0078]

[0079] where, is the real trajectory point, and are the mean and variance predicted by the model respectively, is the number of samples.

[0080] In the process of supervised training, in order to improve the generalization ability of the model, first, the public datasets nuScenes, nuPlan and Argoverse, etc. are unified into the data format required by the model. In this case, Gaussian noise is added to the trajectory point position to simulate the error of the actual sensor, and the formula is as follows:

[0081]

[0082] where, is the mean of 0 and the variance of Gaussian distribution. In addition, part of the time steps in the trajectory data are randomly deleted to enhance the robustness of the model to sensor frame loss scenarios. Finally, we adjust the sampling probability for the long-tail scenarios (such as turning, overtaking, etc.) that are relatively rare in driving data to ensure the diversity and balance of the training data.

[0083] The real vehicle data transfer learning module fine-tunes the weights of the trajectory encoder and the map encoder generated by the public data set supervised training module. First, load the weights of the trajectory encoder and the map encoder generated by the public data set supervised training module as the initial parameters of fine-tuning. Filter the trajectory points in the real vehicle data, including noise filtering (such as Kalman filtering). Based on the processed real vehicle trajectory data, fine-tune the trajectory encoder and the map encoder. To prevent catastrophic forgetting during the transfer learning process, a weight regularization strategy is introduced. This strategy constrains the changes in model weights during the fine-tuning process, ensuring that useful features learned are preserved during the update process, while guiding the model to better adapt to new data.

[0084] The quantization parameter calibration step determines the quantization range of the parameters that need to be quantized in the model (including scaling factor and zero point) through training of a small amount of data set. Through training of a small amount of sample, the range of quantization parameters is adapted to the data characteristics of the current task, so as to ensure that the quantized model can still maintain high accuracy. The goal of calibration is to minimize the error before and after quantization.

[0085] Quantization-aware training (QAT) is to insert a pseudo-quantization node during training, so that the model can "perceive" the impact of quantization on model accuracy during training. Insert pseudo-quantization node: insert pseudo-quantization node in each layer of the model (such as convolutional layer, fully connected layer, etc.). Train using the model with pseudo-quantization node. During training, adjust the model weights and biases through the backpropagation algorithm, so that the model can maintain high accuracy in the quantized environment. QAT considers the impact of quantization on the model during training, so that the trained model better adapts to the quantized hardware deployment.

[0086] The fixed-point model conversion step converts the floating-point model after quantization-aware training into a fixed-point model for efficient operation in embedded hardware or other resource-constrained environments. Quantization model conversion: convert the floating-point model after QAT to a fixed-point model through quantization algorithms. Finally, deploy the converted fixed-point model to the target hardware platform for inference calculation.

[0087] The performance evaluation unit is designed to comprehensively monitor the training effect and convergence of the model during floating-point model training and quantization-aware training, and to provide direction for further optimization. The unit includes three main modules: validation set indicator self-test program, key scene visualization program, and public data set official indicator test program. Through the coordinated work of the three programs, the performance evaluation unit can provide detailed feedback for the training and optimization of the model, and ensure that the model exhibits excellent performance in complex application scenarios.

[0088] The validation set indicator self-test program is designed to monitor the model's performance on the validation set in real time during training, evaluate the training effect and convergence of the model. The key scene visualization program is used to provide an intuitive evaluation of the model's prediction in complex scenarios. The public data set official indicator test program is used to ensure the comparability and universality of the model under standard benchmarks (such as public data sets and evaluation indicators).

[0089] The present application provides a system and method for supporting fast iteration and quantization deployment of trajectory prediction algorithm, which realizes offline testing of trajectory prediction model in hardware-in-the-loop test platform. Through offline testing of backfilling data, the accuracy and stability of the prediction model are evaluated, thereby promoting the iterative optimization of the trajectory prediction algorithm. Figure 4 Three core modules of the hardware-in-the-loop test platform are demonstrated: first, the data backfilling program, which is used to play back the collected open road driving data, simulating the perception process of the autonomous driving system; then the trajectory prediction program, which is based on backfilling data for model inference, predicting the trajectory of traffic participants; finally, the model performance evaluation program, which is used to calculate and evaluate the performance of the trajectory prediction model, including qualitative and quantitative analysis. The above programs work together to realize the verification and optimization of the trajectory prediction model in the autonomous driving system.

