Testing Method for Pedestrian Trajectory Prediction and Emergency Braking in an Autonomous Driving Platform

Abstract

The present invention belongs to the field of data processing, and discloses a test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform; it includes pre-segmenting the real road environment and obtaining the level of road segments according to the calculation of the segmentation coefficient, allocating resources to the segments according to the level and collecting multi-source heterogeneous data; introducing importance optimization in the Transformer model to generate possible prediction trajectories of individuals, using the possible prediction trajectories to construct graph data, introducing an Autoencoder anomaly detection mechanism to screen out high-risk trajectories, and outputting high-risk scenarios; combining the road geometric model with the high-risk scenarios to form a scenario library, loading the scenarios in the simulation platform and conducting autonomous driving tests, and combining the real vehicle test data for cross-verification of virtual and real data to obtain the performance evaluation of the algorithm; realizing pedestrian trajectory prediction and emergency braking during the autonomous driving process.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and more specifically, to a testing method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform. Background Art

[0002] The rapid development of autonomous driving technology has brought about revolutionary changes in transportation modes. Pedestrian trajectory prediction and emergency braking, as two key technologies, play a vital role in ensuring the safe and reliable operation of autonomous vehicles.

[0003] During testing, pedestrian behavior is highly uncertain and dynamic, especially in complex traffic scenarios (such as sudden crossings and group avoidance). Traditional prediction methods struggle to accurately capture pedestrian intentions and interactions. In real-world road environments, high-risk scenarios (such as accidents or near-accidents) are often sparse and difficult to accurately identify, and traditional methods can overlook potential risks. The complexity of real-world road environments leads to high data collection and processing costs, and uneven resource allocation can lead to insufficient data in key areas.

[0004] In view of this, the present invention proposes a testing method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for testing pedestrian trajectory prediction and emergency braking in an autonomous driving platform, comprising: step S1: pre-segmenting a real road environment and obtaining the levels of road segments based on the calculation of segmentation coefficients, allocating resources to the segments based on the levels, and collecting multi-source heterogeneous data;

[0006] Step S2: Importance optimization is introduced into the Transformer model to generate possible predicted trajectories of individuals. Graph data is constructed using the possible predicted trajectories. The GNN model processes the graph data to obtain the interaction behavior trajectories after individual association. The Autoencoder anomaly detection mechanism is introduced to screen out high-risk trajectories and output high-risk scenarios.

[0007] Step S3: Combine the road geometry model with high-risk scenarios to form a scenario library, load the scenarios into the simulation platform and conduct autonomous driving tests. Combine the real vehicle test data with virtual and real data for cross-validation to obtain performance evaluation of the algorithm.

[0008] Step S4: Arrange the test process and results on a visual interface.

[0009] Preferably, the method of pre-segmenting the real road environment, obtaining the level of road segments according to the calculation of the segmentation coefficient, allocating resources to the segments according to the level, and collecting multi-source heterogeneous data includes:

[0010] The multi-source heterogeneous data includes pedestrian trajectory data, traffic flow data, environmental data, and scene data, and the scene data includes accident occurrence record data and approximate accident data;

[0011] Pre-divide the real road into segments, and calculate the segmentation coefficient for each segment , where is the total number of influencing factors, is the normalized value of the th influencing factor within the th cycle time, and the range is , is the mapping function of , is the dynamic weight of the th influencing factor within the th cycle time;

[0012] Preset the first boundary threshold and the second boundary threshold, and the first boundary threshold is greater than the second boundary threshold. If the segmentation coefficient is greater than the first boundary threshold, it is determined as a high-risk section, and the acquisition devices are deployed densely. If the segmentation coefficient falls within the interval between the first boundary threshold and the second boundary threshold, it is determined as a medium-risk section, and the acquisition devices are deployed at medium density. If the segmentation coefficient is less than the second boundary threshold, it is determined as a low-risk section, and the acquisition devices are deployed at low density;

[0013] Calculate the segmentation coefficient of the th cycle time and make a level judgment. According to the judgment output result, dynamically deploy the acquisition devices for the segments in the th cycle time, and collect multi-source heterogeneous data in real time through the acquisition devices.

[0014] Preferably, the high-risk scenario and the acquisition method include:

[0015] Perform cleaning and interpolation preprocessing on the collected multi-source heterogeneous data;

[0016] For each segment, extract the multi-source heterogeneous data within G_U cycle times, and obtain the complete input vector through feature embedding processing;

[0017] Use the Transformer model to predict the possible behaviors of pedestrian individuals and generate the possible prediction trajectories of individuals;

[0018] Use the GNN model to associate the interaction relationships between pedestrians, correct the possible prediction trajectories of individuals, and generate the associated interaction behavior trajectories;

[0019] Introduce the Autoencoder anomaly detection mechanism, remove the interaction behavior trajectories with low occurrence probabilities, retain the high-risk trajectories with high occurrence probabilities, and output high-risk scenarios.

[0020] Preferably, the complete input vector is obtained through feature embedding processing. The method includes:

[0021] Extract the historical trajectory data, environmental context data, and intention data of each pedestrian to form raw data;

[0022] Clean, smooth, and standardize the raw data;

[0023] Independently embed the data of each modality in the raw data, including: for the historical trajectory data of pedestrians, use MLP or RNN to extract temporal features from the trajectory sequence; for the environmental context data, use MLP to embed the environmental data, and for the intention data of pedestrians, use MLP to embed the intention data. Normalize all the embedded features;

[0024] Calculate the mean and variance of the embedded features of each modality, make it close to the standard normal distribution by adjusting the activation function of the embedding layer, calculate the correlation matrix of the embedded features of each modality, and if the correlation exceeds the predefined judgment threshold, use PCA for dimensionality reduction;

[0025] Take the embedded features of the historical trajectory data as the main features, and the embedded features of the environmental context data and intention data as the auxiliary features. Use the Cross-Attention mechanism to capture the interaction between modalities, and perform weighted fusion on the interacted features to obtain the complete input vector.

