Dynamic weight correction and path deviation probability prediction method for vehicle track

Abstract

The invention discloses a dynamic weight correction and path deviation probability prediction method for a vehicle track, which comprises the following steps of: acquiring vehicle data and multi-source dynamic data of the vehicle track in real time through an optical sensor, a radio wave sensor and an inertial navigation sensor, carrying out space-time calibration, extracting obstacle characteristics and road structure characteristics, and predicting the path deviation probability of the vehicle track. The obstacle movement trend is quickly captured through a space-time diagram sequence and a diagram convolutional neural network, real-time obstacle avoidance is realized in combination with dynamic weight adjustment, a driving intention is predicted by using a Bayesian neural network, a trajectory planning strategy is adjusted through weight correction, and uncertainty is reduced by using multi-source data fusion and probabilistic prediction. A closed-loop feedback mechanism continuously optimizes the model, and efficient operation is kept in complex scenes such as intersections and roundabout through real-time weight adjustment and closed-loop feedback, so that the purpose of quickly responding to dynamic obstacles or driving behavior changes can be achieved, and the precision of the path deviation prediction probability is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle trajectory prediction, and particularly relates to a method for dynamically correcting the weights of vehicle trajectories and predicting the probability of path deviation. Background Art

[0002] In an autonomous driving system, the dynamic correction of vehicle trajectories and the prediction of path deviation probability are particularly important. By predicting the deviation probability in real time, the autonomous driving system can trigger emergency avoidance or deceleration strategies in advance. When it is predicted that the vehicle may deviate from the drivable area, the trajectory is regenerated by adjusting the path planning weights to ensure that the vehicle always remains within the safety boundary and reduce the collision risk.

[0003] Currently, the method for dynamically correcting the autonomous driving path is usually based on path planning correction through multi-source data fusion. For example, the patent document with the application number CN119063749A discloses a method, device, and equipment for dynamically correcting autonomous driving path planning. This method determines the drivable area of the current vehicle by obtaining the point cloud data of the millimeter-wave radar and the image data of the vision camera; constructs a grid map with composite weights, and searches for the global reference path based on the grid map through the Dijkstra algorithm; uses the dynamic programming method to determine the sampling point set of the current vehicle from the global reference path, fits a spline curve to the sampling point set, and convexifies the drivable area through the point sampling method based on the spline curve to obtain the convex space boundary of the drivable area; constructs a quadratic programming problem based on the convex space boundary, solves it to obtain the expected trajectory; when it is detected that the reference trajectory deviates from the drivable area, adjusts the weights of the sampling point set and the quadratic programming problem, and corrects the expected trajectory, which can improve the accuracy of the issued expected trajectory.

[0004] However, when using the dynamic programming method to determine the sampling point set of the current vehicle from the global reference path, the global reference path is searched based on the grid map through the Dijkstra algorithm. The grid map usually relies on sensors (such as millimeter-wave radars) for real-time scanning and updating. However, the positions of dynamic obstacles (such as pedestrians and moving objects) change frequently, resulting in the map update speed being unable to match the environmental changes, and the adaptability to the dynamic environment is poor. Therefore, we need to propose a method for dynamically correcting the weights of vehicle trajectories and predicting the probability of path deviation to solve the above existing problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for dynamically correcting the weights of vehicle trajectories and predicting the probability of path deviation, which can quickly respond to changes in dynamic obstacles or driving behaviors, improve the accuracy of the path deviation prediction probability, and thus improve the adaptability to the dynamic environment, so as to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention adopts the following technical solutions: Dynamic Weight Correction and Path Deviation Probability Prediction Method for Vehicle Trajectory, including the following steps: S1. Real-time collect the vehicle's own data and multi-source dynamic data of the vehicle trajectory through optical sensors, radio wave sensors, and inertial navigation sensors, perform spatio-temporal calibration, and extract obstacle features and road structure features; S2. Construct a spatio-temporal graph sequence with obstacles as nodes and spatial distance and time correlation as edges; S3. Combine obstacle features and road structure features to predict the ideal trajectory of the vehicle without obstacle interference; S4. Predict the future trajectory of obstacles through graph convolution and spatio-temporal attention mechanism, and fuse it with the ideal trajectory and road structure features to form an interactive trajectory set; S5. Define a path deviation probability threshold based on the road structure or navigation path, and use a Bayesian neural network to predict the path deviation probability and uncertainty interval of future frames; S6. Define a state space according to the prediction results of the Bayesian neural network, and correct the data fusion weight according to the defined state space; S7. Real-time monitor the corrected trajectory, calculate the deviation from the actual trajectory, and iteratively correct the weight strategy according to the deviation to form a closed-loop feedback of prediction-evaluation-correction.