[0090] The specific hardware-in-the-loop test steps include:

[0091] (1) Open road driving and real vehicle data collection and saving. To fully evaluate the performance of the trajectory prediction model, the collected traffic scenarios include pedestrians, vehicles, cyclists (including motorcycles and bicycles), and cover various types of traffic scenarios such as intersections and two-way straight roads. The safety officer drives the vehicle to the domestic open road that meets the above scene requirements, starts the recording function of the specific software on the vehicle control platform, collects more than 1000 seconds of perception positioning data, and saves the perception data collected on the open road as a data package.

[0092] (2) Real vehicle data playback. The collected data packets are played back and the perception package data is published using the specific software pre-installed on the notebook computer. The notebook computer is connected to the embedded platform through an Ethernet cable, accesses the embedded platform using the Xshell tool, cross-compiles the subscription and perception package data parsing program, and transmits the executable file to the embedded platform to run the subscription and parsing script, thereby realizing the subscription and parsing of the Protobuf-based perception data.

[0093] (3) Running the prediction model. The generation of the fixed-point model, the cross-compilation of the board-end reasoning program, and the writing of the configuration file are completed in advance, and the executable file and the configuration file are transmitted to the embedded platform. The reasoning script of the prediction model is run to receive the perception data and perform the prediction model reasoning, and the perception and prediction results are saved in real time.

[0094] (4) Sample processing and test sample construction. The saved serialized perception data is processed into prediction true value data, which is processed into test samples together with the prediction results. The total number of samples of the non-motor vehicle trajectory prediction model in the test sample is not less than 1000, of which the number of pedestrian samples is not less than 500 and the number of cyclists is not less than 500.

[0095] (5) Check the calculation index. The test index program is written in advance, and the prediction time domain and prediction accuracy index of the non-motor vehicle are calculated respectively.

[0096] The accuracy of the non-motor vehicle lateral trajectory prediction is evaluated by the lateral average distance error (ADE-x). The specific calculation formula is:

[0097]

[0098] Wherein, N is the number of samples, T is the total prediction step, represents the longitudinal position coordinate of the prediction trajectory in the ith sample at time t, and is the longitudinal true position coordinate at time t in the ith sample.

[0099] The average distance error (Average Distance Error, ADE) is used to evaluate the average distance error between the predicted trajectory and the true trajectory of the non-motor vehicle, and to measure the accuracy of the trajectory prediction algorithm. Here, the Euclidean distance is used to calculate the trajectory point error.

[0100]

[0101] Wherein, N is the number of samples, T is the total prediction step, represents the position of the prediction trajectory in the ith sample at time t, and is the true position at time t in the ith sample.

[0102] Figure 5The overall framework of an automatic driving real vehicle test platform is demonstrated. The automatic driving real vehicle test platform is composed of multiple key components, including a vehicle-mounted sensing device, an upstream perception program, a trajectory prediction program, a downstream regulation and control program, and a drive-by-wire execution mechanism, which work together to complete online testing of a trajectory prediction model, specifically including the following steps:

[0103] (1) The vehicle-mounted sensing device is responsible for perceiving the surrounding environment and publishing raw observation data in real time. It includes a camera, a laser radar, a combination navigation system, and the like. These sensors work together to collect environmental data in real time and publish the raw observation data to the upstream perception program through a vehicle-mounted computing platform.

[0104] (2) The upstream perception program. The upstream perception program runs on the vehicle-mounted computing platform and is responsible for processing data from the vehicle-mounted sensing device and generating output results of the perception model. It subscribes to raw observation data in real time, processes the raw data through various perception algorithms, detects and tracks dynamic and static targets, and publishes the processed perception data in a standard format (such as proto format) for use by the downstream trajectory prediction program.

[0105] (3) The trajectory prediction program. The trajectory prediction program runs on the vehicle-mounted computing platform, receives perception result data from the upstream perception program, and performs trajectory prediction. The main process includes: data preprocessing, coordinate transformation, data normalization, time alignment, and the like on the received perception data; model inference, based on the preprocessed historical trajectory and map data, running the deployed trajectory prediction model to predict the future trajectory of the traffic participant; post-processing of the inference result, including trajectory smoothing, confidence estimation, and the like, to ensure the accuracy and stability of the prediction result; publishing model prediction result data, publishing the predicted trajectory result to the downstream regulation and control program.

[0106] (4) The downstream regulation and control program. The downstream regulation and control program is responsible for processing regulation and control algorithms based on the trajectory prediction result to generate appropriate control instructions. These control instructions will then be sent to the drive-by-wire execution mechanism to guide the operation of the vehicle. Its main functions include: receiving prediction result data: receiving prediction result data from the trajectory prediction program, including the predicted trajectory, position, speed, and the like of the target object. Regulation and control algorithm processing: based on the prediction result, executing the regulation and control algorithm to generate a suitable driving trajectory or control command for the vehicle.