[0026] Preferably, the method of using the Transformer model to predict the possible behaviors of individual pedestrians and generate the possible prediction trajectories of individuals includes:

[0027] Set the input of the Transformer encoder in the Transformer model as the input vector, and the output as the encoded features, denoted as encoded features. The input of the conditional variational autoencoder in the Transformer model is the encoded features, the embedded features of the environmental context data, and the embedded features of the intention data. The output is the parameters of the latent distribution. Obtain the Gaussian expression form of the latent distribution according to the parameters, and sample the latent variables from the latent distribution, and then generate F_O possible future trajectories;

[0028] During the process of sampling the latent variables from the latent distribution, introduce importance optimization, that is, calculate the importance weights of each latent variable, , is the encoded feature, is the embedded feature of the intention data, is the th trajectory generated probability, is the prior distribution of the latent variable, represents the th latent variable, is the index of the trajectory, is the index of the latent variable;

[0029] Arrange the importance weights of each latent variable in descending order, select the top A_O latent variables, denote them as the possible predicted trajectories, and then output the possible predicted trajectories;

[0030] Set trajectory constraints, including diversity constraints, physical constraints, and intention constraints. The diversity constraint is to calculate the minimum Euclidean distance between any two possible predicted trajectories at all time steps and maximize the minimum Euclidean distance. The physical constraint is that the acceleration and speed of the pedestrian are both less than the maximum acceleration and maximum speed. The intention constraint is to minimize the Euclidean distance between the trajectory and the target position.

[0031] Preferably, the method of using the GNN model to associate the interaction relationships between pedestrians, correct the possible predicted trajectories of individuals, and generate the associated interaction behavior trajectories includes:

[0032] Extract the distances between pedestrians and the distances between pedestrians and vehicles or roadblocks from multi-source data, denote them as interaction data, and combine the interaction data with the possible predicted trajectories of individuals as the input data of the GNN model;

[0033] Construct a blank graph, represent each pedestrian as a node in the graph. The node features include the possible predicted trajectories of individuals, environmental context data, and intention data. Represent the distances between pedestrians as edges in the graph, and the edge features include distance information. Also represent the distances between pedestrians and vehicles or roadblocks as edges, and the edge features include the positions and speeds of the vehicles or roadblocks to form graph data;

[0034] Use the graph neural network GNN to process the graph data. The GNN updates the features of each node through multi-layer graph convolution or graph attention mechanism. During the update process, introduce the social force model constraint and output the updated node features;

[0035] Map the updated node features back to the trajectory space and output the interaction behavior trajectories of each pedestrian.

[0036] Preferably, the method of introducing the Autoencoder anomaly detection mechanism, removing the interaction behavior trajectories with low possibility of occurrence, retaining the high-risk trajectories with high possibility of occurrence, and outputting the high-risk scenarios includes:

[0037] Use an Autoencoder time series anomaly detection model, with the input being the interaction behavior trajectories of each pedestrian and the output being the anomaly scores of each interaction behavior trajectory;

[0038] Based on the anomaly threshold and trajectory features, define high-risk behavior types as including sudden crossing, direction change, and group avoidance. Set the anomaly threshold and trajectory features for each high-risk type. Among them, the trajectory feature of sudden crossing is that the pedestrian trajectory crosses the lane line, the trajectory feature of direction change is that the speed direction of the pedestrian trajectory undergoes a sudden change, and the trajectory feature of group avoidance is that the trajectories of more than one pedestrian change direction simultaneously;

[0039] For the interaction behavior trajectory of each pedestrian, calculate the difference between the anomaly score and the anomaly threshold, and use the Sigmoid activation function to convert it into a probability value in combination with the trajectory features, and record the probability value as the high-risk probability;

[0040] Set a risk threshold. When the high-risk probability exceeds the risk threshold, the interaction behavior trajectory is marked as a high-risk behavior. Extract the environmental context data at the time of the high-risk behavior, combine it with the high-risk behavior to form a high-risk scenario, and output the high-risk scenario.

[0041] Preferably, the method of combining the road geometry model with the high-risk scenario to form a scenario library, loading the scenario in the simulation platform and conducting autonomous driving tests, and combining the real vehicle test data for virtual-real data cross-validation to obtain the performance evaluation of the algorithm includes:

[0042] Based on high-precision maps and sensor data, construct a road geometry model, use weather simulation tools to add weather conditions, generate traffic flow through SUMO, and import it into CARLA to simulate the road traffic environment;

[0043] Import the high-risk scenario into the simulated road traffic environment to form a scenario library, and store the scenario library in a standard format;

[0044] Load the scenario library in the simulation platform, load a test vehicle in the scenario library, arrange an autonomous driving algorithm on the test vehicle and run it;

[0045] Record the test data and conduct virtual-real data cross-validation. The performance evaluation of the algorithm includes trajectory error and performance index. The trajectory error is the absolute difference between the interaction behavior trajectory of each pedestrian predicted by the autonomous driving algorithm on the test vehicle and the high-risk scenario in the scenario library, and the method for obtaining the performance index is to accumulate the braking distance, stopping distance, and offset of the vehicle;

[0046] Output the performance evaluation of the algorithm.

[0047] Preferably, the method of arranging the test process and results on the visualization interface includes:

[0048] Display the operation of the simulation platform on the visualization interface;

[0049] Parallelly set R_P test vehicles, run the R_P test vehicles together in the scenario library, and conduct tests for S_F time periods;

[0050] Plot the performance evaluation of the algorithm on the R_P test vehicles over S_F time periods, and arrange the chart on the visualization interface for intuitive observation by users.

[0051] Preferably, the visualization interface includes the end page of a computer or the display interface of a mobile device, and interacts with the user through the interface.

[0052] Technical effects and advantages of the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform of the present invention:

[0053] 1. By collecting pedestrian trajectories, traffic flow, environmental and scenario data (including accident and near-accident data), the fusion of multi-modal data is achieved. Through feature embedding and the Cross-Attention mechanism, the interaction relationships between modalities are captured, enhancing the integrity and expressive ability of the input vector. Using the Transformer model combined with the conditional variational autoencoder, multiple possible future trajectories are generated, and high-likelihood trajectories are screened through importance optimization, improving the prediction accuracy. Trajectory constraints (diversity, physical and intention constraints) are introduced to ensure that the predicted trajectories conform to the physical laws and behavioral logics of the real scenario.