[0007] Preferably, the optical sensor uses a high-definition camera to collect image data, and identifies the obstacle category and position through a target detection algorithm; the radio wave sensor uses a millimeter-wave radar to obtain the obstacle distance, speed, and reflection intensity; the inertial navigation sensor obtains the vehicle's own motion reference through an inertial measurement unit and a global navigation satellite system as the reference coordinate system for trajectory prediction.

[0008] Preferably, during spatio-temporal calibration, align the timestamps of different sensors through a hardware clock, then unify the sensor data into the world coordinate system, and complete the coordinate transformation through an extrinsic parameter matrix.

[0009] Preferably, when extracting obstacle features, combine the camera image and millimeter-wave radar point cloud, associate the 2D bounding box with the 3D obstacle cluster through spatio-temporal calibration, detect the obstacle category and position collected by the optical sensor with a target detection model, extract the obstacle distance, speed, and reflection intensity collected by the millimeter-wave radar through a clustering algorithm, and finally output the obstacle position, size, speed, category, and spatio-temporal context information; When extracting road structure features, match the vehicle pose with a high-precision map, extract static elements such as lane lines and traffic signs, segment the image with a semantic segmentation model, identify the drivable area and lane lines, and combine the map and visual information to output the road structure features of the lane line equation and drivable area mask.

[0010] Preferably, the process of constructing the spatio-temporal graph sequence is as follows: A1. Define each obstacle as a node and connect the nodes with spatio-temporal correlation; A2. Calculate the spatial distance weight, time interval weight, and comprehensive edge weight according to the defined nodes; A3. Connect the graphs of consecutive time steps along the time axis to form a dynamic graph sequence.

[0011] Preferably, the ideal trajectory is generated based on the center line of the lane or the navigation path, combined with the vehicle kinematic model, to generate an ideal trajectory without obstacle interference.

[0012] Preferably, the process of predicting the future trajectory of an obstacle is as follows: B1. Use a graph convolutional network to process the spatio-temporal graph to capture the interaction features between obstacles; B2. Introduce an attention mechanism to focus on key obstacles and calculate the attention weights between nodes; B3. Based on the captured interaction features between obstacles and the attention weights, predict the future trajectory through a recurrent neural network.

[0013] Preferably, the method of defining the path deviation probability threshold is to set the threshold based on the road structure and vehicle state. If the lateral deviation of the trajectory from the reference path exceeds the set threshold, it is considered a deviation; if the lateral deviation of the trajectory from the reference path is within the set threshold range, it is considered not a deviation.

[0014] Preferably, when predicting the path deviation probability of future frames, use a Bayesian neural network to predict the path deviation probability of future frames. Take the set of interaction trajectories, vehicle state information, and environmental features as the input of the Bayesian neural network, estimate the posterior distribution of path deviation through a probability model, obtain multiple prediction results through multiple samplings, calculate the mean and variance of the path deviation probability, and provide a probabilistic prediction result through the Bayesian neural network to quantify the uncertainty of the prediction.

[0015] Preferably, the state space includes the deviation probability, uncertainty, and sensor reliability index output by the Bayesian neural network; when correcting the data fusion weight, based on the state space, adopt an adaptive weighted fusion algorithm to dynamically adjust the weight model in combination with the deviation probability in the state space.