[0107] (5) The drive-by-wire execution mechanism. The drive-by-wire execution mechanism is responsible for receiving the control results published by the downstream regulation and control program and driving the various operations of the vehicle through precise execution mechanisms. The specific functions include: receiving control result data: receiving control commands published by the downstream regulation and control program, such as acceleration, steering angle, brake force, and the like. Executing control commands: through electric control systems (such as throttle, brake, steering control modules), realizing dynamic control of the vehicle.

[0108] The platform can quickly verify the accuracy and real-time performance of trajectory prediction algorithms, providing strong support for performance improvement and iteration of autonomous driving systems. At the same time, through closed-loop control, the safety and stability of the autonomous driving system are ensured.

[0109] This embodiment realizes systematic breakthrough through the coordinated innovation of each module, and the beneficial effects mainly manifest in the following dimensions:

[0110] First, efficient compatibility and closed-loop update of multi-source heterogeneous data. In view of the poor model generalization caused by single data source and heterogeneous format in the prior art, the invention constructs a unified data representation framework through a standardized traffic flow data format conversion unit. This unit converts public datasets such as Argoverse and nuScenes and multi-source heterogeneous data collected from real vehicles into a standardized format containing coordinate systems, lane topologies, and kinematic parameters of traffic participants. This mechanism improves cross-dataset training efficiency and forms a closed-loop ecosystem of "data collection-model training-scenario feedback".

[0111] Second, dual improvement of model training and quantization efficiency and accuracy. Traditional methods have high risk of overfitting and poor hardware adaptability in the model training phase. The invention supports a multi-stage training strategy of self-supervised pre-training-supervised training-transfer learning-quantization-aware training, and uses advanced model training techniques such as Masked Trajectory Modeling in the self-supervised pre-training phase and the introduction of a dynamic curriculum learning mechanism in the supervised training phase. The convergence speed of both floating-point models and fixed-point models has been effectively improved, and the robustness and generalization ability of the model have also been improved.

[0112] Finally, the reliability of model real vehicle deployment testing is improved through a dual verification mechanism of hardware-in-the-loop testing and real vehicle testing. Hardware-in-the-loop testing is based on an embedded twin environment, simulating real hardware resource constraints in the laboratory stage to identify potential deployment risks such as model inference latency anomalies and excessive memory usage, significantly reducing the debugging cost in the real vehicle testing stage. Real vehicle testing verifies the cooperativity of prediction models and control systems in dynamic traffic scenarios through the closed-loop linkage of vehicle-mounted computing platforms and line control execution mechanisms, ensuring that the model output trajectory meets the vehicle dynamics constraints and actual road rules.

[0113] Embodiment 4

[0114] The application further provides a computer readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card memory (for example, an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an App application store, and the like, which stores a computer program, and the program is executed by a processor to realize corresponding functions. The computer readable storage medium of the embodiment is executed by the processor to realize the method for supporting the trajectory prediction model of the autonomous vehicle.

[0115] It should be noted that, according to the needs of implementation, each step / component described in the present application can be split into more steps / components, or two or more steps / components or part of the operation of the steps / components can be combined into a new step / component, so as to realize the purpose of the present application.

[0116] The size of the serial number of each step in the above embodiment does not mean the order of execution, the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0117] It should be understood that those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes shall fall within the protection scope of the appended claims of the present application.

Claims

1. A system that supports rapid iteration and quantitative deployment of trajectory prediction algorithms, characterized by: include: The data sample processing module is used to integrate multi-source heterogeneous data, build a dynamically expandable sample library through a standardized format conversion mechanism, and generate a standardized pre-processed trajectory prediction sample set; The sample library includes public datasets and real-car datasets collected in stages; The model training and quantization module is used to perform multi-stage training on the pre-built autonomous vehicle trajectory prediction model using the pre-processed trajectory prediction sample set, including floating-point model training followed by quantization-aware training. Quantization-aware training includes quantizing and calibrating the parameters in the trajectory prediction model after floating-point model training, inserting pseudo-quantization nodes into the trajectory prediction model to generate a pseudo-quantization model, fine-tuning the parameters of the pseudo-quantization model, and finally converting it into a fixed-point model. Specifically, floating-point model training involves pre-training the vehicle trajectory encoder and map encoder in the trajectory prediction model through a self-supervised learning mechanism to extract the latent semantic features of the trajectory and map, and using a label-based supervised learning strategy on a public dataset, combined with data augmentation technology, to provide a trajectory prediction model with strong generalization ability for transfer learning of the real vehicle dataset. Finally, transfer learning is performed using the real vehicle dataset to train the floating-point parameters in the trajectory prediction model that are adapted to the real vehicle data domain. The model deployment and testing module is used to perform hardware-in-the-loop testing on the trajectory prediction model after multi-stage training to complete pre-verification of the model inference latency and resource usage, and to conduct autonomous driving real-vehicle testing to dynamically expand the sample library to further iterate and optimize the trajectory prediction model based on the test results.