[0054] 2. Using the GNN model to construct graph data, combined with the social force model constraint, the interaction relationships between pedestrians are captured, individual trajectories are corrected, and the robustness of the prediction results is improved. Through graph convolution or graph attention mechanisms, the ability to update node features is enhanced to adapt to dynamic interactions in complex scenarios. The Autoencoder anomaly detection mechanism is introduced, combined with the characteristics of high-risk behavior types (such as sudden crossing, direction change, group avoidance), to remove low-likelihood trajectories and retain high-risk trajectories. By calculating the anomaly score and high-risk probability, the risk level is quantified to ensure the accuracy and practicality of high-risk scenarios.

[0055] 3. Through the dynamic calculation of the segmentation coefficient, small road segments are divided into high, medium, and low risk levels, and the density of acquisition devices is adjusted according to the levels, optimizing the data acquisition efficiency. Dynamically deploy acquisition devices to adapt to changes in the road environment, improving the real-time and pertinence of data acquisition. Combining high-precision maps, weather simulation, and traffic flow simulation, a realistic simulation environment is constructed, and high-risk scenarios are loaded through the scenario library, improving the comprehensiveness of the test. Through cross-validation of virtual and real data, combined with trajectory error and performance indices (braking distance, stopping distance, offset), the performance of the autonomous driving algorithm is comprehensively evaluated.

[0056] This solution solves the key problems in pedestrian trajectory prediction and emergency braking tests through multi-model collaboration, dynamic resource allocation, high-risk scenario screening, and virtual-real combination testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic structural diagram of the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform of the present invention;

[0058] Figure 2 It is a schematic step diagram of the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Embodiment 1

[0060] Please refer to Figure 1 and Figure 2 As shown, the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform of this embodiment includes:

[0061] During the test, pedestrian behavior has a high degree of uncertainty and dynamics, especially in complex traffic scenarios (such as sudden crossing, group avoidance, etc.). Traditional prediction methods are difficult to accurately capture pedestrian intentions and interaction behaviors. In the real road environment, high-risk scenarios (such as accidents or near-accidents) are often sparse and difficult to accurately identify, and traditional methods may ignore potential risks.

[0062] Step S1: Pre-segment the real road environment, obtain the level of the road segment according to the calculation of the segmentation coefficient, allocate resources to the segment according to the level, and collect multi-source heterogeneous data;

[0063] The method of pre-segmenting the real road environment, obtaining the level of the road segment according to the calculation of the segmentation coefficient, allocating resources to the segment according to the level, and collecting multi-source heterogeneous data includes:

[0064] Multi-source heterogeneous data includes pedestrian trajectory data (such as behavioral data like the position, speed, acceleration, and sudden emergence of pedestrians), traffic flow data (such as vehicle density, speed distribution, etc.), environmental data (such as natural environmental impact data records affecting vehicle driving like weather, light, road surface conditions, etc.), and scenario data. Scenario data includes accident occurrence record data (such as accident record data like vehicle rollover, hitting pedestrians, etc.) and near-accident data (such as event record data like pedestrians suddenly crossing, vehicles making emergency brakes, sharp turns, etc. that occur but do not cause accidents).

[0065] In a real road environment, the setting of the segmentation coefficient is a key issue. The purpose is to divide the road into different regions in advance by managers according to experience or requirements based on factors such as road complexity, accident occurrence probability, traffic flow, etc., so as to reasonably allocate data collection resources and focus of attention according to the segmentation coefficient. The setting of the segmentation coefficient needs to comprehensively consider various influencing factors, and divide the real road into small segments in advance, and calculate the segmentation coefficient for each small segment , where is the total number of influencing factors, is the th normalized value of the th influencing factor in the th cycle time, with a range of , is the mapping function for to capture the non-linear influence of factors, is the dynamic weight of the th influencing factor in the th cycle time, satisfying

[0066] , reflecting the change in the importance of factors in different scenarios; the segmentation coefficient is a comprehensive indicator used to measure the complexity and risk level of a certain section of the road. The higher the value, the more complex or risky the section is, and finer-grained segmentation and higher resource investment (such as sensor density, data collection frequency) are required.

[0066] Suppose there are 5 influencing factors: road curvature , accident frequency , traffic flow density , near-accident frequency , environmental complexity , is used to capture the non-linear influence of each factor on segmentation.

[0067] For example, the accident frequency has little impact on risk at low frequencies, but the risk increases sharply at high frequencies. Therefore, an exponential function is used to transform the accident frequency and the near-accident frequency , such as , ; The Sigmoid function is suitable for scenarios where the impact changes smoothly within a certain range, so it can be used for calculating road curvature. ; The power function is suitable for scenarios where the impact shows a power-law growth, so it can be used for calculating traffic flow density. ; The piecewise function is suitable for scenarios where the impact has different behaviors in different intervals, such as environmental complexity, so it can be used for calculating environmental complexity. , where environmental complexity can directly obtain the numerical value of environmental complexity using machine learning models (such as random forest, gradient boosting tree, neural network). The input of the model can be the collected environmental data.

[0068] Dynamic weight is dynamically adjusted according to the real-time scenario to adapt to changes in traffic flow, weather, etc. The specific method is based on machine learning-based adjustment. For example, use online learning algorithms (such as reinforcement learning) to dynamically update the weight. The goal of reinforcement learning is set to maximize the coverage rate of high-risk sections. By dynamically adjusting the weight, the segmentation coefficient can adapt to real-time scenario changes. For example, on rainy days, the weight of environmental complexity increases, making the segmentation coefficient pay more attention to the impact of weather and road surface conditions. During the traffic peak period, the weight of traffic flow density increases, making the segmentation coefficient pay more attention to the risks of congested sections. The dynamic weight adjustment improves the flexibility and real-time performance of the model, and can better cope with the dynamically changing road environment, such as sudden weather changes and sudden traffic accidents.

[0069] Through the segmentation coefficient, high-risk sections (such as accident-prone sections, congested sections) can be more accurately identified, thereby guiding finer-grained road segmentation and resource allocation. And it can adapt to environmental changes, improving the pertinence of data collection and the effectiveness of monitoring. The combination of the non-linear model and dynamic weight makes the segmentation coefficient more robust and generalization ability, suitable for different road types (such as urban roads, rural roads, highways).