[0016] The method for dynamically correcting the weight of vehicle trajectories and predicting the path deviation probability proposed by the present invention has the following advantages compared with the prior art: 1. The present invention rapidly captures the motion trends of obstacles through spatio-temporal graph sequences and graph convolutional neural networks, realizes real-time obstacle avoidance by combining dynamic weight adjustment, predicts driving intentions using Bayesian neural networks, adjusts the trajectory planning strategy through weight correction, reduces uncertainty by using multi-source data fusion and probabilistic prediction, and continuously optimizes the model through a closed-loop feedback mechanism. Through real-time weight adjustment and closed-loop feedback, it can operate efficiently in complex scenarios such as intersections and roundabouts, so as to achieve the purpose of being able to quickly respond to dynamic obstacles or changes in driving behavior, improve the accuracy of path deviation prediction probability, and thus improve the adaptability to dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Shows a flowchart according to an embodiment of the present invention; Figure 2 Shows a flowchart for constructing a spatio-temporal graph sequence according to an embodiment of the present invention; Figure 3 Shows a flowchart for predicting the future trajectory of an obstacle according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] 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. The specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention. 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.

[0019] The present invention provides a method for dynamically correcting weights and predicting path deviation probability of a vehicle trajectory as shown in Figures 1-3 and includes the following steps: S1. Real-time collect the vehicle's own data and multi-source dynamic data of the vehicle trajectory through an optical sensor, a radio wave sensor, and an inertial navigation sensor, perform spatio-temporal calibration, and extract obstacle features and road structure features; The optical sensor uses a high-definition camera to collect image data and identifies the types and positions of obstacles through a target detection algorithm; the radio wave sensor uses a millimeter-wave radar to obtain the distance, speed, and reflection intensity of obstacles; the inertial navigation sensor obtains the vehicle's own motion reference through an inertial measurement unit and a global navigation satellite system as a reference coordinate system for trajectory prediction. Integrate an optical sensor (such as a camera), a radio wave sensor (such as a millimeter-wave radar), and an inertial navigation sensor (IMU) to collect the vehicle's own state (speed, acceleration, attitude) and trajectory data (position, direction) in real time. The multi-source data fusion captures the obstacle position, speed, and road structure features (such as lane lines, curvature), providing a basis for dynamic weight adjustment; When performing spatio-temporal calibration, the timestamps of different sensors are aligned through the hardware clock, and then the data of each sensor is unified into the world coordinate system. The coordinate transformation is completed through the extrinsic parameter matrix. The coordinate transformation formula is: , where is the point in the transformed coordinate system, is the rotation matrix, is the translation vector, is the point in the sensor coordinate system. Through spatio-temporal calibration operations, the time delay and spatial deviation of multi-sensor data are eliminated, ensuring data consistency. Spatio-temporal calibration avoids prediction errors caused by data misalignment and improves the accuracy of subsequent steps; When extracting obstacle features, combining the camera image and the millimeter-wave radar point cloud, associating the 2D bounding box with the 3D obstacle cluster through spatio-temporal calibration, and using the object detection model to detect the category and position of the obstacles collected by the optical sensor. The object detection model is set to the YOLOv8 model. The distance, speed, and reflection intensity of the obstacles collected by the millimeter-wave radar are extracted through the clustering algorithm. Finally, the obstacle position, size, speed, category, and spatio-temporal context information are output; When extracting road structure features, matching the vehicle pose with the high-precision map, extracting static elements such as lane lines and traffic signs, segmenting the image with the semantic segmentation model. The semantic segmentation model is the DeepLabv3 model, identifying the drivable area and lane lines, and combining the map and visual information to output the road structure features such as the lane line equation and the drivable area mask.

[0020] S2. Using obstacles as nodes and the spatial distance and time correlation as edges to construct a spatio-temporal graph sequence. For example, in the vehicle following scenario, the leading vehicle is used as a node, and its relative distance and speed difference with the following vehicle are used as edge weights. The spatio-temporal graph sequence effectively captures the motion trends and interaction patterns of obstacles, providing structured support for predicting their future trajectories; The spatial distance and time correlation jointly define the edge weights, reflecting the real-time influence between obstacles; Such as Figure 2 shown, the construction process of the spatio-temporal graph sequence is as follows: A1. Define each obstacle as a node and connect the nodes with spatio-temporal correlation; A2. Calculate the spatial distance weight, time interval weight, and comprehensive edge weight according to the defined nodes. The calculation formula for the spatial distance weight is: , where is the spatial distance weight value, and are the positions of two nodes respectively, is the spatial scale parameter; The time interval weight calculation formula is as follows: , where is the time interval weight value, and are the positions of two nodes respectively, is the time scale parameter; The comprehensive edge weight calculation formula is as follows: , where is the comprehensive edge weight value; A3. Connect the graphs of consecutive time steps along the time axis to form a dynamic graph sequence; Model the dynamic environment as a structured graph to facilitate capturing the spatio-temporal interaction relationships between obstacles. The graph sequence can represent the temporal dynamics of the environment and provide an interaction modeling basis for subsequent trajectory prediction; S3. Combine the obstacle characteristics and road structure characteristics to predict the ideal trajectory of the vehicle without obstacle interference; The ideal trajectory is based on the lane centerline or navigation path and combines the vehicle kinematic model to generate an ideal trajectory without obstacle interference. The ideal trajectory prediction formula is as follows: , where is the velocity component in the x direction, is the velocity component in the y direction, is the vehicle speed, is the vehicle steering angle, is the vehicle wheelbase, is the vehicle heading angle, is the heading angle change rate of , that is, the angular velocity; Taking the ideal trajectory as a reference provides a reference for subsequent deviation probability prediction and weight correction. The planning-based method considers road constraints to ensure the feasibility of the predicted trajectory.