2. The system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms according to claim 1, characterized in that: Hardware-in-the-loop testing specifically uses an embedded twin environment to pre-verify model inference latency and resource occupancy; autonomous driving real-vehicle testing specifically uses the on-board computing platform and wire-controlled actuators to implement closed-loop scenario adaptability testing of the trajectory prediction model. The road scene data collected during the real-vehicle testing process will be fed back to the data sample processing module for dynamic expansion of the sample library.

3. The system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms according to claim 1, characterized in that: Multi-source heterogeneous data includes public road data and self-collected real vehicle data. The data sample processing module generates a public data set by summarizing and integrating the public road data, and integrates the real vehicle data collected in stages in real scenes by on-board sensor equipment into a real vehicle data set.

4. The system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms according to claim 3, characterized in that: On-board sensor equipment is used to perceive the surrounding environment and publish raw observation data in real time. It mainly includes cameras, lidar, inertial measurement units and wheel speed meters.

5. The system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms according to claim 1, characterized in that: The model training and quantization module also includes performance evaluation, which is used to conduct performance evaluation in stages during floating-point model training and quantization-aware training, monitor the model's training effect and convergence during the process, and provide intuitive evaluation and optimization direction of the model's prediction ability in complex scenarios.

6. The system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms according to claim 1, characterized in that: During hardware-in-the-loop testing, a data re-injection program is run to obtain and replay newly collected perception result data from the real vehicle data set, including motor vehicle and non-motor vehicle perception results, vehicle status information, and road structure information; the perception result data is pre-processed on a pre-built twin computing platform, and the trajectory of the pre-processed data is predicted through a multi-stage trained trajectory prediction model. The trajectory prediction results are then post-processed and the trajectory prediction result data is published.

7. The system for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms according to claim 6, characterized in that: During hardware-in-the-loop testing, a model performance evaluation program is also run to evaluate the model prediction performance of the prediction results and perception result data. This includes quantitative performance evaluation and qualitative performance evaluation. The quantitative performance evaluation is calculated based on three indicators: trajectory prediction accuracy, trajectory prediction time domain, and trajectory prediction delay. The qualitative performance evaluation uses specific rendering tools to visualize the real-time perception results and prediction results.

8. A method for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms, characterized in that: The following steps are involved: Integrate multi-source heterogeneous data, build a dynamically expandable sample library through a standardized format conversion mechanism, and generate a standardized pre-processed trajectory prediction sample set; The sample library includes public datasets and real-car datasets collected in stages; The pre-built autonomous vehicle trajectory prediction model is trained in multiple stages using a pre-processed trajectory prediction sample set. This includes first performing floating-point model training and then performing quantization-aware training. Quantization-aware training involves quantizing and calibrating the parameters in the trajectory prediction model after floating-point model training, inserting pseudo-quantization nodes into the trajectory prediction model to generate a pseudo-quantization model, fine-tuning the parameters of the pseudo-quantization model, and finally converting it into a fixed-point model. Specifically, the floating-point model training involves pre-training the vehicle trajectory encoder and map encoder in the trajectory prediction model through a self-supervised learning mechanism to extract the latent semantic features of the trajectory and map. Furthermore, a label-based supervised learning strategy is used on a public dataset, combined with data augmentation technology, to provide a trajectory prediction model with strong generalization capabilities for transfer learning on the real vehicle dataset. Finally, transfer learning is performed using the real vehicle dataset to train the floating-point parameters in the trajectory prediction model that are adapted to the real vehicle data domain. The trajectory prediction model after multi-stage training is subjected to hardware-in-the-loop testing to complete the preliminary verification of model inference latency and resource occupancy, and autonomous driving real vehicle testing is carried out to dynamically expand the sample library to further optimize the trajectory prediction model based on the test results.

9. A computer storage medium, characterized in that A computer program executable by a processor is stored therein, and the computer program executes the method for supporting rapid iteration and quantitative deployment of trajectory prediction algorithms as described in claim 8.

Images