[0070] Preset the first boundary threshold and the second boundary threshold, and the first boundary threshold is greater than the second boundary threshold. If the segmentation coefficient is greater than the first boundary threshold, it is determined as a high-risk section, and the acquisition equipment is deployed densely. If the segmentation coefficient belongs to the interval between the first boundary threshold and the second boundary threshold, it is determined as a medium-risk section, and the acquisition equipment is deployed moderately. If the segmentation coefficient is less than the second boundary threshold, it is determined as a low-risk section, and the acquisition equipment is deployed sparsely.

[0071] Calculate the segmentation coefficient of the th cycle time and make a grade judgment. According to the judgment output result, dynamically deploy the acquisition equipment for the small section in the th cycle time, and collect multi-source heterogeneous data in real time through the acquisition equipment.

[0072] In the above description, it is assumed that the predefined first boundary threshold is 0.3 and the second boundary threshold is 0.6, and there are 30 available acquisition devices. For high-risk road sections, 15 can be deployed to comprehensively monitor the road environment; for medium-risk road sections, 10 can be deployed to observe key main positions; for low-risk road sections, 5 can be deployed. The acquisition devices are those used to obtain multi-source heterogeneous data, such as sensors like lidar, cameras, millimeter-wave radars, GPS, etc. The lidar collects point cloud data for target detection and trajectory tracking, the cameras collect image data for target recognition and scene segmentation, the millimeter-wave radar collects the distance, speed, and angle of targets, and the GPS / IMU collects vehicle positioning and motion states. These sensors can be worn on movable devices (such as drones or robotic dogs) to facilitate rapid mobile scheduling and achieve adaptive adjustment of resources and maximized utilization of benefits.

[0073] Step S2: Introduce importance optimization in the Transformer model to generate the possible predicted trajectories of individuals, use the possible predicted trajectories to construct graph data, the GNN model obtains the interactive behavior trajectories after individual association by processing the graph data, introduce the Autoencoder anomaly detection mechanism to filter out high-risk trajectories, and output high-risk scenarios;

[0074] High-risk scenarios and acquisition methods include:

[0075] Perform cleaning and interpolation preprocessing on the collected multi-source heterogeneous data; cleaning is used to remove noise and outliers in the multi-source heterogeneous data, and median filtering is used for processing. Interpolation is used to interpolate missing data, and linear interpolation or spline interpolation methods can be adopted.

[0076] Input data: For each small segment, extract the multi-source heterogeneous data within G_U cycle times, and obtain the complete input vector through feature embedding processing;

[0077] The method to obtain the complete input vector through feature embedding processing includes:

[0078] Extract the historical trajectory data of each pedestrian (including position, speed, and acceleration within the historical cycle time), environmental context data (such as road type, obstacle distribution, and weather conditions), and pedestrian intention data (such as inferring the possible moving direction by analyzing the pedestrian's head orientation or gesture through image data) to form the original data;

[0079] Perform cleaning, smoothing, and standardization processing on the original data to reduce the influence of noise and outliers;

[0080] For example, for the historical trajectory data of pedestrians, use Kalman filtering or moving average filtering to smooth the position, speed, and acceleration data, reduce sensor noise, use statistical methods (such as Z-score) or density-based anomaly detection (such as DBSCAN) to identify and remove outliers, and standardize the trajectory data (zero mean, unit variance); for environmental context data, perform one-hot encoding or label encoding on discrete variables (such as road type, weather condition), standardize continuous variables such as obstacle distribution, and use interpolation (such as linear interpolation) or mean / median to fill in missing environmental data; for pedestrian intention data, use a pre-trained object detection model (such as YOLOv5 or Faster R-CNN) to extract the head orientation and gesture features of pedestrians, post-process the extracted features (such as smoothing or threshold filtering) to reduce detection errors, and standardize the intention features.

[0081] Perform independent embedding on the data of each modality in the original data to generate high-quality feature representations, including: for the historical trajectory data of pedestrians, use MLP or RNN (such as LSTM, GRU) to extract temporal features from the trajectory sequence; for environmental context data, use MLP to embed environmental data (such as road type, obstacle distribution, weather), and for pedestrian intention data, use MLP to embed intention data (such as head orientation, gesture). Normalize all the embedded features; calculate the mean and variance of the embedded features of each modality, and make them close to the standard normal distribution by adjusting the activation function of the embedding layer (such as using ReLU or Tanh). Calculate the correlation matrix of the embedded features of each modality, check for redundancy, and if the correlation exceeds a predefined judgment threshold (such as (correlation |r| > judgment threshold 0.8)), then use PCA for dimensionality reduction (the embedding layer structure can also be adjusted);

[0082] Feature distribution analysis ensures the numerical stability of the embedded features. Correlation analysis reduces feature redundancy.

[0083] Use the embedded features of the historical trajectory data as the main features, and the embedded features of the environmental context data and the intention data as auxiliary features. Use the Cross-Attention mechanism to capture the interactions between modalities, and perform weighted fusion on the interacted features to obtain a complete input vector. The fusion based on Cross-Attention optimizes the feature concatenation method, avoiding information redundancy or dimensional explosion caused by direct concatenation.

[0084] Define the Cross-Attention mechanism , calculate the interaction features between the trajectory and the environment , similarly, calculate the interaction features between the trajectory and the intention , obtain the input vector after weighted fusion ;

[0085] Among them, is the query vector, is the key vector, is the value vector, is the dimension of the key vector, is the embedded feature of the historical trajectory data, is the embedded feature of the environmental context data, , and are learnable weight matrices, and are learnable weights, normalized by softmax.