[0021] S4. Predict the future trajectories of obstacles through graph convolution and spatio-temporal attention mechanism, and fuse them with the ideal trajectory and road structure characteristics to form an interactive trajectory set; As Figure 3 shown, the process of predicting the future trajectories of obstacles is as follows: B1. Use the graph convolutional network to process the spatio-temporal graph to capture the interaction features between obstacles. The formula for capturing the interaction features between obstacles is: , where is the node feature of the +1 layer, is the weight matrix, is the activation function, is the The layer node features is the degree matrix is the adjacency matrix with self-loops added; B2. Introduce an attention mechanism to focus on key obstacles and calculate the attention weights between nodes; The calculation formula for the attention weight is: where is the attention weight of node towards node representing the degree of attention, is the query vector of node is the key vector transposed, is the exponential form of the similarity score, is the normalization term used to sum the similarity scores of all nodes related to node to ensure that the sum of the attention weights is 1; B3. Based on the captured interaction features between obstacles and the attention weights, predict the future trajectory through a recurrent neural network. The future trajectory prediction formula is: where is the predicted position of obstacle at future time step, is the historical feature vector of the th obstacle from time step 1 to t, is the composite module of the graph convolutional network and the attention mechanism, used to extract the spatial interaction features and key information weights of obstacles; is the recurrent neural network used to process the time series features to generate future trajectory predictions; The algorithm formula for fusing the future trajectory of the obstacle with the ideal trajectory and the road structure features is: where is the interaction trajectory, is the ideal trajectory, is the ideal trajectory weight, is the attention weight of obstacle is the predicted trajectory of obstacle , and N is the total number of obstacles; Efficiently handle spatio-temporal dependencies through graph convolution and the attention mechanism, improve the accuracy of obstacle trajectory prediction. The set of interaction trajectories covers multiple possible scenarios, providing redundant options for path planning.

[0022] ​​S5. Define a path deviation probability threshold based on the road structure or navigation path, and use a Bayesian neural network to predict the future frame path deviation probability and uncertainty interval; The path deviation probability threshold is defined based on the road structure and vehicle state. If the lateral deviation of the trajectory from the reference path exceeds the set threshold, it is considered a deviation. If the lateral deviation of the trajectory from the reference path is within the set threshold, it is considered not to deviate; When predicting the future frame path deviation probability, use a Bayesian neural network (BNN) to predict the future frame path deviation probability. The interactive trajectory set, vehicle state information (such as speed, position, attitude), and environmental features (such as road structure, obstacle distribution) are used as the input of the BNN. The posterior distribution of path deviation is estimated through a probability model. The BNN is implemented using the Monte Carlo dropout (MC Dropout) or variational inference method. Multiple prediction results are obtained through multiple samplings, and the mean and variance of the path deviation probability are calculated. A probabilistic prediction result is provided through the Bayesian neural network, quantifying the uncertainty of the prediction (such as the confidence interval), providing confidence support for decision-making, and enabling the system to evaluate risks through probabilistic prediction; When predicting the uncertainty interval, based on the multiple prediction results of the Bayesian neural network, calculate the mean and variance of the path deviation probability to form an uncertainty interval. The mean represents the estimated value of the path deviation probability, and the variance reflects the degree of uncertainty of the prediction; S6. Define a state space based on the prediction results of the Bayesian neural network, and correct the data fusion weights according to the defined state space; The state space includes the deviation probability, uncertainty, and sensor reliability index output by the Bayesian neural network. When correcting the data fusion weights, based on the state space, an adaptive weighted fusion algorithm is adopted, and the weight model is dynamically adjusted in combination with the deviation probability in the state space. The formula expression for dynamically adjusting the weight model is: , where is the updated weight after combining the trajectory credibility and the weight adjustment amount, is the weight before update, is the learning rate that controls the weight update step size, is the path deviation probability, is the weight adjustment amount of the th sensor; Dynamically adjust the sensor weights according to the environmental uncertainty to improve the data fusion accuracy. Enhance the weights of reliable sensors in high deviation probability scenarios to reduce the risk of misjudgment; S7. Real-time monitor the corrected trajectory, calculate the deviation from the actual trajectory, and iteratively correct the weight strategy according to the deviation to form a closed-loop feedback of prediction-evaluation-correction; The weight policy iteration correction formula is as follows: , where is the value of the weight parameter after the +1 -th iteration update, is the weight parameter value at the -th iteration, is the learning rate used to control the step size of each weight update, is the derivative vector of the objective function with respect to , is the objective function used to measure the error between the model prediction result and the actual value; Through a sequence of spatio-temporal graphs constructed with obstacles as nodes and spatio-temporal correlations as edges, the real-time position, speed, and interaction relationships of obstacles can be directly captured. Without waiting for the grid map to be updated, it can reflect environmental changes. By using graph convolution and spatio-temporal attention mechanisms to predict the future trajectories of obstacles, it can perceive the movement trends of obstacles in advance. Through dynamic fusion weight correction, it can adjust the fusion ratio of optical sensor, radio wave sensor, and inertial sensor data in real time according to the path deviation probability and uncertainty interval. For example, when the obstacle trajectory prediction shows high dynamics (such as a vehicle cutting into the current lane), the weight of radar data can be increased, reducing the dependence on the static position of obstacles in the grid map; if the Bayesian network predicts a high path deviation probability, the weight of inertial navigation data can be increased to strengthen the vehicle kinematic constraints and reduce the dependence on the road boundaries in the grid map; By predicting - evaluating - correcting the deviation between the corrected trajectory and the actual trajectory in a closed-loop in real time, the weight policy is dynamically adjusted. For example: if it is found that the vehicle deviates from the path due to the grid map not updating the obstacle position in time, the closed-loop feedback can quickly reduce the weight of the grid map data to avoid error accumulation; through multiple iterations, the system can learn a better weight allocation strategy in dynamic obstacle scenarios, such as preferentially using the obstacle trajectory predicted by radar rather than the historical position in the grid map. Through this mechanism, the impact of the grid map update delay can be adaptively compensated. That is, when the environmental dynamics is high, such as in a dense pedestrian scenario, the system automatically increases the weight of the prediction model and reduces the dependence on real-time grid updates; when the environment is relatively static, it restores the trust in sensor data to balance the computational efficiency and accuracy.

[0023] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for dynamically correcting the weight of a vehicle trajectory and predicting the probability of path deviation, characterized in that: It includes the following steps: S1. Real-time collect multi-source dynamic data of the vehicle's own data and the vehicle's trajectory through an optical sensor, a radio wave sensor, and an inertial navigation sensor, perform spatio-temporal calibration, and extract obstacle features and road structure features; S2. Construct a spatio-temporal graph sequence with obstacles as nodes and the spatial distance and time correlation as edges; S3. Combine the obstacle features and road structure features to predict the ideal trajectory of the vehicle without obstacle interference; S4. Predict the future trajectory of the obstacle through graph convolution and spatio-temporal attention mechanism, and fuse it with the ideal trajectory and road structure features to form an interactive trajectory set; S5. Define a path deviation probability threshold based on the road structure or navigation path, and use a Bayesian neural network to predict the path deviation probability and uncertainty interval of future frames; S6. Define the state space according to the prediction result of the Bayesian neural network, and correct the data fusion weight according to the defined state space; S7. Real-time monitor the corrected trajectory, calculate the deviation from the actual trajectory, and iteratively correct the weight strategy according to the deviation to form a closed-loop feedback of prediction-evaluation-correction.