[0086] Use the Transformer model to predict the (multiple) possible behaviors of individual pedestrians and generate the possible prediction trajectories of the individual;

[0087] Use the Transformer model to predict the possible behaviors of individual pedestrians and generate the possible prediction trajectories of the individual. The methods include:

[0088] Set the input of the Transformer encoder in the Transformer model as the input vector. The Transformer uses the multi-head self-attention mechanism to focus on the important information at different time steps in the trajectory, such as speed mutation or acceleration change, and capture the temporal dependence relationship in the trajectory data. The output is the encoded feature, denoted as the encoded feature, which is used to represent the latent pattern of the pedestrian trajectory. The input of the encoder in the conditional variational autoencoder (CVAE) in the Transformer model is the encoded feature, the embedded feature of the environmental context data, and the embedded feature of the intention data. The output is the parameters of the latent distribution. According to the parameters, the Gaussian expression form of the latent distribution is obtained, that is, the Gaussian distribution, which can also be said to be the expression of the latent distribution. Sample the latent variables from the latent distribution, and input the latent variables, the encoded feature, the embedded feature of the environmental context data, and the embedded feature of the intention data into the decoder together. Use the MLP to initialize the trajectory sequence and use the Transformer decoder to iteratively generate the future trajectory, and then generate F_O possible future trajectories;

[0089] During the process of sampling the latent variables from the latent distribution, introduce importance optimization, that is, calculate the importance weight of each latent variable , is the encoded feature, is the embedded feature of the intention data, is the th trajectory The generated probability, is the prior distribution of the latent variable, denotes the th latent variable, is the index of the trajectory, is the index of the latent variable;

[0090] Arrange the importance weights of each latent variable in descending order, select the top A_O latent variables, denoted as the possible predicted trajectories, and then output the possible predicted trajectories;

[0091] Reduce the number of samplings through importance sampling and improve the computational efficiency. Use simulation tools (such as CARLA, SUMO) to verify the generated trajectories. Compare the similarity between the simulated trajectories and the real trajectories,

[0092] Set trajectory constraints, including diversity constraints, physical constraints, and intention constraints. The diversity constraint is to calculate the minimum Euclidean distance between any two possible predicted trajectories at all time steps and maximize the minimum Euclidean distance to increase diversity. The physical constraint is that the acceleration and speed of the pedestrian are both less than the maximum acceleration and maximum speed. The intention constraint is to minimize the Euclidean distance between the trajectory and the target position. Encourage the trajectory to move towards the target direction. The target position is the destination that the pedestrian wants to reach.

[0093] CVAE first maps the encoded features to the latent space to generate the distribution of latent variables; then samples from the latent distribution, combines the environmental context and intention data, and generates multiple possible future trajectories. To enhance the diversity of the trajectories, introduce diversity constraints to encourage the generated trajectories to cover different behavioral patterns, such as going straight, crossing, or changing direction. To ensure the authenticity of the trajectories, introduce behavioral constraints, such as restricting the acceleration of the trajectory within a reasonable range, or ensuring that the trajectory moves towards the target direction. The output can specifically be that multiple possible future trajectories are generated for each pedestrian, such as 5 trajectories, and each trajectory includes the position coordinates within a future period of time, such as the position changes within the next 5 seconds.

[0094] In summary, the advantages of the above process are enhanced diversity (CVAE etc. ensure that the generated trajectories cover different behavioral patterns), improved accuracy (Transformer decoder, behavioral constraints, and intention embedding improve the prediction accuracy), enhanced computational efficiency (importance sampling reduces the number of samplings and optimizes the computational performance), and behavioral constraints (physical constraints and intention constraints ensure that the trajectories conform to real-world laws and enhance practicality).

[0095] Use the GNN model to associate the interaction relationships between pedestrians, correct the possible predicted trajectories of individuals, and generate the associated interactive behavior trajectories;

[0096] Associate the interaction relationships between pedestrians using a GNN model, correct the possible predicted trajectories of individuals, and generate the interaction behavior trajectories after association. The method includes:

[0097] Extract the distances between pedestrians and the distances between pedestrians and vehicles or roadblocks from multi-source data, record them as interaction data, and combine the interaction data with the possible predicted trajectories of individuals as the input data of the GNN model;

[0098] Construct a blank graph, represent each pedestrian as a node in the graph, the node features include the possible predicted trajectories of individuals, environmental context data, and intention data, represent the distances between pedestrians as edges in the graph, the edge features include distance information, and also represent the distances between pedestrians and vehicles or roadblocks as edges, the edge features include the positions and speeds of vehicles or roadblocks, to form graph data;

[0099] Use the graph neural network GNN to process the graph data and capture the interaction relationships between pedestrians. The GNN updates the features of each node through multi-layer graph convolution or graph attention mechanism. During the update process, introduce the social force model constraint. To ensure that the corrected trajectories conform to physical laws, the GNN needs to consider the attraction and repulsion forces between pedestrians. For example, when two pedestrians are too close, avoidance behavior may occur. The GNN also considers the interaction between pedestrians and obstacles. For example, when a pedestrian approaches a vehicle, the moving direction may change. Consider the interaction forces and environmental forces between pedestrians, such as avoiding trajectory overlap with obstacles. The GNN outputs the updated node features;

[0100] Map the updated node features back to the trajectory space, correct the individual trajectory prediction results, generate the trajectories after interaction, record them as interaction behavior trajectories (behavior features after interaction), and output the interaction behavior trajectories of each pedestrian. For example, 5 trajectories, each trajectory includes the position coordinates in the future for a period of time, reflecting the influence of group interaction;

[0101] The social force model is a model used to simulate and analyze crowd behavior, and it has been widely applied in fields such as crowd simulation, traffic flow analysis, and building design.

[0102] Introduce the Autoencoder anomaly detection mechanism to remove the interaction behavior trajectories with low occurrence probabilities, retain the high-risk trajectories with high occurrence probabilities, and output high-risk scenarios.

[0103] The method of introducing the Autoencoder anomaly detection mechanism to remove the interaction behavior trajectories with low occurrence probabilities, retain the high-risk trajectories with high occurrence probabilities, and output high-risk scenarios includes:

[0104] Use the Autoencoder time-series anomaly detection model to analyze the trajectories after interaction, identify high-risk behaviors. The input is the interaction behavior trajectory of each pedestrian, and the output is the anomaly score of each interaction behavior trajectory. The specific principle is to compress the trajectory into low-dimensional features and then attempt to reconstruct the original trajectory. By comparing the differences between the reconstructed trajectory and the original trajectory, the anomaly score is calculated. The higher the anomaly score, the more likely the trajectory contains abnormal behaviors, such as sudden crossing or group avoidance.

[0105] Based on the anomaly threshold and trajectory features, define high-risk behavior types as including sudden crossing, direction change, and group avoidance. Set the anomaly threshold and trajectory features for each high-risk type. Among them, the trajectory feature of sudden crossing is that the pedestrian trajectory intersects with the lane line, the trajectory feature of direction change is that the speed direction of the pedestrian trajectory suddenly changes, and the trajectory feature of group avoidance is that the trajectories of more than one pedestrian change direction simultaneously.

[0106] The anomaly threshold for each high-risk type can be determined through historical data statistics, such as analyzing approximate accident data.

[0107] For the interaction behavior trajectory of each pedestrian, calculate the difference between the anomaly score and the anomaly threshold, and use the Sigmoid activation function to convert it into a probability value in combination with the trajectory features. Denote the probability value as the high-risk probability. For example, 。

[0108] Set the risk threshold. When the high-risk probability exceeds the risk threshold (e.g., 0.5), the interaction behavior trajectory is marked as a high-risk behavior. Extract the environmental context data at the time of the high-risk behavior and combine it to form a high-risk scenario, and output the high-risk scenario. For example, the high-risk scenario includes the trajectory after the interaction of pedestrians (e.g., position coordinates within the next 5 seconds), high-risk behavior types (e.g., sudden crossing or group avoidance), and environmental context data (e.g., road type, obstacle distribution, and weather conditions).

[0109] In the above process, first, train the Transformer model alone to optimize the accuracy and diversity of individual trajectory prediction. Then, train the GNN model alone to optimize the correction effect of the trajectory after interaction. Finally, jointly train the Transformer, GNN, and anomaly detection models to optimize the overall performance.

[0110] The optimization objectives are to ensure accurate individual trajectory prediction, covering a variety of behavior patterns, to ensure that the trajectory after interaction reflects the group interaction law and avoid unreasonable trajectories, and to ensure accurate high-risk behavior detection, reducing false positives and false negatives.

[0111] Divide the data into a training set, a validation set, and a test set. For example, 80% is used for training, 10% for validation, and 10% for testing. The training set is set according to the input and output of the model. The validation set is used to adjust the model parameters, and the test set is used to evaluate the model performance.

[0112] Use the gradient descent method to gradually optimize the model to ensure its convergence. During the training process, regularly check the performance of the model on the validation set, adjust the learning rate or other parameters to avoid overfitting. In addition, for the case where there are few samples of high-risk behaviors, data augmentation or weighting methods can be used to increase the model's attention to high-risk behaviors.

[0113] The evaluation metrics can be in the following aspects:

[0114] 1. Trajectory prediction accuracy: Evaluate the average distance error between the predicted trajectory and the real trajectory, such as the average deviation between the predicted position and the actual position. Evaluate the final position error, such as the deviation between the end point of the predicted trajectory and the actual end point.

[0115] 2. High-risk behavior detection performance: Evaluate the precision of high-risk behavior detection, that is, the proportion of correctly identified high-risk behaviors among all identified high-risk behaviors. Evaluate the recall rate, that is, the proportion of correctly identified high-risk behaviors among all actual high-risk behaviors. Combine the precision and recall rate to calculate a comprehensive performance metric, such as the F1 score.

[0116] Step S3: Combine the road geometry model with high-risk scenarios to form a scenario library, load the scenarios in the simulation platform and conduct autonomous driving tests, and perform cross-validation of virtual and real data in combination with real vehicle test data to obtain the performance evaluation of the algorithm.

[0117] Combine the road geometry model with high-risk scenarios to form a scenario library, load the scenarios in the simulation platform and conduct autonomous driving tests, and perform cross-validation of virtual and real data in combination with real vehicle test data to obtain the performance evaluation of the algorithm. The methods include:

[0118] Based on high-precision maps and sensor data, construct a road geometry model (generally including lane line, intersection, and obstacle distribution models, and obstacles can be vehicles, roadblocks). The high-precision maps and sensor data can be obtained in advance or locally extracted from multi-source heterogeneous data. Use weather simulation tools to add weather conditions, such as harsh conditions like rain, snow, haze, and night. Generate traffic flow through SUMO and import it into CARLA to simulate the road traffic environment and achieve the simulation of a real urban traffic environment;

[0119] Import high-risk scenarios into the simulated road traffic environment to form a scenario library, store the scenario library in a standard format; such as OpenSCENARIO, support docking with the simulation platform, and regularly update the scenario library to supplement newly discovered boundary conditions.

[0120] Load the scenario library in a simulation platform (such as Carla, SUMO), load the test vehicle in the scenario library, deploy and run the autonomous driving algorithm on the test vehicle; the autonomous driving algorithm is the pedestrian trajectory prediction and emergency braking algorithm. Ensure that the test vehicle is exactly the same as the actual vehicle.

[0121] Record the test data and perform cross-validation of virtual and real data. The performance evaluation of the algorithm includes trajectory error and performance index. The trajectory error is the absolute difference between the interaction behavior trajectories of each pedestrian predicted by the autonomous driving algorithm on the test vehicle and the high-risk scenarios and those in the scenario library. The method for obtaining the performance index is to accumulate the braking distance, stopping distance, and offset of the vehicle;

[0122] Among them, the trajectory error is obtained by subtracting the interaction behavior trajectories of each pedestrian predicted by the autonomous driving algorithm from those in the scenario library and then taking the absolute value. The absolute difference in high-risk scenarios refers to the absolute difference between the number predicted by the autonomous driving algorithm on the test vehicle and that in the scenario library. The braking distance of the vehicle is the road length from the vehicle starting to avoid the pedestrian to the start of braking. The stopping distance is the distance between the vehicle and the pedestrian when it finally stops. The offset is the relative offset position of the vehicle after braking and stopping relative to the start of braking.

[0123] Output the performance evaluation of the algorithm.

[0124] Step S4: Arrange the test process and results on the visualization interface;

[0125] The method for arranging the test process and results on the visualization interface includes:

[0126] Display the operation of the simulation platform on the visualization interface;

[0127] Parallelly set R_P test vehicles, run the R_P test vehicles together in the scenario library, and conduct tests for S_F time periods;

[0128] Plot the performance evaluations of the algorithms on the R_P test vehicles for S_F time periods into a chart and arrange the chart on the visualization interface for intuitive observation by users.

[0129] The visualization interface includes the end page of a computer or the display interface of a mobile device and interacts with the user through the interface.