2. The method for dynamically correcting the weight of a vehicle trajectory and predicting the path deviation probability according to claim 1, characterized in that: The optical sensor uses a high-definition camera to collect image data, and identifies the obstacle category and position through a target detection algorithm; The radio wave sensor uses a millimeter-wave radar to obtain the obstacle distance, speed, and reflection intensity; the inertial navigation sensor obtains the vehicle's own motion reference through an inertial measurement unit and a global navigation satellite system, as the reference coordinate system for trajectory prediction.

3. The method for dynamically correcting the weight of a vehicle trajectory and predicting the path deviation probability according to claim 2, characterized in that: When performing spatio-temporal calibration, align the timestamps of different sensors through a hardware clock, then unify the sensor data into the world coordinate system, and complete the coordinate transformation through an external parameter matrix.

4. The method for dynamically correcting the weight of a vehicle trajectory and predicting the path deviation probability according to claim 3, characterized in that: When extracting obstacle features, combine the camera image and the millimeter-wave radar point cloud, associate the two-dimensional bounding box with the three-dimensional obstacle cluster through spatio-temporal calibration, detect the obstacle category and position collected by the optical sensor with a target detection model, extract the obstacle distance, speed, and reflection intensity collected by the millimeter-wave radar through a clustering algorithm, and finally output the obstacle position, size, speed, category, and spatio-temporal context information; When extracting road structure features, match the vehicle pose with the high-precision map, extract static elements such as lane lines and traffic signs, segment the image with a semantic segmentation model, identify the drivable area and lane lines, and combine the map and visual information to output the road structure features of the lane line equation and the drivable area mask.

5. The method for dynamically correcting the weight of a vehicle trajectory and predicting the path deviation probability according to claim 4, characterized in that: The process of constructing the spatio-temporal graph sequence is as follows: A1. Define each obstacle as a node and connect the nodes with spatio-temporal correlation; A2. Calculate the spatial distance weight, time interval weight, and comprehensive edge weight according to the defined nodes; A3. Connect the graphs of consecutive time steps along the time axis to form a dynamic graph sequence.

6. The method for dynamically weighting correction of vehicle trajectory and prediction of path deviation probability according to claim 5, wherein: The ideal trajectory is based on the lane centerline or navigation path, and combines the vehicle kinematic model to generate an ideal trajectory without obstacle interference.

7. The method for dynamically weighting correction of vehicle trajectory and prediction of path deviation probability according to claim 6, characterized in that: The process of predicting the future trajectory of the obstacle is as follows: B1. Use a graph convolutional network to process the spatio-temporal graph and capture the interaction features between obstacles; B2. Introduce an attention mechanism to focus on key obstacles and calculate the attention weight between nodes; B3. Based on the captured interaction features between obstacles and attention weights, predict the future trajectory through a recurrent neural network.

8. The method for dynamically weighting and correcting vehicle trajectories and predicting path deviation probabilities according to claim 7, characterized in that: The path deviation probability threshold is defined based on the road structure and vehicle state. If the lateral deviation of the trajectory from the reference path exceeds the set threshold, it is considered a deviation. If the lateral deviation of the trajectory from the reference path is within the set threshold, it is considered not to deviate.

9. The method for dynamically correcting the weight of a vehicle trajectory and predicting the path deviation probability according to claim 8, characterized in that: When predicting the path deviation probability of future frames, use a Bayesian neural network to predict the path deviation probability of future frames. Take the set of interaction trajectories, vehicle state information, and environmental features as the input of the Bayesian neural network. Estimate the posterior distribution of path deviation through a probability model, obtain multiple prediction results through multiple samplings, calculate the mean and variance of the path deviation probability, and provide a probabilistic prediction result through the Bayesian neural network to quantify the uncertainty of the prediction.

10. The method for dynamically correcting the weight of a vehicle trajectory and predicting the path deviation probability according to claim 9, characterized in that: The state space includes the deviation probability, uncertainty, and sensor reliability index output by the Bayesian neural network. When correcting the data fusion weights, based on the state space, adopt an adaptive weighted fusion algorithm to dynamically adjust the weight model in combination with the deviation probability in the state space.

Images