[0130] The user can perform interactive operations such as clicking, querying, and downloading the inspection report through the computer end page (computer display screen, where the computer here refers to an intelligent device, and the mobile device can be a mobile phone or a tablet). Embodiment 2

[0131] Please refer to Figure 1As shown in the figure, for the parts not described in detail in this embodiment, refer to the description in Embodiment 1. A test system for pedestrian trajectory prediction and emergency braking in an autonomous driving platform is provided, including:

[0132] Data acquisition and reconstruction module: used to calculate the segmentation coefficient according to the pre-segmented real road, obtain the level of small segments and perform resource allocation, and collect multi-source heterogeneous data of small segments;

[0133] High-risk scenario screening module: introduce importance optimization in the Transformer model to generate possible prediction trajectories of individuals, use the possible prediction trajectories to construct graph data, the GNN model obtains the interaction behavior trajectories after individual association by processing the graph data, introduce the Autoencoder anomaly detection mechanism, screen out high-risk trajectories, and output high-risk scenarios;

[0134] Algorithm evaluation module: combine the road geometric model with high-risk scenarios to form a scenario library, load the scenarios in the simulation platform and conduct autonomous driving tests, and perform cross-validation of virtual and real data in combination with real vehicle test data to obtain the performance evaluation of the algorithm;

[0135] Visualization module: arrange the test process and results on the visualization interface.

[0136] Through the collaboration of multiple models, the combination of Transformer, GNN, and Autoencoder respectively solves the problems of individual prediction, interaction modeling, and anomaly detection, forming a complete prediction and screening process. Based on the road grading and device dynamic deployment based on the segmentation coefficient, the data acquisition efficiency is optimized and the cost is reduced.

[0137] Through anomaly detection and high-risk behavior characteristics, high-risk scenarios are accurately screened, improving the pertinence of the test. The cross-validation of simulation and real vehicle tests ensures the reliability and practicality of the test results. The application of trajectory constraints, importance optimization, and social force model improves the physical rationality and behavioral logic of the prediction results. Embodiment 3

[0138] This embodiment publicly provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it realizes the running mode of the above-mentioned test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform.

[0139] Since the electronic device introduced in this embodiment is the electronic device used in the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform of this application embodiment, based on the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform introduced in this application embodiment, those skilled in the art can understand the specific implementation manner of the electronic device in this embodiment and its various variations. Therefore, the specific implementation of how this electronic device implements the method in this application embodiment will not be introduced in detail here. As long as those skilled in the art implement the electronic device used in the test method for pedestrian trajectory prediction and emergency braking in the autonomous driving platform of this application embodiment, it falls within the scope of protection of this application.

[0140] The above formulas are all dimensionless and take their numerical calculations. The formula is a formula obtained by collecting a large amount of data for software simulation to approximate the real situation. The preset parameters and threshold selection in the formula are set by those skilled in the art according to the actual situation.

[0141] The above are only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for ordinary technical users in this technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. A test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform, characterized in that, Including: Step S1: Pre-segment the real road environment, obtain the level of road segments according to the calculation of the segmentation coefficient, allocate resources to the segments according to the level, and collect multi-source heterogeneous data; Step S2: Introduce importance optimization in the Transformer model to generate possible prediction trajectories of individuals. Use the possible prediction trajectories to construct graph data. The GNN model obtains the interaction behavior trajectories after individual association by processing the graph data. Introduce the Autoencoder anomaly detection mechanism to filter out high-risk trajectories and output high-risk scenarios; Use the Transformer model to predict the possible behaviors of pedestrian individuals and generate possible prediction trajectories of individuals. The method includes: Set the input of the Transformer encoder in the Transformer model as the input vector, and the output as the encoded features, denoted as encoded features. The input of the conditional variational autoencoder in the Transformer model is the encoded features, the embedded features of the environmental context data, and the embedded features of the intention data, and the output is the parameters of the latent distribution. Obtain the Gaussian expression form of the latent distribution according to the parameters, sample latent variables from the latent distribution, and then generate F_O possible future trajectories; During the process of sampling latent variables from the latent distribution, importance optimization is introduced, that is, calculating the importance weights of each latent variable , is the encoded feature, is the embedded feature of the environmental context data, is the embedded feature of the intent data, is the th trajectory generated probability, is the prior distribution of the latent variable, represents the th latent variable, is the index of the trajectory, is the index of the latent variable; Arrange the importance weights of each latent variable in descending order, select the first A_O latent variables, denoted as possible prediction trajectories, and then output the possible prediction trajectories; Set trajectory constraints, including diversity constraints, physical constraints, and intention constraints. The diversity constraint is to calculate the minimum Euclidean distance between any two possible prediction trajectories at all time steps and maximize the minimum Euclidean distance. The physical constraint is that the acceleration and speed of the pedestrian are both less than the maximum acceleration and maximum speed. The intention constraint is to minimize the Euclidean distance between the trajectory and the target position; Step S3: Combine the road geometry model with the high-risk scenarios to form a scenario library, load the scenarios in the simulation platform and conduct autonomous driving tests, and combine the real vehicle test data for cross-validation of virtual and real data to obtain the performance evaluation of the algorithm; Step S4: Arrange the test process and results on the visualization interface.

2. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 1, characterized in that, The method of pre-segmenting the real road environment, obtaining the level of road segments according to the calculation of the segmentation coefficient, allocating resources to the segments according to the level, and collecting multi-source heterogeneous data includes: The multi-source heterogeneous data includes pedestrian trajectory data, traffic flow data, environmental data, and scenario data. The scenario data includes accident occurrence record data and approximate accident data; Pre-divide the real road into small segments and calculate the segmentation coefficient for each small segment , where is the total number of influencing factors, is the normalization value of the th influencing factor in the th cycle time, and the range is , is the mapping function for , is the dynamic weight of the th influencing factor in the th cycle time; Preset a first boundary threshold and a second boundary threshold, and the first boundary threshold is greater than the second boundary threshold. If the segmentation coefficient is greater than the first boundary threshold, it is determined as a high-risk section, and the acquisition equipment is densely deployed. If the segmentation coefficient belongs to the interval between the first boundary threshold and the second boundary threshold, it is determined as a medium-risk section, and the acquisition equipment is medium-densely deployed. If the segmentation coefficient is less than the second boundary threshold, it is determined as a low-risk section, and the acquisition equipment is low-densely deployed; Calculate the segmentation coefficient of the th cycle time and perform level judgment. According to the judgment output result, dynamically deploy the acquisition devices for the small segments on the th cycle time, and collect multi-source heterogeneous data in real time through the acquisition devices.

3. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 2, characterized in that, The high-risk scenarios and the acquisition method include: Perform cleaning and interpolation preprocessing on the collected multi-source heterogeneous data; For each small segment, extract multi-source heterogeneous data within G_U cycle times, and obtain a complete input vector through feature embedding processing; Use the Transformer model to predict the possible behaviors of pedestrian individuals and generate possible prediction trajectories of individuals; Use the GNN model to associate the interaction relationships between pedestrians, correct the possible prediction trajectories of individuals, and generate the interaction behavior trajectories after association; Introduce the Autoencoder anomaly detection mechanism to remove the interaction behavior trajectories with low probability of occurrence, retain the high-risk trajectories with high probability of occurrence, and output high-risk scenarios; 4. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 3, characterized in that, The method for obtaining a complete input vector through feature embedding processing includes: Extract the historical trajectory data, environmental context data, and intention data of each pedestrian to form raw data; Clean, smooth, and standardize the raw data; Independently embed the data of each modality in the raw data, including: for the historical trajectory data of pedestrians, use MLP or RNN to extract temporal features from the trajectory sequence; for the environmental context data, use MLP to embed the environmental data, and for the intention data of pedestrians, use MLP to embed the intention data, and normalize the embedded features; Calculate the mean and variance of the embedded features of each modality, make it close to the standard normal distribution by adjusting the activation function of the embedding layer, calculate the correlation matrix of the embedded features of each modality, and if the correlation exceeds the predefined judgment threshold, use PCA for dimensionality reduction; Take the embedded features of the historical trajectory data as the main features, and the embedded features of the environmental context data and intention data as auxiliary features, use the Cross-Attention mechanism to capture the interaction between modalities, and perform weighted fusion on the interacted features to obtain a complete input vector; 5. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 4, characterized in that The method for using the GNN model to associate the interaction relationships between pedestrians, correct the possible prediction trajectories of individuals, and generate the interaction behavior trajectories after association includes: Extract the distances between pedestrians and the distances between pedestrians and vehicles or roadblocks from multi-source data, denoted as interaction data, and combine the interaction data with the possible prediction trajectories of individuals as the input data of the GNN model; Construct a blank graph, represent each pedestrian as a node in the graph, the node features include the possible prediction trajectories of individuals, environmental context data, and intention data, represent the distances between pedestrians as edges in the graph, the edge features include distance information, and also represent the distances between pedestrians and vehicles or roadblocks as edges, the edge features include the positions and speeds of vehicles or roadblocks, to form graph data; Use the graph neural network GNN to process the graph data, and GNN updates the features of each node through multi-layer graph convolution or graph attention mechanism. During the update process, introduce the social force model constraint and output the updated node features; Map the updated node features back to the trajectory space and output the interaction behavior trajectories of each pedestrian; 6. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 5, wherein, The method for introducing the Autoencoder anomaly detection mechanism to remove the interaction behavior trajectories with low probability of occurrence, retain the high-risk trajectories with high probability of occurrence, and output high-risk scenarios includes: Use the Autoencoder time series anomaly detection model, with the input being the interaction behavior trajectories of each pedestrian and the output being the anomaly scores of each interaction behavior trajectory; Based on the anomaly threshold and trajectory features, define high-risk pedestrian types as including sudden crossing, direction change, and group avoidance. Set the anomaly threshold and trajectory features for each high-risk type. Among them, the trajectory feature of sudden crossing is that the pedestrian trajectory crosses the lane line, the trajectory feature of direction change is that the speed direction of the pedestrian trajectory undergoes a sudden change, and the trajectory feature of group avoidance is that the trajectories of more than one pedestrian change direction simultaneously; For the interaction behavior trajectory of each pedestrian, calculate the difference between the anomaly score and the anomaly threshold, and use the Sigmoid activation function to convert it into a probability value in combination with the trajectory features. Denote the probability value as the high-risk probability; Set the risk threshold. When the high-risk probability exceeds the risk threshold, the interaction behavior trajectory is marked as a high-risk behavior. Extract the environmental context data at the time of the high-risk behavior and combine it to form a high-risk scenario, and output the high-risk scenario.

7. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 6, characterized in that, The method of combining the road geometry model with the high-risk scenario to form a scenario library, loading the scenario in the simulation platform and conducting autonomous driving tests, and combining the real vehicle test data for virtual-real data cross-validation to obtain the performance evaluation of the algorithm includes: Based on the high-precision map and sensor data, construct a road geometry model, use the weather simulation tool to add weather conditions, generate traffic flow through SUMO, and import it into CARLA to simulate the road traffic environment; Import the high-risk scenario into the simulated road traffic environment to form a scenario library, and store the scenario library in a standard format; Load the scenario library in the simulation platform, load the test vehicle in the scenario library, and deploy and run the autonomous driving algorithm on the test vehicle; Record the test data and conduct virtual-real data cross-validation. The performance evaluation of the algorithm includes the trajectory error and the performance index. The trajectory error is the absolute difference between the interaction behavior trajectory of each pedestrian predicted by the autonomous driving algorithm on the test vehicle and the high-risk scenario and the scenario library. The method for obtaining the performance index is to accumulate the braking distance, stopping distance, and offset of the vehicle; Output the performance evaluation of the algorithm.

8. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 7, characterized in that, The method of arranging the test process and results on the visualization interface includes: Display the operation of the simulation platform on the visualization interface; Set R_P test vehicles in parallel, run the R_P test vehicles together in the scenario library, and conduct tests for S_F time periods; Plot the performance evaluation of the algorithm on the R_P test vehicles for S_F time periods into a chart and arrange the chart on the visualization interface for intuitive observation by users.

9. The test method for pedestrian trajectory prediction and emergency braking in an autonomous driving platform according to claim 8, wherein The visualization interface includes the end page of a computer or the display interface of a mobile device and interacts with the user through the interface.

Images